Key takeaways

  • AI can access every project document and still use the wrong decision. A year of requirements, feedback, and meeting notes is a record of what happened—not necessarily an accurate account of what the business has approved today.
  • The cost is not just an incorrect answer. Outdated assumptions can travel from a product requirement into a launch plan, from a sales decision into a proposal, or from a customer commitment into execution. Faster output does not compensate for weaker decision quality.
  • The operating fix is decision ownership, supported by AI—not another document repository. AI can help detect changes, reconstruct rationale, surface conflicts, and assess downstream implications. Accountable people must resolve what is authoritative, and teams need that updated context wherever they work.

Access permissions determine what AI may read. They do not establish which business decisions are current, why they changed, or who owns them. That distinction may determine whether AI produces useful work—or more work that needs correcting.

01The problem we encountered with AI was not a lack of AI

In a recent engagement with a leading global strategy consulting firm, advanced AI tools were already available across a roughly 50-person organization developing and launching products for clients. Engineering and QA had become more efficient at execution. Product managers were handling more work. Project information was accessible through tools such as Coda, Jira, spreadsheets, and meeting notes.

The problem was not getting AI connected to the information. It was knowing what that information meant today.

Over about 12 months, client demos, market feedback, and senior-partner input had changed requirements and decisions. Some earlier decisions remained embedded in tickets and pages even after the business had moved on. A senior leader still used a detailed spreadsheet to track decisions and delivery activities; reconciling that view with downstream work consumed substantial effort.

An AI assistant could retrieve both the original requirement and the later discussion. What it could not reliably establish from retrieval alone was which one governed the next deliverable, who had authorized the change, or what else needed to change as a result.

We implemented reusable AI-assisted workflows that helped teams reconstruct decision histories, identify contradictory or potentially superseded information, and bring unresolved questions to accountable people. These workflows could carry clarified decisions and their rationale into subsequent work, rather than requiring every manager to independently reconstruct the same history.

Among the participating product and business teams, we observed greater decision consistency, less repeated review, and stronger visibility into evolving requirements. Product quality and delivery confidence also improved, although substantial human effort contributed and no controlled organization-wide impact study was conducted.

The important limitation was equally instructive: keeping approved working context continuously still required managers to do meaningful manual work.

Read the case study: Improving Decision Quality in AI-Enabled Product Teams.1

02Your document history is not your current operating context

Most enterprise AI discussions begin with a technology question: What systems should we connect, how much information can the model retrieve, and what permissions should it have?

Those questions matter. They are not sufficient.

The quality of an AI-assisted business deliverable depends in part on the assumptions and decisions that inform it. A decision made six months ago may have been entirely reasonable when approved. Following a customer demonstration or a shift in competitive positioning, it may no longer apply. The old document can remain factually accurate as a historical record while being wrong as guidance for today's work.

This is not necessarily a search failure, a data-security failure, or even a model failure. It is a failure to distinguish historical organizational knowledge from current business authority.

Research on enterprise retrieval provides a technical parallel. In an October 2026 paper, Garima Agrawal and colleagues describe a multi-agent development environment in which similarity-based retrieval returned a nearly identical rule from the wrong service. Their proposed knowledge layer organizes artifacts by producer, component, service, domain, and type before narrowing retrieval. The lesson is that finding similar content is not the same as finding applicable content. Importantly, the paper does not test whether AI can establish which evolving business decision has been approved or who owns it; those remain distinct operating requirements.2

Consider a sales proposal built from last quarter's positioning. An AI can produce a polished presentation, accurate citations, and persuasive prose—while missing the pricing or market decision the commercial leadership team made last Tuesday. The result may look productive until someone has to review and rebuild it.

The same failure can appear in product requirements, customer deployment plans, investment analyses, and operational decisions.

The next advantage is not giving AI access to everything. It is making the decisions that matter clearly enough for AI—and the people working with it—to apply consistently.

03Four questions every AI-assisted workflow should answer

Instead of treating project context as a pile of files, we believe leaders should manage it as an evolving set of business decisions. Four questions expose whether that discipline exists:

1 · What Is Current?

