企業 AI 要獲得自主權，必須先建立累積業務脈絡

Imagine it’s Monday morning, hours before the quarterly forecast meeting. A finance leader asks an AI agent, “Why did gross margin decline? Which assumptions need to change? And what should the company do next?” Within seconds, the agent pulls information from multiple systems, identifies a handful of drivers and recommends cutting discretionary spending or revisiting the hiring plan. It’s clear, detailed and plausible. Nothing about the answer looks obviously wrong.

That is what makes it so dangerous. 

Pulling back the curtain reveals several missing details. The agent used the wrong version of the forecast. It missed an approved accounting adjustment from the previous quarter. It failed to take into account that one region follows different business rules. None of the underlying data was necessarily wrong; the agent simply did not know how the organization interpreted that data or how the pieces fit together. The problem is not with the model that another one can easily solve. It’s that the model does not know this business. 

Much of enterprise AI development has concentrated on model capability and agent autonomy. But a model cannot arrive at a reliable, company-specific conclusion unless it understands how the organization defines its numbers and connects them to the way the business operates. 

The missing layer is cumulative business context, or the governed understanding of what enterprise data means: its definitions, relationships, assumptions, calculations, workflows, exceptions and the decisions that accumulate around it over time.

That context develops through planning, forecasting, scenario modeling, variance analysis, corrections, feedback, close activity and approved decisions. Without it, organizations risk giving AI more responsibility without giving it enough business understanding to carry that responsibility reliably. 

AI Capability is Outpacing Enterprise Readiness

According to Stanford’s 2026 AI Index, 88% of surveyed organizations used AI in 2025, yet agent deployment remained in the single digits across nearly all business functions. 

These statistics demonstrate that enterprises have access to AI capabilities, but many are not ready to let AI perform consequential work. Leaders know they need to build with AI, but they still need to determine whether they are building at the right layer of the problem.

A Model Can Read Data Without Understanding the Business

Foundation models are effective at pattern recognition, synthesis, natural-language interaction and exploring large datasets. While those are meaningful advances, the model itself is becoming less differentiating. 

The harder enterprise problem is teaching AI what a company’s information means and importantly how it is used in decision-making and analysis.

Much of the conversation surrounding AI governance has focused on clean data. While this is a necessary step, it’s insufficient when scaling AI for broad enterprise adoption. Two clean datasets can still use different definitions, time periods and business rules. 

IBM’s 2025 CEO Study found 25% of AI initiatives delivered expected ROI, and only 16% scaled enterprise wide. That pattern is consistent with a larger implementation challenge: access to models and data does not ensure that an AI system can reconcile the organizational rules that determine how information should be used.

More capable models can easily be procured. Company-specific understanding has to be built. 

Context at Retrieval is Not Cumulative Context

Much of what is described as context in enterprise AI is still supplied through retrieval at the moment of a request. The system finds a relevant document, record or data point it, uses it to generate an answer and begins again the next time it’s asked. 

Retrieval is valuable, but when used alone does not create cumulative organizational memory. It carries little forward, meaning the AI can encounter the business as a stranger in each new interaction.

It’s similar to onboarding a brilliant new analyst every morning. The analyst understands finance and knows how to interrogate a dataset, but they still need to learn how this particular company defines its metrics, which forecast is authoritative, why an exception was approved and which assumptions repeatedly prove wrong.

Cumulative context serves a different purpose. Rather than resetting with every prompt or planning cycle, context compounds over time. Each investigation, scenario, correction, approval and decision can contribute to a governed understanding of how the organization operates. This means later cycles can build on what earlier cycles established instead of rediscovering the same relationships, logic and ways of working with every prompt. 

That does not mean retaining everything indefinitely. Useful enterprise context must be deliberately scoped, correctable, permission-aware and traceable to its source. Organizations need to preserve what matters while ensuring that outdated assumptions, temporary exceptions and incorrect interpretations can be identified and changed over time.

Trust Operates Beneath the Model

The model is the most visible part of an AI system, but it’s rarely the part that determines whether an enterprise can trust the outcome. 

A capable model can investigate a variance, explore possible explanations and recommend an action. Whether an enterprise can rely on that recommendation depends on the architecture surrounding that model. 

Gartner found that organizations reporting successful AI initiatives invest up to four times more, as a percentage of revenue, in foundations such as data quality, governance, AI-ready talent and change management than organizations experiencing poor outcomes.

The governed context layer is part of that foundation, sitting between source systems, models and business actions. 

This is also where intelligence and determinism must diverge. AI can investigate, draft and recommend. Once an action has been reviewed and approved, execution should move through governed calculations, permission checks, write-back controls and audit trails. The flexibility that makes a model useful for exploration is not always appropriate when committing a change to the business.

A model can recommend what should happen. The surrounding architecture, including human approval and system controls, determines what is allowed to happen.

Constraints That Make Agents Useful

Agent deployment is happening faster than the controls around it. KPMG found that the share of organizations deploying at least some AI agents increased from 11% in the first quarter of 2025 to 42% by the third quarter. Yet Gartner reported that only 13% of organizations strongly agreed they had the right agent-governance structures in place.

That gap matters because capability and authority are separate design decisions. A system may be technically able to access a database, revise a forecast or trigger a workflow. That does not mean it should be permitted to do those things independently. 

Organizations should apply three constraints before delegating meaningful work.

• Establish Shared Meaning

Definitions, metrics, hierarchies, fiscal calendars and calculation rules must be reconciled before an agent can reason consistently across them. 

AI should not be expected to silently resolve disagreements the organization itself has not addressed. Without shared meaning, it may treat conflicting definitions as interchangeable and produce an answer that is coherent but operationally wrong.

• Decide What Carries Forward

Organizations must define which corrections, approvals, decisions and outcomes become reusable context. They also need policies governing retention, access and correction. 

Useful memory is selected and governed, not accumulated indiscriminately. Otherwise, an outdated assumption or one-off exception can quietly alter future recommendations.

• Separate Recommendation from Commitment

Analysis, a recommendation, a proposed change, human-approved execution and bounded autonomous execution represent different levels of authority. Agents should have identities and permissions appropriate to each level.

Evaluation must also cover the full chain of action: what information was retrieved, which context was used, which tools were called, what changed and whether that change can be reversed. 

Testing only the final answer misses most of the ways an agentic workflow can fail.

Capability describes what an agent can do, and governance defines what it’s allowed to do.

Autonomy Requires More Than Capability

In the opening margin scenario, speed is only useful if the system also knows which forecast is authoritative, which exception was approved, what action it may take and who remains accountable. 

I believe enterprise AI will earn trust only when it carries forward the right context and can show how it arrived at a consequential result. 

Organizations should build that context and control layer before expanding agent autonomy. The work will shift, but accountability will not.