Decision Intelligence

AI Can Be Technically Right and Organizationally Wrong

Why consequential AI needs to understand not only what is happening in the business, but what the organization has decided to do about it.

By Chris Tambos · Published August 18, 2026

Executive Summary

Enterprise AI is becoming increasingly capable of understanding the state of a business: customers, products, revenue, margin, inventory, demand, capacity, forecasts, and operating performance. For many organizations, building the reliable, connected, governed data foundation required to support that capability is still unfinished work.

But as those foundations improve, another constraint emerges.

AI can know what is happening in the business without knowing what the organization has decided to do about it.

Accurate data and sound analysis can describe the consequences of different actions, but they do not always determine which outcome the organization should prefer. Leadership may already have made choices about strategy, customers, capacity, priorities, commitments, or accepted trade-offs that materially change the appropriate action.

The organizational problem itself is not new. What AI changes is the speed, scale, and authority with which decisions can now be recommended and increasingly acted upon.

The opportunity is therefore larger than better retrieval. Consequential AI increasingly needs two kinds of grounding: reliable business state and relevant organizational state — what has been decided, committed, prioritized, put into motion, and authorized. Meetings, documents, email, and other organizational knowledge remain important, but primarily as evidence that helps resolve, verify, explain, or challenge that state.

This becomes more consequential as AI moves from recommending actions to taking them. A system may be highly confident in its local analysis while lacking knowledge of a commitment, conflict, or unresolved decision elsewhere in the organization. Autonomy should therefore depend not only on analytical confidence, but also on whether the relevant organizational state, dependencies, boundaries, and authority are sufficiently resolved.

This is an emerging constraint, not a claim that organizational grounding is already the dominant enterprise AI problem. Many companies still have more fundamental data, governance, and production-AI work to do.

But leaders can begin preparing now: make current strategy and priorities explicit and available to enterprise AI, preserve consequential decisions, ground AI in sources the organization trusts, and build toward a decision environment that connects strategy, execution, authority, business state, and delegated AI action.

The long-term opportunity is an enterprise in which people and AI share a more reliable understanding of both what is happening in the business and what the organization has decided to do about it.

The argument in four claims

1. Accurate analysis does not always determine the preferred action.

Reliable enterprise data and sound analysis can show what is happening and what different actions are likely to produce. They do not always determine which outcome the organization should prefer.

The state of the business does not itself supply the organization’s preference.

2. AI can know the business without knowing what has been decided.

The organization may already have made a choice that materially changes the right action, while that choice remains distributed across meetings, plans, documents, systems, and working knowledge.

Giving AI access to the company’s knowledge is not the same as telling it what the company has decided and what is still in force.

3. The next opportunity is to make organizational state reliably usable by AI.

The goal is not simply to give AI more context. It is to help AI resolve current decisions, commitments, priorities, execution state, exceptions, and authority from the larger body of organizational information.

The destination is not simply more context. It is a better-resolved decision environment.

4. Organizational grounding becomes part of the autonomy boundary.

As AI moves from recommendation to action, locally correct decisions can conflict with commitments, constraints, unresolved trade-offs, or decisions elsewhere in the organization.

Autonomy should depend not only on confidence in the analysis, but also on the organizational scope and consequence of the action.

Figure 1 — The Decision Environment AI Needs.

The gap after trusted data

For decades, companies have invested in making the state of their business increasingly visible and machine-readable.

ERP systems capture transactions. CRM systems capture customers and commercial activity. Supply-chain and planning systems capture demand, inventory, capacity, and orders. Data warehouses and modern cloud platforms connect those systems. Semantic layers increasingly give the data consistent business meaning. AI can now reason over more of that information than ever before.

All of that matters. For many companies, getting this foundation right is still unfinished work. Fragmented data, inconsistent definitions, weak governance, and disconnected systems remain more immediate constraints than the problem this paper addresses.

But as those foundations improve, another question emerges:

Can AI know what is happening in the business without knowing what the organization has decided to do about it?

Consider the difference between two kinds of enterprise state.

The state of the business describes what is happening: customers, products, demand, revenue, margin, inventory, suppliers, capacity, forecasts, working capital, and operating performance.