Which requirement, assumption, policy, or decision governs this work now? What has been explicitly superseded, and what remains unresolved?

2 · Why Did It Change?

What customer evidence, market development, technical finding, or leadership judgment altered the earlier direction? Without the rationale, the team can repeat the same argument or misapply the new decision.

3 · Who Owns It?

Who is accountable for the decision, who can approve a change, and who must act on it? Ownership must be grounded in the business domain. A product lead may own a product trade-off; a pricing leader may own commercial terms; a compliance owner may control a regulated policy.

4 · What Else Does It Affect?

Which deliverables, downstream decisions, customer commitments, and team workflows need to be revisited when the decision changes?

The fifth, practical step is to make unanswered questions visible to the right person—not have AI invent authority where none has been established.

This is a decision-quality discipline. It is not a proposal to replace human judgment or make every manager a records administrator.

04Why this belongs with business leaders, not only IT

IT and platform teams must govern data access, security, reliability, and integration. But they cannot be the sole arbiters of whether a sales promise, product priority, or client requirement is still the right business decision.

That responsibility already exists with the people running those functions. What is missing in many workflows is a low-friction way to express their decisions as current, usable context for both colleagues and AI tools.

The business case for better workflows is stronger than a technology-only argument. McKinsey's March 2025 State of AI survey found that workflow redesign had the largest estimated relationship with reported generative-AI EBIT impact among 25 attributes tested. Yet only 21% of respondents at organizations using generative AI said their organizations had fundamentally redesigned at least some workflows. These are self-reported, correlational survey findings—not evidence that context ownership by itself increases EBIT.3

A 2026 Harvard Business School working paper examining 202 AI-active U.S. enterprises also identified operating-model practices around shared information, coordination, feedback, and oversight, with higher behavioral scores associated with better operating performance. It does not directly test manager-approved context or establish causation.4

In a September 2026 Harvard Business Review essay, Gleb Tsipursky argued that middle managers are pivotal in translating AI initiatives into day-to-day work. His analysis reinforces the importance of the operating layer, although it does not demonstrate that any particular context-management mechanism works.5

That is why simply telling managers to maintain another knowledge base will not solve the problem.

AI must reduce the effort required to identify a proposed change, compare it with existing decisions, show potential downstream consequences, and prepare an informed human review. The accountable person confirms the decision. The updated working context then needs to reach the teams and AI workflows that rely on it.

This is the operating opportunity: centralize the current decision and its rationale without centralizing every human judgment or forcing everyone into the same tool.

05The difference between access and accountability

Enterprise platforms increasingly offer sophisticated permissions, retrieval, connectors, and governance features. Those capabilities remain necessary; there is no basis for assuming that every vendor has ignored context management.

The more useful question for a CEO or CPO is operational rather than competitive: even if the platform can support the right controls, has the organization defined who owns the current meaning of its information, how changes are approved, and how they propagate into work?

Access tells the system what it may use. Accountability tells the organization what it should stand behind.

Harvard Business Review authors Rohan Narayana Murty and Ravi Kumar S have argued that organizational context can become a differentiator as AI capabilities become more widely available. That is a useful strategic thesis, though it is not proof that any particular software or operating model solves the problem.6

For a business leader, the economic consequence is straightforward in principle. Work that is generated quickly but based on a superseded decision can require more review and rework, and may create quality or customer risk. The size of that cost needs to be measured in the relevant workflow, not assumed.

06A practical test before another major AI investment

Choose one decision-intensive workflow where business direction changes regularly: product requirements, proposal development, account planning, or customer deployment.

Take one material decision that changed recently. Then ask the team and its AI assistants to establish:

  • Which version of the decision is currently authoritative?
  • Why did it change, and who approved or owns it?
  • Which deliverables and teams should have been affected?
  • How quickly did the new decision reach the people and AI workflows using the old one?
  • How much time went into finding, reviewing, and correcting work that relied on superseded assumptions?

That small exercise often reveals the real gap more clearly than another inventory of AI licenses.

