Written by Konda — Paddy & Water Inc. / Strategy

AI Agents & Predictive Management — An Introduction

― Ending “reactive management” and building an organization where AI looks ahead ―

June 2026

Introduction: Why AI Agents and Predictive Management, Now

0.1　The Chronic Illness of “Reactive Management Structures”

At many companies, management meetings have become a forum for “looking back at last month’s results.” The reasons sales declined, the reasons inventory piled up, the reasons complaints increased — all of it is discussed only “after the fact.” This happens because management information exists only in the “present and past,” with no data pointing toward the “future.” We call this structure “reactive management.” Its defining trait is that action is taken only after a problem becomes visible, meaning countermeasure costs are high, timing is late, and prevention is impossible.

Reactive Management vs. Predictive Management — What’s the Difference

[Reactive Management] Management meeting topic: “Sales fell by XX million yen last month. What was the cause?” Time orientation of information: Past (mainly monthly/weekly reports) Starting point of decision-making: After a problem becomes apparent Typical costs: Emergency response, inventory write-offs, complaint handling, lost opportunity [Predictive Management] Management meeting topic: “There are early signs that sales will fall by XX million yen next month. What should we do this week?” Time orientation of information: Present + future (real-time data + AI forecasts) Starting point of decision-making: Before a problem becomes apparent Typical cost reduction: Prevention, inventory optimization, maximizing opportunity → AI agents are the engine that converts “reactive management” into “predictive management”

0.2　Why AI Agents Are a “Game Changer”

Since 2023, generative AI (ChatGPT, Claude, Gemini, etc.) has spread rapidly, but the “AI agent” is the next stage beyond it. If generative AI is a tool that “answers questions,” an AI agent is a system that, given a goal, autonomously thinks, plans, acts, and produces results. This difference makes a decisive difference in how AI is used in management.

Stage of AI evolution	What it can do	Level of management use	Representative examples
1st generation: Rule-based AI (– 2015)	Judges and classifies within rules defined by humans	Automation of routine tasks (FAQ response, scoring)	Credit scoring, spam filters
2nd generation: Machine learning AI (2015–2022)	Automatically learns patterns from data and forecasts	Demand forecasting, anomaly detection, recommendations	SAP IBP demand forecasting, manufacturing line anomaly detection
3rd generation: Generative AI (2022–2024)	Answers questions in natural language, generates text and code	Efficiency in information gathering, document drafting support, knowledge search	ChatGPT, Claude, Copilot (document drafting support)
4th generation: AI agents (2024–)	Given a goal, autonomously plans, uses tools, executes, and verifies	Realization of predictive management; business process automation via multi-agent systems	SAP Joule, AutoGen, LangGraph, Agentforce

0.3　Who This Book Is For, and How It Is Organized

• Executives and management planning staff: readers who want to understand how AI agents will change management decision-making. Focus especially on the Introduction, Chapter 1, Chapter 2, and Chapter 7

• DX promotion and IT department leaders: readers who want to grasp the technical composition of AI agents and the adoption roadmap. Focus especially on Chapter 1, Chapter 4, Chapter 5, and Chapter 6

• Managers in operational departments (procurement, manufacturing, finance, sales): readers who want a concrete picture of how AI can be used in their own department. Focus especially on Chapter 3 and Chapter 4

• Those responsible for SAP/ERP implementation: readers considering integration design with SAP Business AI. Focus especially on Chapter 6

Chapter 1: What Is an AI Agent? — Concepts and Composition

1.1　Definition of an AI Agent

An AI agent is an AI system that, “when given a goal, perceives its environment (Perception), reasons and forms an action plan (Reasoning & Planning), acts using tools (Action), and observes the results to autonomously iterate until the goal is achieved (Observation & Iteration).” What fundamentally distinguishes it from conventional AI is that it can autonomously solve problems without a human giving instructions one by one.

The AI Agent Operating Cycle (the ReAct Loop)

[Setting the Goal] Receives a goal from a human, e.g., “Come up with a plan to cut next month’s manufacturing costs by 5%” | ▼ [Reasoning (inference and planning)] The AI autonomously decomposes the task into subtasks: ① Analyze the past three months of cost data ② Identify the major cost drivers (materials, energy, outsourcing, cost of defects) ③ Rank the items with the greatest potential for reduction ④ Generate options for reduction measures ⑤ Run simulations of the measures ⑥ Document the recommended plan | ▼ [Action (using tools and executing)] Autonomously selects and executes the necessary tools: → Queries SAP Material Ledger data → Analyzes cost structure with statistical analysis tools → Simulates material costs with a supply/demand planning system → Automatically generates an Excel report | ▼ [Observation (reviewing and evaluating the results)] Checks the results obtained and self-evaluates “is this sufficient to achieve the goal?” If not sufficient → returns to Reasoning and re-plans (autonomous loop) If sufficient → presents the final result to the human

1.2　The Four Components of an AI Agent

Component	Role	Specific examples
① Large Language Model (LLM: the Brain)	The core engine responsible for reasoning, planning, natural language understanding, and generation. Autonomously judges “what to do next” from the given goal and context	GPT-4o / Claude 3.5 / Gemini 1.5 / SAP’s in-house LLM. Combining multiple LLMs is also increasingly common
② Tools (the Hands)	The “hands” through which the agent interacts with external systems, data, and APIs. The more tools available, the broader the range of tasks it can handle	Database queries, web search, running calculations, code generation, calling SAP APIs (BAPI/REST), sending email, file operations, supplier portal integration
③ Memory / Context	A mechanism that retains past interactions, learned results, and intermediate outputs, enabling long-term task execution. There are two kinds: “short-term memory” and “long-term memory”	Short-term: conversation history, task progress. Long-term: vector databases (internal knowledge, past cases) — e.g., past engineering change order (ECO) cases, negotiation history with business partners, product quality records
④ Orchestrator (the Director)	A mechanism that coordinates and controls multiple agents and tools to efficiently carry out complex tasks. The linchpin of multi-agent systems	LangChain / LangGraph / AutoGen / CrewAI; agent orchestration on SAP BTP

1.3　Single Agents and Multi-Agent Systems

Simple tasks can be handled by a single AI agent (a single-agent system), but complex management challenges require a “multi-agent system” in which multiple agents divide roles and collaborate. In predictive management for manufacturing, a demand-forecasting agent, a procurement-planning agent, a cost-analysis agent, and a quality-monitoring agent share information and coordinate in real time to support globally optimal decision-making.