The state of the organization describes the organizational reality surrounding those facts: what leadership is trying to accomplish, what has already been decided, what initiatives are underway, what commitments have been made, which trade-offs have been accepted, what assumptions have changed, and who has authority to act.

This is not a new organizational problem. Companies have always struggled with competing priorities, incomplete coordination, and locally rational decisions that conflict with broader commitments. AI does not create that problem. What changes is the speed, scale, and authority with which decisions can now be recommended and increasingly acted upon.

Those two states increasingly need to meet.

An AI system can have accurate data, perform sound analysis, and produce a rational recommendation while still missing organizational information that materially changes what the business should do.

That is what I mean when I say:

AI can be technically right and organizationally wrong.

The distinction begins with a deceptively simple question: even with perfect information about the business, do the facts always determine the right action?

1. Accurate analysis does not always determine the preferred action

Imagine a company making a pricing decision.

Its data is excellent. It can estimate price elasticity, volume, contribution margin, competitive response, and market-share effects with considerable accuracy.

At one price, margin rises but volume declines. At another, the company sacrifices some margin to gain share. A third supports a premium positioning strategy at the expense of near-term profit.

Which price is correct?

The analysis can describe the consequences of each choice. It cannot, by itself, determine which consequence the organization should prefer.

Herbert Simon made this distinction decades ago, separating factual premises from the value premises involved in choosing among alternatives.¹ For enterprise AI, the implication is simple: better facts and better analysis do not necessarily supply the organizational objective that makes one outcome preferable to another.

If leadership has already established that objective, along with the constraints and acceptable trade-offs, AI may be able to optimize the choice extremely well.

For example:

Maximize contribution margin while keeping volume loss below a defined threshold and maintaining our premium market position.

Now the organization has supplied information the pricing data itself did not contain.

This is an important boundary for the argument. Many decisions can and should be automated once objectives and constraints are sufficiently clear.

The narrower point is:

The state of the business does not always supply the organization’s preference.

And sometimes the organization has already made that choice.

The AI simply does not know it yet.

2. AI can know what is happening without knowing what has been decided

Consider a company evaluating one of its largest customer relationships.

The company has done the hard work on its data. Revenue is accurate. Customer and product hierarchies are connected. Cost-to-serve, rebates, returns, freight, support costs, and payment terms are reflected in a reliable view of profitability.

The analysis is clear: the customer is destroying economic value.

An AI system could reasonably recommend repricing the account, reducing its service level, or exiting the relationship.

Illustrative scenario

Suppose leadership has already reviewed the same economics and deliberately chosen to protect the relationship because of its strategic value or another commitment the company has made.

The numbers say the customer should go. The organization has already decided otherwise.

Better customer data does not solve this problem. It may make the wrong recommendation more persuasive.

That gives us a useful test.

If the missing information is another transaction, cost, customer attribute, or operating fact that should have been integrated into the analysis, this is still a data problem.

But if the business data is complete and the missing information is a decision the organization has already made, we are dealing with something different.

The system understands the state of the business. It does not reliably understand the relevant state of the organization.

That decision may exist in an account plan, presentation, meeting, email, or the working knowledge of the responsible leaders. AI will become increasingly capable of retrieving those sources.

But retrieval does not settle the question.

Suppose the system finds an older strategy saying the relationship should be protected, a newer presentation recommending reconsideration, and a later email saying no change should be made yet.

Which reflects the decision now in force? Which was only a proposal? Who had authority? Did one decision supersede another? What scope does it apply to?

These are not simply retrieval questions. They are questions about organizational state.

Giving AI access to the company’s knowledge is not the same as telling it what the company has decided and what is still in force.²

The obvious response is: if an important decision has been made, why not record it so the AI can use it?

That is exactly the right direction.

The challenge is not to give AI every artifact the organization has produced. It is to make the organizational state that matters to consequential decisions more reliable and usable.

That leads to the next question: what, exactly, should that state contain?

3. The next step is making organizational state reliably usable

For AI, the decision environment has two grounding layers: the state of the business and the current state of the organization. Organizational knowledge helps maintain and explain the latter.

The state of the business includes customers, products, orders, inventory, capacity, costs, revenue, margin, forecasts, and other operating facts represented in enterprise systems.