If the answer depends on one overloaded leader remembering the history, the organization has a continuity problem. If the relevant records disagree and no one can establish ownership, it has a decision-quality problem. If the approved change is known but cannot be propagated across workstreams, it has an operating-model problem.

The next step is to design a workflow in which AI prepares the evidence and change analysis, an accountable person resolves material uncertainty, and the current working decision becomes reusable across teams. Evaluate whether that process improves decision consistency, reduces re-review, and raises output quality without introducing excessive maintenance overhead.

07The question to take into your next leadership meeting

We are moving toward an environment where almost anyone can produce a sophisticated AI-assisted deliverable quickly. But the business still has to decide whether that deliverable is based on the right assumptions and current decisions.

The scarce resource is no longer just the capacity to produce. It is the organization's ability to maintain and apply sound judgment as circumstances change.

Before investing in more AI capability, ask one question:

When a consequential business decision changes today, who makes sure every team—and every AI workflow acting for that team—knows what changed, why, and what to do differently?

If the answer is unclear, more capable AI may simply produce more polished work based on yesterday's decisions.

That is the operating problem worth solving.

References

  1. StatsLateral — “Improving Decision Quality in AI-Enabled Product Teams,” case study.
  2. Garima Agrawal et al. — “Reimagining Enterprise RAG for Multi-Agent Systems: A Knowledge Layer That Organizes Before It Retrieves,” arXiv:2610.32851, submitted October 23, 2026. A case of service-level disambiguation in technical artifacts; does not establish the authority of changing human decisions.
  3. Alex Singla, Alexander Sukharevsky, Lareina Yee, Michael Chui, and Bryce Hall — “The State of AI: How Organizations Are Rewiring to Capture Value,” McKinsey & Company, March 12, 2025. Survey-based associations, not causal findings.
  4. Ellora Sarkar, Dilina Abudurexiti, Elisabeth Beltinger, Jamie Daubenspeck, and Marco Iansiti — “Organizing AI at Enterprise Scale: Agentic Operating Models and Firm Outcomes,” Harvard Business School Working Paper No. 27-038, 2026. Observational associations; does not directly test manager-approved context.
  5. Gleb Tsipursky — “Middle Managers Will Make or Break AI Adoption,” Harvard Business Review, September 1, 2026. Commentary on managers as an adoption and workflow-translation layer.
  6. Rohan Narayana Murty and Ravi Kumar S — “When Every Company Can Use the Same AI Models, Context Becomes a Competitive Advantage,” Harvard Business Review, February 18, 2026. Argument for organizational context as a competitive differentiator.

Frequently asked questions

Why does AI give wrong answers even when it has access to the right files?

Access permissions determine what AI may read, but they do not establish which business decisions are current, why they changed, or who owns them. A year of requirements, feedback, and meeting notes is a record of what happened — not necessarily an accurate account of what the business has approved today — so an AI assistant can retrieve both an original requirement and a later change without reliably knowing which one governs the next deliverable.

What four questions should every AI-assisted workflow answer?

What is current: which requirement, assumption, policy, or decision governs this work now? Why did it change: what evidence or judgment altered the earlier direction? Who owns it: who is accountable, who can approve a change, and who must act on it? And what else does it affect: which deliverables, downstream decisions, customer commitments, and workflows must be revisited? Where the answers are unclear, the practical step is to make the open question visible to the right person rather than have AI invent authority.

Who should own the context AI uses — IT or the business?

IT and platform teams must govern data access, reliability, and integration, but they cannot be the sole arbiters of whether a sales promise, product priority, or client requirement is still the right business decision. That responsibility already sits with the people running those functions. What is often missing is a low-friction way for them to express their decisions as current, usable context for colleagues and AI tools.

How can a company test whether it has a decision-ownership problem?

Choose one decision-intensive workflow where business direction changes regularly, take one material decision that changed recently, and ask the team and its AI assistants which version is currently authoritative, why it changed and who owns it, which deliverables and teams should have been affected, how quickly the new decision reached people still using the old one, and how much time went into correcting work that relied on superseded assumptions.