Agent type	Tasks it suits	Why a multi-agent system becomes necessary
Single agent	Generating a specific report; analyzing a single data source; routine queries and responses; one-directional information gathering	A single agent has inherent limits: it “cannot process in parallel,” “cannot hold all specialized knowledge in one place,” and “cannot cross-check itself”
Multi-agent (role-division type)	Analysis spanning multiple systems; chained planning across demand, procurement, production, and finance; decision support requiring multiple areas of expertise	A “demand-forecasting specialist agent,” a “procurement-optimization specialist agent,” and a “financial-impact-calculation agent” each execute their specialized tasks in parallel, and finally an integration agent consolidates the results
Multi-agent (checking type)	Verification of important decisions; risk assessment and regulatory compliance checks; analysis from multiple viewpoints	A “proposal agent” puts forward an idea, a “critic agent” raises objections and points out risks, and an “arbiter agent” makes the final call — a structure that mimics human team decision-making

Chapter 2: What Is Predictive Management? — Concept and Design

2.1　The Definition of Predictive Management and Its Three Layers

Predictive management is “a management model in which AI predicts the future from present data, allowing humans to act ahead of that future.” Predictive management is composed of the following three layers.

Layer	Name	Content	Role of the AI agent
Layer 1	Descriptive (what happened)	Aggregates and visualizes past and present data to grasp “what is happening now” (traditional BI reports and dashboards)	Automating data collection, aggregation, and visualization — “automatically producing yesterday’s production-results summary”
Layer 2	Predictive (what will happen)	AI predicts the future from patterns in past data — “What will demand be next month?” “When will the next failure occur?”	Demand forecasting, early anomaly detection, risk scoring — “forecasting next month’s demand fluctuation with 98% accuracy”
Layer 3	Prescriptive (what should be done)	Takes the “prediction” and autonomously proposes and executes “what should be done about it” — the most advanced form of predictive management	The AI agent drafts, simulates, and recommends measures — “In preparation for rising demand, Part A should be reordered within the coming week: recommended quantity XX units, estimated cost YY0,000 yen”

2.2　The Four Walls Blocking Predictive Management

Behind the fact that many companies say they “want to pursue predictive management” yet fail to achieve it are four structural walls. Designing an AI agent implementation begins with breaking down these four walls.

Wall	Content	Solution through AI × data integration
① The wall of data (silos)	Data is fragmented by department. Sales data, manufacturing data, financial data, and supplier data are each locked inside separate systems	A data integration platform (data lake, SAP BTP, data fabric) unifies the silos. AI agents query multiple data sources in real time across the organization
② The wall of quality (dirty data)	Data is incomplete, inaccurate, or inconsistently defined — “the same item code differs by department,” “entries are frequently missing”	Automated data cleansing and anomaly detection by AI agents, combined with master data management (SAP MDG), continuously improves data quality
③ The wall of speed (latency)	With monthly or weekly reporting cycles, “predictive information doesn’t arrive in time for management decisions”	Combining real-time data streams (SAP HANA in-memory, Kafka) with AI agents brings the decision-making lag down to nearly zero
④ The wall of interpretability (explainability)	The “reasons” behind AI’s predictions and recommendations are unclear, so executives cannot trust them or use them in decision-making	Implementing explainable AI. Agents are built with a function that explains, in natural language, “why this demand forecast” — including the underlying data and factors

2.3　The Data Maturity Required for Predictive Management

The level of predictive management a company can achieve depends on its “data maturity.” Before introducing AI agents, it is essential to accurately assess your own organization’s data maturity.

Level	State of the data	What AI can do	Typical company situation
L1: Fragmented	Managed department-by-department in Excel/Access; data definitions are not unified	Rule-based automation only; simple aggregation and visualization	“Sales in Salesforce, inventory in Excel, finance in SAP” — the data is not connected
L2: Integrated	Data is consolidated in the ERP; quality is reasonably assured	Demand forecasting and anomaly detection via machine learning become possible; BI dashboards function	Core data is unified in an ERP such as SAP, but real-time capability and external data integration remain insufficient
L3: Real-time	ERP data can be referenced in real time; IoT and external data are connected	Real-time forecasting, automatic alerts, and scenario simulation by AI agents become possible	A real-time analytics foundation is in place on SAP S/4HANA + BTP; IoT data integration is partially realized
L4: Predictive optimization	Company-wide data is integrated, clean, and accessible in real time	Autonomous forecasting, decision-making, and execution by multi-agent systems become possible — full predictive management	A digitally advanced manufacturer with a digital thread that extends to include suppliers

Chapter 3: Seven Frontlines of Management That AI Agents Are Transforming

This chapter organizes, with cutting-edge case studies, what AI agents are transforming right now across the seven core functions of management: sales, procurement, production, inventory, accounting, human resources, and quality management. Read on not simply as a set of isolated functional improvements, but with the perspective that “predictive management” is achieved only once these seven functions are connected.

3.1　Sales — From “Gut Feeling and Experience” to “Prediction and Optimization”

Sales organizations are among the areas that benefit earliest and most tangibly from AI agents. Data and AI dramatically sharpen decisions about “which customer, when, and with what proposal” to approach.

Area of application	Specific AI agent actions	Leading examples / impact
Lead scoring & prioritization	Integrates CRM data, web behavior logs, email open rates, deal history, and external company data to automatically rank the “top 20 customers most likely to close” every morning. Also presents the basis for the score (e.g., “viewed the website’s pricing page three times last week”)	Increases the win rate without reducing the number of deals sales reps pursue. Cases report a 30–40% improvement in average lead-conversion rate using Salesforce Einstein
Dynamic pricing (real-time price optimization)	Integrates competitor prices, supply-demand balance, customer price sensitivity, inventory status, and contract-renewal timing in real time to calculate “the optimal price to offer this customer today.” Automatically flags deals that require discount approval	Amazon changes the price of the same product hundreds of times a day. In manufacturing too, AI is increasingly setting order prices dynamically, raising win rates while preserving margins
Churn prediction (preventing cancellation and attrition)	Monitors changes in purchase frequency and transaction volume, inquiry content, and payment-delay patterns to detect in advance that “this customer has an 80% probability of churning within three months.” Automatically drafts retention proposals ahead of time	Standard practice at telecom and SaaS companies. In manufacturing, alerts fire before recurring parts orders stop, letting sales get ahead of the problem. Some cases protect 3–8% of revenue through churn reduction
Next Best Action	AI analyzes the deal stage, the customer’s purchasing cycle, and patterns from similar past deals to automatically suggest to the sales rep “the action to take this week” (run a demo, submit a quote, bring in a manager). Also presents the priority and rationale for the suggestion	With SAP Sales Cloud + AI, the “next recommended action” is automatically displayed for every deal. Reduces the coaching burden on sales managers while improving results for junior reps
Demand-linked pricing and promotion planning	Forecasts demand 13–26 weeks ahead from past sales results, market trends, competitor activity, and macroeconomic data, and automatically plans “when, for which product, and which promotion to run.” Links in real time with production and inventory planning	In consumer packaged goods, AI integration of demand forecasting and promotion planning has cut stockout rates by 50% in some cases. In manufacturing too, pre-order forecast accuracy feeds directly into production leveling

3.2　Procurement — From Cost Cutting to “Strategic Sourcing Intelligence”

Procurement and sourcing is an area where AI is transforming everything from price negotiation to supplier selection to risk management. AI agents integrate market information, supplier data, and internal purchasing data to support, at a strategic level, decisions about “what to buy, when, and from whom.”