The current state of the organization includes the relevant decisions, commitments, priorities, initiatives, exceptions, authority, and execution state that the organization currently treats as governing action.³

Meetings, messages, plans, documents, and other organizational knowledge remain important, but primarily as evidence from which organizational state can be resolved, verified, explained, or challenged.

The opportunity for enterprise AI is not merely to retrieve more organizational knowledge.

It is to resolve enough current organizational state that AI can reason over it alongside the state of the business.

Consider a product launch.

Leadership has committed to a date. Capacity has been protected. Resources have been assigned. A commercial plan is underway. Other work may have been deferred because the launch has priority.

No single system may contain that full reality. A roadmap may hold the date. A project system may show the work underway. A resource plan may reflect staffing. A steering meeting may contain the trade-offs leadership accepted.

An AI system capable of reading all of those artifacts would possess a great deal of organizational knowledge. What it needs for consequential reasoning is something more resolved: the commitment, its scope, its owner, the authority behind it, the resources it constrains, and the conditions under which it could change.

Not every organizational question will have a single resolved answer. Priorities can conflict, decisions can remain open, and legitimate ambiguity can persist. Where organizational state is genuinely unresolved or contested, the objective should not be to manufacture certainty. It should be to make that uncertainty visible.

This is also why high-level strategy, while valuable, is not enough.

In enterprise AI work, I have seen the decision environment improve when AI was given corporate priorities, strategic initiatives, and leadership principles. That context helped, but it remained too coarse to answer more specific questions about what had actually been decided, what work was already underway, what resources had been committed, and which trade-offs leadership had accepted.

Strategy provides direction. Decisions and execution state show how that direction is currently being translated into action.

And those decisions do not exist independently.

A product-launch commitment can affect capacity allocation. That can affect production and sourcing. Those choices can affect inventory and working capital. A commercial commitment can change service decisions. A channel decision can change which products should receive inventory, promotion, or support.

Organizational decisions propagate. They change the conditions under which other decisions should be made.

This does not mean everything in an enterprise must be connected to everything else, or that AI needs a perfect model of every downstream consequence. Most decisions can remain bounded.

What matters is whether the system can recognize the material dependencies that change the appropriateness of a recommendation or action.⁴

These capabilities begin to suggest what a corporate second brain could become.

Not simply a repository that remembers what the organization has said, but a capability that helps maintain a reliable understanding of what the organization currently knows, has decided, committed to, prioritized, and put into motion, together with the material relationships among those choices.

How that capability is built will almost certainly change. It may draw from strategy and execution systems, meetings, planning tools, enterprise data, deliberately captured decisions, and technologies still emerging. Over time, AI itself may do much of the work of identifying decisions, detecting conflicts, recognizing supersession, and maintaining organizational state.⁵

The architecture is not the important part of the argument.

The requirement is.

The destination is not simply more context. It is a better-resolved decision environment.

That distinction becomes much more consequential when AI stops merely recommending what the organization should do and begins acting on its behalf.

4. When recommendation becomes action

When AI recommends, missing organizational context can be a relevance problem. When AI acts, it can become a control problem.

A recommendation that ignores a commitment or priority may waste attention or be rejected by an experienced leader. An autonomous system can go further: changing a price, adjusting an order, reallocating inventory, issuing a customer concession, modifying a schedule, or triggering another business process.

Companies have always dealt with locally rational decisions that conflict with broader priorities or commitments. AI does not create that interdependence. It can operationalize it at much greater speed, scale, and authority.

That makes an important distinction increasingly relevant:

analytical uncertainty and organizational uncertainty are not the same thing.

Analytical uncertainty asks whether the system understands the business problem well enough.

How reliable is the demand forecast? How confident is the price-elasticity estimate? How complete is the customer data?

Organizational uncertainty asks something different.

Is another commitment already in force? Is an initiative underway that changes the appropriate action? Does an exception apply? Has another function been promised the same constrained resource? Has authority for this trade-off actually been delegated?

An AI system can have very low analytical uncertainty and still face material organizational uncertainty.