Area of application	Specific AI agent actions	Leading examples / impact
Automated spend analytics	AI automatically categorizes all purchasing data (invoices, POs, credit card statements) using natural language processing, visualizing “is this cost fixed or variable,” “are there duplicate orders,” and “what percentage of spend is concentrated with a given supplier”	Enables full analysis of a million purchase records per month — impossible manually. SAP Ariba’s spend-analytics AI automatically categorizes over 90% of previously unclassified spend items. Reports indicate an average of 7–12% in newly discovered cost-reduction opportunity
Real-time supplier risk monitoring	Integrates supplier financial scores, rating changes, news coverage, social media, and shipping data to update a risk score daily. Issues advance alerts such as “this supplier shows signs of financial deterioration: bankruptcy risk rising” or “port strikes may affect delivery”	Demand for supply-chain risk monitoring AI has surged since the pandemic. Tools such as Resilinc and Riskmethods (integrated with SAP Ariba) have even put into practice a feature that automatically suggests alternative supplier candidates
AI-assisted price negotiation and benchmarking	AI integrates and analyzes market price databases for similar parts and materials, past negotiation records, and raw-material cost trends to present buyers with an “appropriate price range” and “angles for negotiation.” AI automatically compares quotes and ranks them by lowest price plus quality score	Procurement AI tools such as GEP and Zycus autonomously issue RFQs to multiple suppliers, collect responses, and compare them — buyers make only the final decision. Cases report average sourcing-cost reductions of 5–15%
Automated contract analysis and risk extraction	AI automatically reads procurement contracts and highlights “unfavorable clauses,” “terms that deviate from the standard,” and “renewal deadlines and cancellation conditions.” Builds a database of contract terms by supplier for comparative analysis	Contract-analysis AI powered by LLMs cuts legal review time by 70%. In M&A due diligence in particular, thousands of contracts can be scanned within days
Tail-spend management (automating low-value purchases)	AI dynamically sets rules to auto-approve and auto-order low-value (“tail”) purchases. For example: “if under 50,000 yen, from an approved supplier, and within budget,” the AI places the order automatically with no human approval needed	60–80% of procurement staff time is spent processing small orders. Automating this with AI lets buyers focus on strategic sourcing

3.3　Production — From “Building to Schedule” to “Continuously Building Optimally”

AI agents in the production domain change both “planning” and “execution on the ground.” AI analyzes real-time data from IoT sensors to keep equipment, people, materials, and energy in a constantly optimal state.

Area of application	Specific AI agent actions	Leading examples / impact
AI production scheduling (constraint-satisfaction optimization)	Considers order information, equipment utilization, worker shifts, material inventory, and tool availability all at once to generate the “optimal production sequence for today, next week, and next month” within minutes. Instantly reschedules in response to equipment failure or sudden order changes	More accurate and faster to compute than conventional APS. At large manufacturers, AI-driven updates to production scheduling have increasingly delivered a 20% cut in changeover time and a 15% rise in production efficiency. SAP’s PPO (Production Planning Optimization) is in the process of being AI-enabled
Predictive maintenance	IoT sensors continuously collect equipment vibration, temperature, current, and acoustic data. AI detects patterns that deviate from the normal range and predicts, for example, “an 80% probability of failure within 72 hours for this piece of equipment.” Automatically suggests the optimal maintenance timing and the parts required	Offered as a product by Siemens, GE, Hitachi, and others. Numerous cases in Japan and abroad report equipment downtime cut by up to 50% and maintenance costs cut by 25–30%. Some plants at Schaeffler (automotive parts) have achieved zero unplanned downtime through predictive-maintenance AI
Digital-twin-linked production optimization	Synchronizes a digital copy (digital twin) of the physical plant, equipment, and product in real time. AI repeatedly runs simulations on the digital twin to discover “which settings, sequence, and conditions produce the highest output and quality”	In the Siemens Xcelerator digital twin, AI autonomously optimizes manufacturing parameters. At BMW’s digital-twin factory, planning time for production-line changes has been cut by 90%. Every problem is solved in the virtual factory before physical production begins
AI-optimized energy consumption control	Monitors power consumption by piece of equipment across the plant in real time. AI automatically controls consumption — e.g., “at the current production volume, output from this piece of equipment can be reduced,” or “avoid peak power consumption during hours when electricity prices are high.” Also integrates management of solar generation and battery storage	Electricity costs account for 5–15% of manufacturing cost. AI-driven energy management cut cooling power by 40% at Google’s data centers. Manufacturing sites have reported 10–20% energy-cost reductions as well
Autonomous coordinated control of robots and cobots	AI coordinates the movements of multiple robots, collaborative robots (cobots), and human workers in real time, autonomously reallocating tasks as conditions change. Vision AI recognizes the position and orientation of parts in real time and automatically corrects assembly precision	More than 2,000 robots work in AI-coordinated fashion at Amazon’s fulfillment warehouses. At Volkswagen, cobot adoption pushed door-assembly defect rates close to zero. Dynamic AI-driven line control continues to evolve at Tesla’s highly automated factories as well

3.4　Inventory — AI Automatically Resolving “Too Much and Too Little”

Inventory management is an optimization problem with an inherent dilemma: holding too much incurs capital cost, while holding too little causes lost sales. AI agents model demand uncertainty, lead-time variation, and supplier risk together to dynamically calculate optimal inventory levels.

Area of application	Specific AI agent actions	Leading examples / impact
Multi-echelon inventory optimization	Jointly optimizes inventory across multiple tiers — headquarters warehouse, regional distribution centers, stores, and external warehouses. Calculates “where, what, and how much to hold” for minimum total cost, and automatically generates replenishment instructions	Supported by SAP IBP (Inventory Optimization), o9 Solutions, and others. In retail, cases have cut total company-wide inventory by 20–30% while simultaneously lowering the stockout rate. In manufacturing, inter-site transfer costs are also factored into inventory placement optimization
Demand sensing (ultra-short-term demand detection)	Collects POS data, e-commerce access logs, order logs, weather, and social-media trends daily/hourly, and AI detects “signals of a demand surge next week” in advance. Substantially improves accuracy 1–3 weeks out compared with statistical forecasting	P&G, Unilever, and others have put demand sensing into practice. Many cases report a 30–50% reduction in short-term forecast error. In manufacturing too, inventory and production can be adjusted ahead of a sudden increase in orders
Dynamic optimization of safety stock	AI dynamically replaces the fixed safety-stock formula (safety factor × standard deviation of demand × √lead time). Learns supplier-specific lead-time variability, seasonality, and special factors to continuously update the optimal safety stock for each item	Moves from the rule-of-thumb “about two months’ worth just in case” to a safety stock grounded in actual data. Cases achieve both a 20–40% inventory reduction and an improved stockout rate at the same time
Obsolescence-risk prediction (preventing write-offs before they happen)	Integrates inventory age, demand trends, product life cycle, and engineering change order (ECO) information so AI can predict in advance that “this part carries a high risk of being scrapped unused within 90 days.” Proposes options such as “prioritize consumption now / transfer to another site / negotiate a return with the supplier”	Manufacturers write off tens of millions to hundreds of millions of yen in obsolete inventory every year. Multiple studies estimate that AI-based obsolescence prediction can cut such write-off losses by 50–70%
Autonomous replenishment	When inventory falls below a set threshold, AI automatically calculates the order quantity, supplier, and timing, and auto-generates a purchase order in the ERP. Human approval is required only for exceptions (urgent, large-quantity, or new-supplier orders)	Walmart has fully automated replenishment ordering for some products. In manufacturing, automated ordering of MRO (Maintenance, Repair & Operations) supplies is also spreading. Buyers are freed to spend their time on strategic sourcing

3.5　Accounting & Finance — From “Finance That Reports” to “Finance That Predicts”

Accounting and finance is one of the areas where “automation of routine work” has advanced earliest through AI. Beyond automated journal entry, reconciliation, and report generation lies a higher-order role now emerging for AI: predicting the company’s future financial condition so management can act ahead of it.

Area of application	Specific AI agent actions	Leading examples / impact
Autonomous monthly close	AI performs automatic classification of journal entries, reconciliation, variance analysis, and provisioning calculations. Automates over 80% of the routine work in month-end closing, escalating only anomalies and exceptions to humans. AI also checks all account-balance reconciliations	Cases in Japan and abroad using SAP S/4HANA + AI have shortened the monthly close from 10 days to within 5. At one large manufacturer, a draft financial statement is now completed automatically by the morning of closing day
Intelligent invoice processing	AI reads invoices arriving on paper, as PDFs, or by email, and automatically handles amount and line-item extraction, tax-category matching, reconciliation against purchase orders, and approval routing. Only exceptions (amount mismatches, no matching PO) are routed to humans	Automating PO matching and three-way matching (PO, goods receipt, invoice) has cut AP processing effort by 70% in many cases. SAP Intelligent Robotic Process Automation (iRPA) combined with AI is the primary toolset
Cash-flow forecasting and treasury optimization	AI integrates receivables collection forecasts (calculating each customer’s payment pattern and delay probability), payables schedules, and capital-expenditure plans to automatically update a “13-week cash flow forecast” every week. Also proposes the timing for investing surplus cash or drawing on credit lines	Global companies use AI to manage cash positions across dozens of countries in an integrated way. AI autonomously optimizes treasury while factoring in FX risk and interest-rate movements, changing the CFO’s role from “waiting for data” to “reviewing AI’s proposals”
Tax compliance automation	AI learns the rules for VAT, corporate tax, transfer pricing, and customs duties, and automatically determines the tax category the moment a transaction occurs. AI automatically ingests tax-law changes in each country and automatically assesses the impact on existing transactions	Tax filing at multinational companies consumes tens of thousands of hours annually. SAP Tax Compliance + AI detects tax errors in advance. KPMG and Deloitte are rolling out tax AI products that automate a large share of tax work
Advanced fraud detection and internal controls	AI continuously monitors every accounting transaction — journal entries, approvals, payments, and expense reimbursements — statistically detecting “anomalous transaction patterns” in real time. Forensic analysis is also automated	AI automatically performs internal-control checks for J-SOX compliance and automatically collects audit evidence. PwC’s AI audit tool can verify 100% of transactions without sampling. Early fraud-detection rates have improved substantially

3.6　Human Resources — From “Managing People” to “Predicting and Developing People’s Potential”

AI agents in HR are evolving from “efficiency in recruiting” toward “maximizing the potential of talent.” Beyond quantitative data, AI draws on time-series changes in meeting remarks, written text, and performance to predict which employees “will grow,” which “carry flight risk,” and who is “well suited to a given role.”

Area of application	Specific AI agent actions	Leading examples / impact
Recruiting and screening automation	AI learns an “ideal candidate profile” for a given role from profiles of past successful hires. AI automatically screens and scores applications. Interview scheduling and coding-test grading are also automated	Unilever and L’Oréal have deployed large-scale recruiting AI, cutting applicant-screening time by 75%. However, bias in AI hiring — unfairly disadvantaging certain attributes — has also been pointed out, making fairness audits essential
Attrition prediction and retention measures	AI integrates and analyzes attendance patterns, changes in overtime hours, shifts in 1-on-1 comments, performance-review trends, sentiment scores from internal social channels, and salary-market comparisons to predict “an 80% risk that this employee will leave within three months.” Alerts managers and HR in advance and proposes retention measures	According to LinkedIn research, the cost of losing a high performer is 1.5–2.5 times their annual salary. SAP SuccessFactors’ Retention Risk analysis has reportedly achieved over 80% attrition-prediction accuracy at some companies
Skills mapping & growth-path design	AI builds a database of every employee’s skills, experience, qualifications, and project track record. Automatically proposes “this employee’s skill gaps,” “the experience and training needed for their next stage of growth,” and “the position they will be suited to in three years”	Supported by SAP SuccessFactors Skills Graph and Workday Skills Cloud. Unilever reports that 70% of its employees have adopted AI-suggested skill-development paths, also boosting employee motivation
Workforce planning	AI integrates the business plan, attrition forecasts, skill gaps, and market hiring difficulty to simulate “how many people, with what skills, will be short three years from now.” Also automatically estimates the cost comparison between hiring, development, and outsourcing	At Japanese companies facing a shrinking, aging population, identifying “the talent shortage 3–5 years out” early and building a development/hiring strategy has become urgent. AI-driven workforce planning prevents the “we didn’t notice the shortage until it hit” scenario
Performance-review support & bias detection	AI analyzes each reviewer’s rating tendencies (lenient, harsh, halo effect, recency effect) and warns, “this evaluation may reflect XX bias.” Provides objective supporting evidence based on behavioral data (project contributions, number of code reviews, volume of documents produced)	AI visualizes rating distortions caused by seniority or personal rapport with a manager, supporting the move to a fairer evaluation system. It is especially useful for supplementing “hard-to-see contributions” with data in remote-work settings

3.7　Quality Management — From “Detection Through Inspection” to “Prevention Before It Happens”

AI is accelerating the shift in quality management from the reactive approach of “inspecting after production” toward a predictive approach that “prevents quality problems before they occur.” AI agents learn manufacturing parameters, sensor data, and past defect patterns to monitor quality risk in real time.