And sometimes that uncertainty reflects the organization itself: a trade-off has not yet been resolved, priorities genuinely conflict, or the authority to decide is unclear.

High confidence in the analysis does not automatically justify autonomous action.

Illustrative scenario

Consider a promotional-pricing system.

The system identifies a product as an excellent candidate for promotion. It has reliable elasticity estimates, accurate margin economics, good inventory visibility, and a strong forecast of promotional lift. Within the problem it has been asked to solve, the recommendation is excellent.

But the product does not exist in isolation.

It may be scheduled for removal from one sales channel. Manufacturing capacity may have been protected for another launch. Finance may be working to reduce inventory and working capital. A service commitment elsewhere may depend on the same constrained supply.

None of those decisions necessarily makes the promotion wrong.

A promotion could help accelerate inventory through a channel the company is exiting. Or it could stimulate demand that requires replenishment into a channel leadership has already decided to leave. It could produce attractive promotional economics while consuming capacity the organization has committed elsewhere.

The point is that several locally rational decisions can interact.⁶

A locally correct AI action can still be organizationally wrong when it crosses decisions, commitments, or constraints the system has not adequately accounted for.

The answer is not to require a person to approve every AI action.

The better operating principle is graduated autonomy.⁷

If the promotion system is operating on a product with ample inventory and capacity, no active channel transition, no strategic exception, and no conflicting commitment, and the price change is within delegated financial limits, the system should act.

If the same action touches a product undergoing a channel transition, requires capacity protected for another commitment, or exceeds delegated authority, the quality of the promotional forecast may not have changed at all.

What changed is the organizational scope of the decision.

That is a reason to route the decision to an authority capable of resolving the broader trade-off.

Enterprise AI does not need to understand every initiative or downstream consequence before it can act. Most decisions can remain bounded.

The relevant question is whether an action remains inside a decision environment where the objective, organizational state, material dependencies, constraints, and authority are sufficiently resolved for the consequence at stake.

Where those conditions are met, automate aggressively.

Where an action crosses a material commitment, conflict, exception, or authority boundary, or where the relevant organizational state remains materially unresolved, route it.

This is also why simply giving an AI system permission to use a tool is not enough.

Permission answers whether the system may perform an action. Organizational grounding helps determine whether that action is appropriate under the organization’s current commitments and decisions.

Autonomy should depend not only on confidence in the local analysis, but also on the organizational scope and consequence of the action.

The objective is not more human intervention. It is a decision environment strong enough to support more autonomy where autonomy is warranted.

What this argument does not claim

The argument in this paper is intentionally bounded.

It is not a claim that enterprise data problems are solved, or that organizational state is already the dominant enterprise AI constraint. For many organizations, fragmented systems, inconsistent definitions, weak governance, incomplete integration, and moving AI into production remain more immediate problems. Organizational grounding becomes more important as those foundations mature and AI takes on more consequential work.⁸

It is also not an argument for simply giving AI more information. More documents, longer context windows, and broader access do not automatically create a better decision environment. Relevance, currency, authority, and scope matter.⁹

Nor does consequential AI require a complete model of the enterprise. Most decisions can remain bounded. The requirement is to recognize the material dependencies and boundary crossings that can change whether an action is appropriate.

This is not an argument for universal human approval or permanent manual maintenance. Where objectives, relevant state, constraints, and authority are sufficiently resolved, AI should increasingly be able to act. AI should also become increasingly capable of helping identify decisions, reconcile information, detect changed assumptions, and maintain organizational state.

And this paper does not prescribe a technical architecture. Semantic layers, enterprise search, knowledge graphs, strategy and execution systems, decision records, and AI agents may all play a role. Formal causal models, decision graphs, and other implementation approaches may become useful in particular settings, but they are not required for the argument here.

The underlying requirement is simpler:

Consequential AI needs a decision environment rich enough to understand both what is happening in the business and the organizational reality that can change what should be done about it.

There is also an evidence limitation. Direct empirical evidence of production AI systems causing measurable enterprise harm specifically because they lacked organizational state is still limited. The argument therefore rests on established research about decision-making, organizational interdependence, coordination, and local versus system-level optimization, together with emerging evidence from increasingly autonomous AI systems.