Area of application	Specific AI agent actions	Leading examples / impact
AI vision inspection (appearance and dimensional inspection)	AI analyzes camera images in real time to automatically detect scratches, discoloration, chips, and dimensional anomalies. Faster, more accurate, and capable of running 24 hours a day compared with conventional visual inspection or contact sensors. Detection results are automatically logged in quality records	Rapidly spreading in semiconductors, LCD panels, automotive parts, food, and pharmaceuticals. Cognex, Keyence, Omron, and others are the leading players. Cases report detection accuracy that exceeds human visual inspection while cutting inspection cost by 60–80%
In-process quality predictive control	Collects temperature, pressure, rotation speed, material lot, and worker ID in real time during manufacturing. AI predicts mid-process that “at the current manufacturing conditions, the defect rate will be three times normal,” and automatically adjusts conditions or halts the line as an emergency measure — aiming for zero post-production scrap	Mitsubishi Electric’s e-F@ctory series and Bosch’s AI-driven quality management have cut in-process defect rates by 60–80% in some cases. At Daikin Industries, combining manufacturing parameters with AI has achieved a zero-defect production line
Automated root-cause analysis (RCA)	When a quality problem occurs, AI automatically collects equipment, material, operator, process-condition, and environmental data and estimates the root cause. Presents a reasoned finding such as, “this defect stems from a combination of wear on equipment XX, elevated humidity, and material lot XX”	Conventional RCA takes a quality engineer days to weeks to analyze. AI can present candidate root causes within hours. Some organizations have also put into practice a flow where AI auto-drafts the CAPA (corrective and preventive action) for the quality team to review
Automated customer-complaint analysis and response	AI automatically classifies customer complaints (quality, delivery, billing, operator error, etc.) and routes them to the responsible department. AI presents similar past complaints and their resolutions. From the content of a complaint, it automatically determines whether the issue is “a product-design problem or a manufacturing-process problem” and files an engineering change request (ECR)	With SAP QM + AI, 80% of the process from complaint intake to initial response is automated. Some cases have cut the average time to first response to customers from 72 hours to under 4 hours. Trend analysis of complaints also contributes to early identification of recall candidates
Automated supplier-quality scoring	AI automatically aggregates inspection results, on-time delivery rate, quantity discrepancies, and defect-response speed for all parts received from a supplier, continuously updating a real-time quality score. Automatically issues a SCAR (supplier corrective action request) to suppliers whose score declines	A shift from “annual periodic supplier evaluation” to “weekly-updated, real-time quality scoring.” Early detection of problem suppliers prevents quality incidents before they occur. SAP Ariba’s supplier quality management module is in the process of being AI-enhanced

Connecting the Seven Functions Completes “Predictive Management”

[An Image of Cross-Functional Coordination Among AI Agents Across the Seven Functions] The sales AI predicts, “Demand for Product A will rise 20% next month” | ▼ The inventory AI automatically detects, “Current stock cannot meet the demand increase. Parts X and Y are short” | ▼ The procurement AI proposes, “Order additional Part X from Supplier A and Part Y from Supplier B. Presenting the optimal quantity, price, and delivery date” | ▼ The production AI generates, “An optimal schedule to meet the demand increase. Minimum cost through a combination of overtime and outsourcing” | ▼ The quality AI issues an advance warning: “Speeding up production to meet the demand surge risks a higher defect rate. Presenting process conditions that meet it without loosening quality standards” | ▼ The accounting AI instantly calculates, “Additional cost is approximately XX0,000 yen. Can the margin be maintained at XX%? Financial impact assessed immediately” | ▼ The HR AI calculates, “The additional headcount needed to meet the demand increase, comparing the cost of internal resources versus hiring or outsourcing” | ▼ → The executive makes the final GO/NO-GO call only after the AI has finished laying all the groundwork

Chapter 4: A Roadmap for Realizing Predictive Management

4.1　The Four Phases of AI Agent Adoption

Trying to achieve everything at once when introducing AI agents to realize predictive management will always fail. Classifying tasks by “size of impact” and “ease of implementation,” and building capability step by step, is the key to success.

Phase	Approximate duration	What is done	Criteria for success
Phase 1: Automate	3–6 months	Automated generation of routine reports (weekly, monthly); automated data collection and aggregation; automation of simple queries and notifications; making ERP/SAP data available via API. Run PoCs (proofs of concept) on several small-scale tasks	Routine work effort for staff cut by 30% or more; report-generation time cut by 80%; trust in AI is taking root within the organization
Phase 2: Predict	6–12 months	Building and validating demand-forecasting models; full operation of an anomaly-detection system; using predictive dashboards in management meetings; ongoing data-quality improvement	Demand-forecast accuracy improves 10–20% versus before; “AI forecasts” become a topic on the management-meeting agenda; monthly reports of early anomaly detection start appearing
Phase 3: Recommend	12–18 months	AI agents propose responses based on forecasts; scenario-simulation capability is implemented; cross-departmental information sharing and decision support; AI complements human decision-making	Monthly reports of cases where “the AI’s proposal was adopted as-is”; decision-making speed visibly improves; obsolete inventory and lost opportunity are measurably reduced
Phase 4: Execute	18–36 months	AI agents execute within a defined scope without human approval; multi-agent systems are in full operation; humans shift into the role of “monitor, approver, and exception handler” for AI; a continuous learning and improvement loop is established	Routine procurement, inventory adjustment, and schedule changes are executed autonomously by AI; only exceptional cases are escalated to humans; a culture of predictive management takes root in the organization

4.2　How to Prioritize AI Use Cases

To get the greatest impact from limited resources, it is important to evaluate and prioritize use cases using an “impact × feasibility” matrix.

Evaluation criterion	What scores high	What to check
Business impact	Large annual cost savings or revenue gain; directly tied to a critical management issue; the improvement effect can be measured quantitatively	Calculate the current loss amount. Quantify concretely, e.g., “XX0 million yen in obsolete inventory annually” or “Y person-months of effort for the monthly close”
Data readiness	The necessary data already exists in the ERP; data quality is above a certain level; real-time data access is possible	Confirm not only whether the data “exists” but whether it is “usable” (quality, freshness, access rights)
Technical feasibility	Achievable with existing tools and cloud services; can use off-the-shelf AI such as SAP Business AI; the PoC period fits within three months	Avoid building from scratch; make maximum use of off-the-shelf AI services such as SAP Joule, Azure OpenAI, and AWS Bedrock
Difficulty of change management	Minimal disruption to existing operations; easy to gain cooperation from operational staff; risk is small even if it fails	Remove the anxiety that “AI will take away the work of the people on the ground.” Choose a first use case that “makes the staff’s job easier”

4.3　Organizational and Workforce Transformation — Moving to an “Organization That Works Alongside AI”

Even once the technology is ready, predictive management cannot be realized without transforming the organization and its people. Introducing AI agents means fundamentally changing business processes and job definitions.

Role / job	Current main duties	After AI agent adoption	Newly required skills
Demand/supply planner	Creating demand forecasts; drafting inventory plans; managing and adjusting in Excel	Validating the reasonableness of AI forecasts; judging exceptions and special factors; instructing and adjusting AI parameters	AI literacy; business judgment; statistical thinking
Finance/accounting staff	Journal-entry work and aggregation; preparing monthly reports; variance analysis	Reviewing and approving AI-generated reports; explaining analysis to management; drafting improvement actions	Data interpretation; business-problem analysis; communication skills
Procurement/purchasing staff	Comparing quotes and placing orders; negotiating with suppliers; managing delivery dates	Final approval of AI-recommended options; building strategic supplier relationships; risk judgment	Negotiation and relationship-building skills; supply-chain risk judgment; AI tool operation
Production control staff	Creating production schedules; managing and adjusting progress; process improvement	Reviewing AI-generated schedules; managing equipment and personnel exceptions; instructing AI on improvement proposals	Problem-solving ability; process-improvement skills; ability to work with AI
Management planning / CFO	Aggregating and analyzing monthly results; budget management; preparing materials for management meetings	Formulating strategy based on AI forecasts; deciding “what to do next”; evaluating return on investment	Strategic thinking; ability to interpret AI forecasts; risk judgment

Chapter 5: Pitfalls in AI Agent Adoption, and Principles for Success

5.1　Typical Failure Patterns

Failure pattern	Why it happens	How to prevent it
“We’ll bring in AI once the data is ready” syndrome	The company keeps postponing, saying “data quality isn’t good enough yet for AI.” Perfect data never arrives	Don’t demand perfect data from the start. Begin with “what can be done with the data we have now,” running small PoCs repeatedly while improving data quality and applying AI in parallel
Overreliance on “leaving it to AI”	Blindly trusting the AI agent’s forecasts and recommendations to the point of losing independent business judgment — making important decisions “because the AI said so”	Treat AI strictly as a “decision-support aid.” Design the process so a human always gives final confirmation on important decisions, and always require the AI’s recommendation to come with an explanation of “why”
Scope creep	What started small keeps growing as “this too, and that too” gets added, causing the project to balloon and drag on	Fix Phase 1’s scope at exactly three months, one business process, and quantitative success criteria. Move to the next phase only after delivering results
Resistance and distrust on the ground	Fear that “AI will take our jobs” makes frontline staff passive about using AI, or causes them to neglect data entry	Choose a first use case that automates “the task frontline staff dislike the most” (e.g., routine report creation). Staff feeling that “this made things easier” is the greatest driver of adoption
The “stuck at PoC” problem	The proof of concept (PoC) went well, but the project cannot move into production or company-wide rollout	Define “criteria for moving to production” and a “rollout plan” from the PoC design stage. Keep the data, systems, and users as close as possible between PoC and production
Deferring security and governance	Deferring security design in favor of “let’s first get something working,” only to discover problems right before going live	Establish an AI governance policy (what data may be given to AI, who approves AI recommendations) at the design stage

5.2　Five Principles for Successful Predictive Management

Principle ①: Start Small, Learn Fast

Don’t aim from the outset for “predictive transformation across the whole enterprise.” Start with “one issue in one business process.” Deliver results within three months, earn the organization’s trust, and then move to the next step. Waiting for perfection before you move is the single biggest risk.

Principle ②: Business-First, Tech-Second

Define “which business problem to solve” before deciding “which AI to use.” The correct order is: “we want to improve demand-forecast accuracy → what data does that require → what system do we need to obtain that data.”

Principle ③: Clearly Define the Division of Roles Between Humans and AI (Human-in-the-Loop)

Clearly design which “areas the AI executes autonomously” and which “areas a human must always judge.” In particular, designing so that human approval is mandatory for high-value orders, external commitments to customers, and responses to exceptional situations is the foundation of long-term trust.

Principle ④: Build in Explainability by Design

Being able to explain, in natural language, “why the AI produced this prediction or recommendation” is a precondition for earning the trust of executives and frontline staff. A “black-box AI” won’t be used even if it’s accurate. Always design the system to clearly present the basis for a prediction, its confidence interval, and its underlying assumptions.

Principle ⑤: Build in a Continuous Learning and Improvement Loop

An AI model’s real work begins after release. As market conditions and business processes change, model accuracy degrades. Build in, from the design stage, a continuous-improvement mechanism that “monitors the AI’s prediction accuracy monthly and automatically triggers retraining when degradation is detected.”

Chapter 6: SAP × AI Agent Integration Scenarios

6.1　The Big Picture of SAP Business AI

Since 2023, SAP has been rolling out AI capabilities across its entire product line under the umbrella of “SAP Business AI.” At its core are SAP Joule, a natural-language AI assistant, and a group of AI services on SAP BTP (Business Technology Platform). For SAP customers, this is the most accessible gateway to AI-agent-driven predictive management that makes use of the data already accumulated in the ERP.

SAP AI component	Overview of functionality	Main use cases
SAP Joule (natural-language AI assistant)	A generative AI assistant embedded throughout the SAP system. Lets users operate, query, and analyze SAP functionality in natural language. Simply asking “what were the main drivers of last month’s cost variance” returns an analysis	Natural-language queries against management dashboards; natural-language instructions for transactions; automatic generation of reports and documents; suggested responses to anomaly alerts
SAP AI Core (AI model execution platform)	An SAP BTP service for training, deploying, and managing machine-learning models. Can integrate SAP’s own models with external LLMs (GPT, Claude, etc.)	Building and running demand-forecasting models; quality-prediction and anomaly-detection models; deploying custom AI agents; fine-tuning using SAP data
SAP AI Launchpad (AI management portal)	A dashboard for managing the lifecycle of AI models (version control, accuracy monitoring, deployment management)	Centrally managing multiple AI scenarios; continuous monitoring of model accuracy; managing AI governance and audit trails
SAP Datasphere (data integration platform)	A data-fabric platform that integrates SAP and non-SAP system data into a form AI can use	Integrating SAP with Salesforce and external databases; a foundation for real-time data access; integrating data sources for AI agents
SAP Analytics Cloud (SAC) Predictive	Predictive-analytics capability built into SAC. Enables no-code creation of demand-forecast and financial-forecast scenarios	Demand-forecast simulation; driver-based financial-plan forecasting; waterfall and variance analysis

6.2　Predictive Management with SAP Joule in Practice

SAP Joule functions not merely as a way to “turn ERP GUI operations into natural language,” but as an AI agent that “analyzes data across the organization and generates management insight.” Representative use cases are shown below.

Use case	Example natural-language instruction to Joule	What Joule does
Early awareness of inventory risk	“Tell me the top 10 items at risk of stockout next month”	① Retrieves demand-forecast data from IBP ② Cross-checks inventory, open orders, and lead time ③ Calculates and ranks stockout-risk scores ④ Presents response options (order quantity and timing) for high-risk items
Immediate analysis of cost variance	“Analyze the products with the largest manufacturing-cost variance this month and their main causes”	① Retrieves cost-variance data from Material Ledger (CKMLCP/ACDOCA) ② Decomposes it into price variance, quantity variance, and efficiency variance ③ Ranks the top five products by variance size ④ Explains the root cause of each variance in natural language
Supplier risk assessment	“List the suppliers with high procurement risk”	① References the MM material master and vendor master ② Calculates sourcing concentration (dependence on a specific supplier) ③ Analyzes recent on-time delivery rate and quality performance ④ Calculates a risk score and presents recommendations (including alternative sourcing candidates)
Automatic generation of a management forecast report	“Put together a summary of last month’s results and next month’s forecast for this week’s management meeting”	① Aggregates actual sales, profit, inventory, and procurement-cost results ② Calculates next month’s figures using SAC Predictive’s forecasting model ③ Automatically generates comparisons with the prior month and the same month last year ④ Outputs the report as a PowerPoint slide deck

6.3　SAP Integration Design for Custom AI Agents

For advanced scenarios that SAP Joule cannot handle, or use cases that span systems outside of SAP, custom AI agents can be built on SAP BTP. The following are representative integration patterns.

• Pattern ①: Agent integration via the SAP BAPI/OData API. A custom agent runs on a framework such as LangChain and calls SAP’s OData API as a “tool,” autonomously performing tasks such as inventory queries, creation of production orders, and invoice processing

• Pattern ②: Integration using SAP Datasphere as the data hub. Data integrated in Datasphere is used as the AI agent’s vector database (for RAG), letting the AI reference internal knowledge, design documents, and past cases to improve the quality of its answers

• Pattern ③: Multi-agent orchestration on SAP BTP. A “demand-forecasting agent,” a “procurement-planning agent,” and a “financial-impact agent” coordinate via SAP BTP’s service mesh to deliver globally optimal decision support

• Pattern ④: Integrating an external LLM (Azure OpenAI/Claude/Gemini) with SAP data. Azure OpenAI is connected to SAP BTP AI Core, and an AI agent is built that is grounded (via RAG) in SAP ERP data — generating high-accuracy answers while referencing SAP data in its current state

Chapter 7: The Future of Predictive Management, Toward 2030

7.1　Entering the Age of “Agentic AI”

Between 2025 and 2030, AI agents will further expand their autonomy, capability, and scope of coordination. In the trend known as “Agentic AI,” AI will come to possess a higher order of autonomy — not merely executing tasks, but “setting goals, forming plans, mobilizing other AIs and humans, and evaluating outcomes.”

Timeframe	AI agent capability	Impact on manufacturing and management
2024–2025 (present)	Practical use of single agents; spread of embedded AI such as SAP Joule; automation of routine work	The main results so far: automating routine reports, improving demand-forecast accuracy, and speeding up the monthly close. Currently at the stage where “AI proposes something”
2026–2027	Practical use of multi-agent systems; autonomous information-sharing among agents; autonomous execution of longer, more complex tasks	Groups of AI agents autonomously optimize the chained plan spanning demand, procurement, production, and finance. The stage where “AI proposes and the human approves” becomes mainstream
2028–2030	AI takes on goal-setting and evaluation as well; integration with real-time digital twins; seamless human-AI collaboration	Reaches the stage where “AI drafts even the strategy.” Management planning, financial planning, and risk management become a model where “humans adjust the AI’s proposal.” AI fully autonomously executes some routine decisions

7.2　How Predictive Management Changes “The Executive’s Job”

As AI agents take on more of forecasting, analysis, planning, and execution, the substance of the work done by executives and managers will fundamentally change. What decreases is “time spent gathering, aggregating, and analyzing information”; what increases is “time spent judging, creating, and building relationships.”

• Work that decreases: reading, collecting, and aggregating routine reports; investigating root causes after a problem occurs; meetings and inquiries held simply to understand the current state

• Work that increases: judging the validity of AI’s predictions and proposals; human judgment on irregularities and exceptions; medium- to long-term strategic conception; building relationships and negotiating with stakeholders

• Newly required skills: AI literacy (understanding AI’s limits and strengths); data thinking (reading the story behind the numbers); the ability to design and direct agents (designing what to have the AI do)

7.3　Ethics and Governance — Making AI Agents a “Trustworthy Management Partner”

The more deeply AI agents are involved at the core of management, the more important their ethical design and governance framework become. AI that is “fast and smart, but untrustworthy” becomes a management risk. Below are the essential governance requirements for predictive management.

Governance requirement	Overview	Points for implementation
Accountability	All of an AI’s decisions and actions must be traceable, with clarity on who approved what	Log every action taken by an AI agent in an audit trail. Clearly distinguish between “the AI placed the order” and “a human approved it”
Transparency	Stakeholders can understand what the AI is doing and why it reached a given conclusion	Always display the basis for a prediction, the data used, and the confidence interval. An AI that cannot explain “why” cannot be used in production
Fairness	The AI must not give unjust advantage or disadvantage to a particular supplier, customer, or product	Check training data for bias. Conduct regular fairness audits
Human oversight	Design so that a human is always able to be involved in and approve important decisions	Set thresholds for amount and risk; any case exceeding them must always go through a human-approval workflow. Never allow a situation where the AI becomes “uncontrollable”
Data security	Managing the confidentiality and appropriate scope of use of data given to the AI	Strictly separate data sent to external LLMs from data restricted to internal use. Mask personal information and competitively sensitive data

Chapter 8: The World AI Agents Are Changing — Hypotheses and a Vision of the Future

This chapter is not “established fact” but “well-grounded hypothesis.” If we extrapolate current AI-agent trends, how will the society, companies, work, and human roles we live within change? This is a bold, speculative sketch of that future.

8.1　A Monday Morning in 2035 — A Vision of the Future

As a hypothesis, a morning like this might arrive.

A Monday at a Manufacturing Company’s Management Meeting in 2035 (An Imagined Future)

It is 6 a.m. Before the management meeting even begins, the “answer” is already in. The AI agents have kept working all weekend, and by 7 a.m. Monday, the large screen in the meeting room already displays, organized in natural language, “last week’s results across every business, this week’s forecast, recommended actions, and the top three risks.” The demand AI reports: “Demand for Product C in Southeast Asia will surge next month. The basis is our local partner’s 　order-backlog data and the local weather forecast (rising demand for air conditioning due to extreme heat).” The procurement AI has already generated three scenarios for a supplier negotiation to increase production, presenting: “Recommended option: Supplier A can agree to a 30% production increase, six weeks earlier. 　The condition is a payment-terms change from 45 days to 30 days. Financial impact already calculated.” The finance AI adds: “Forecast Q2 margin impact if we meet this demand increase: +1.2%. 　However, there is a 0.4% downside risk given rising container freight rates.” The quality AI warns: “Historical patterns show that increasing production speed raises the defect rate at Process D. 　We recommend limiting the speed increase to 15%; the impact on the schedule is minor.” The agenda of the management meeting is no longer “should we increase production, yes or no,” but simply “do we accept the AI’s proposed condition change (shortening the payment terms).” The executive’s job has been distilled down to “making the judgment” and “taking responsibility for the decision.”

8.2　Hypothesis ① — The Shape of the “Company” Changes

Today’s companies are structured as a pyramid: functional departments connected by layers of management. If AI agents take over the flow of information between departments, planning, and routine work, this structure may be fundamentally called into question.

• Hypothesis: The role of “middle management whose job is to compile information” shrinks. If AI takes on the function of integrating sales, procurement, production, and finance information in real time and producing “globally optimal proposals,” the need for a layer of managers who manually compiled each department’s information declines

• Hypothesis: In its place, the “human-AI collaborative team” becomes the basic unit of management. A small but high-density team, directing and running multiple AI agents, becomes able to produce output equivalent to a traditional large enterprise. This makes possible the “miniaturization and agility of the company”

• Hypothesis: An era arrives where “a company of 500 employees” and “a company of 50 employees plus a large number of AI agents” compete in the same market. The latter carries dramatically lower fixed costs. This rewrites the meaning of “economies of scale” in business strategy

• Hypothesis: The supply chain also changes. If AI can autonomously and fully optimize design, sourcing, manufacturing, logistics, and sales, a “fluid, network-based sourcing model in which AI selects the optimal partner every time” may rise in place of long-term, vertically integrated supply chains

8.3　Hypothesis ② — The Definition of “Work” Changes

When the Industrial Revolution replaced physical labor with machines, humans shifted to “work that requires thinking.” When AI agents replace intellectual, routine work, what will humans shift to?

• Hypothesis: The market value of “the ability to produce the right answer” falls. Since AI can produce correct answers faster and more accurately, the scarcity of “skill in finding the right question that a correct answer presupposes” rises relative to “skill in finding the right answer.” The ability to decide “what should be optimized” becomes the scarcest and most valuable skill

• Hypothesis: The spread of “specialized knowledge” accelerates, even as its value falls. If AI can instantly provide expert knowledge in law, medicine, accounting, and engineering, it becomes hard to differentiate on the sheer volume of knowledge one holds. What differentiates people is the judgment and experience of “which problem to apply that knowledge to, and in what context”

• Hypothesis: What remains as “work only humans can do” converges on four things: ① taking on responsibility (AI has no capacity for accountability); ② empathy, trust, and relationship-building (human connection with customers, employees, and suppliers); ③ creative response to unprecedented problems; and ④ ethical value judgment (deciding, within a social context, what is right)

• Hypothesis: This is not a pessimistic story. It can also be read as the arrival of an era in which humans can focus on “the most human part” of themselves. However, that transition will require a fundamental redesign of education, vocational training, and the social safety net

8.4　Hypothesis ③ — “Management’s Sense of Time” Changes

Management today runs on a time axis set by accounting periods — “monthly, quarterly, annual.” If AI agents continuously process real-time data and keep updating forecasts, this sense of time may change.

• Hypothesis: The management style of “the monthly meeting looking back on last month’s results” disappears. In its place, “event-driven management” becomes mainstream — continuously monitoring how forecasts change this week, today, this very moment, and making decisions only at the instant a forecast changes significantly

• Hypothesis: The concept of the “annual budget” changes. If AI updates demand and financial forecasts weekly and monthly, “rolling forecasts” become the standard in place of a fixed annual budget. This resolves the management distortion where being locked into a budget prevents optimal decision-making

• Hypothesis: The speed of decision-making becomes the most important factor in competitive advantage. Based on the decision-ready material AI has assembled, “who can decide fastest and most accurately” becomes the core of competition between companies. “We don’t have enough information” stops being an excuse

8.5　Hypothesis ④ — The Structure of Inequality Changes

AI agents will not necessarily bring equal benefit to everyone. Rather, they carry the potential to create unprecedented gaps — between companies, people, and countries that can use them well and the infrastructure to support that, and those that cannot.

• Gaps between companies: Companies that adopt AI early and deeply get onto a positive feedback loop — “better data → more accurate forecasts → better decisions → more results → more data.” Companies that fall behind are left outside this loop, and recovery may become difficult

• Gaps between nations: AI capability depends on data volume, AI research strength, computing resources (GPUs/semiconductors), and legal-institutional design. The gap in economic, security, and diplomatic power may widen between countries that possess these in abundance (the United States, China, etc.) and those that do not. This is the background behind the urgency of national AI strategies in countries including Japan

• Gaps between individuals: A new axis of inequality may emerge — “people who can make use of AI” versus “people who are used by AI.” AI literacy, data thinking, and the ability to frame the right question become the most critical skills of the next generation, making the redesign of education urgent

8.6　Hypothesis ⑤ — The Definition of “Intelligence” Is Questioned Again

Here we present the most fundamental question as a hypothesis. Once AI agents begin to possess predictive, judgment, and creative abilities that surpass humans, the question “what is intelligence” becomes important not just philosophically but practically.

AI that surpassed humans at chess (Deep Blue, 1997) broke the definition that “chess proves intelligence.” AI that surpassed humans in medical diagnosis (from 2016 onward) shook the definition that “expert knowledge proves intelligence.” When AI appears that can craft management strategy better than humans, the definition that “strategic thinking is humanity’s advantage” will be called into question as well.

• Hypothesis: Once AI acquires “the ability to frame questions” (whenever that may be), we need to start thinking now about where the “sanctuary of uniquely human intelligence” will remain. If that sanctuary lies in “ethical judgment,” “empathy,” and “the power to create meaning,” then how society, education, and organizations cultivate these becomes the next most critical challenge

• Hypothesis: AI agents may come to behave like “another version of yourself.” Today’s personal AI still feels like an external tool — “asking ChatGPT” — but in the future it may become an extension of the self, where “my AI agent judges and acts on my behalf.” In a world where an AI attends meetings as your proxy, negotiates on your behalf, and keeps learning for you as “another version of you,” where does the boundary of “individual capability” lie?

8.7　What Remains Unchanged, Even So — The Fundamental Value of Being Human

Having laid out hypotheses of change up to this point, I want to close by writing about what does not change.

However good a prediction AI produces, “I trust you” is a human feeling. However efficient a solution AI produces, “there is meaning in pursuing this question” is a meaning that only humans assign. However elaborate a story AI can generate, “this person’s words are genuine” is a trust that arises only from human relationships.

Predictive management and AI agents deliver their greatest power when “humans can spend their time on what humans should be thinking about.” Creating an environment where executives and staff who had been drowning in a sea of routine work can finally face “the questions that truly deserve their attention” — that is the essential purpose of adopting AI agents.

Questions Worth Continuing to Ask in the Age of AI Agents

・Why are we in this business in the first place? (AI cannot decide the “for what purpose”) ・Who does this organization exist for? (Customers, employees, society, shareholders — the priority among them is not something AI should decide) ・Is the answer AI calls “optimal” truly the future we want? (Efficiency and values do not always align) ・If the ability to make skillful use of AI widens inequality, how do we respond? (Who benefits from technology is a choice society makes) ・How do we pass on to the next generation “the ability to ask the right question”? (In an age when AI produces the answers, the person of greatest value is the one who asks a good question) AI agents are a tool — and also a mirror. A mirror that reflects who we are and what we are aiming for, as the ones who use it. An organization that runs from that question will never reach predictive management, no matter how much AI it has. An organization that keeps holding onto that question will be the one that benefits most from AI agents.

End