This should be understood as a forward-looking operating principle, not a claim that the failure mode is already widespread or fully measured.

What leaders should do now

The implications of this argument do not begin with a new AI platform.

They begin with a simpler question:

What does your enterprise AI know about how your company is actually trying to operate?

Most organizations already have much of the relevant information somewhere. The problem is that priorities, initiatives, authority, decisions, and supporting knowledge are unevenly current, unevenly authoritative, and rarely assembled into a reliable decision environment.¹⁰

One prerequisite runs through everything that follows:

A system of record cannot create clarity the organization itself never established.

If leadership is unclear about strategy, priorities, decisions, ownership, or authority, technology cannot manufacture that clarity.

But leaders do not need to solve the entire future architecture today. Start with three practical moves now, while asking your technology and data leaders to build toward two broader capabilities.

Start with three things now

1. Give your enterprise AI the company’s current strategy and priorities

Start with the most basic test:

If an employee asks your enterprise AI, “What are our company’s priorities right now?”, where does the answer come from?

If the answer is an old presentation, a mixture of emails, or whatever the AI happens to retrieve, that is a problem worth fixing.

Identify a governed, current source for the organization’s strategy, major priorities, ownership, and the initiatives that matter most. It does not have to be an OKR platform. It may be a strategy-execution system, a portfolio environment, a governed knowledge site, or another system the organization already uses.

Leadership should be able to say:

This is where the current strategy and priorities live, and this is what we expect both people and AI to trust.

Where priorities are clear, make them explicit, current, and available to the enterprise AI environment rather than forcing every employee to reconstruct them independently.

Leading enterprise platforms are beginning to support this pattern through governed shared sources and scoped grounding. The implementation will vary by environment, but the management principle is durable: common organizational context should increasingly be provided centrally rather than rebuilt employee by employee.¹¹

2. Preserve the consequential decisions that change what the business should do

Organizations are usually much better at recording activity than recording what that activity actually resolved.

They have systems for transactions, customers, projects, objectives, policies, and workflows. But an important management decision can still disappear into a meeting, email thread, presentation, or someone’s memory.

Do not document every decision.

Capture the consequential few: decisions that change priorities, commit meaningful resources, establish an important constraint, create an exception, affect multiple functions, or materially change how future decisions should be made.

A lightweight record can include:

what was decided;
who had authority;
the rationale and important trade-offs;
the scope and effective date;
what it supersedes;
the assumptions behind it;
when it should be reconsidered.

This is one of the least mature parts of the current enterprise landscape. Strategy systems can capture intent, portfolio systems can capture work underway, and policies and workflows can capture rules and authority, but consequential management decisions and their current standing still often fall between systems.¹²

The goal is not bureaucracy. A useful decision record should take minutes, not create another reporting process. AI can increasingly help draft the record from meetings and communications; for consequential decisions, a responsible person confirms what was actually decided.

If a decision will change how other people or systems should act, preserve it somewhere they can reliably find it.

3. Set the grounding rule: consequential AI should rely on authoritative sources

Once strategy and important decisions have trusted homes, generalize the rule.

Giving AI broad access to organizational information is valuable. But broad access is not the same thing as reliable grounding.

Meetings, email, documents, chat, and presentations can contain important evidence and rationale. They can also contain drafts, obsolete assumptions, conflicting proposals, and decisions that were never ratified.

Ask your CIO, CDO, or AI leader:

What sources ground our enterprise AI today? Which are authoritative? What happens when an old deck conflicts with the current decision?

Where the company has a governed source for strategy, an active initiative, a policy, an approval boundary, or a consequential decision, that source should carry more weight than whatever artifact happens to be retrieved first.

The operating principle should increasingly be:

Authoritative organizational state tells AI what is currently in force. Organizational knowledge helps explain, verify, challenge, and update that state.

Current enterprise platforms can increasingly prioritize trusted sources and support curated shared grounding, but they do not remove the organization’s responsibility for currency, supersession, conflict resolution, and stewardship.¹¹

Then build toward two broader capabilities

4. Connect the broader decision environment

Do not start by trying to create one monolithic system. Different systems already hold different parts of the decision environment.

Strategy systems describe intended direction. Portfolio and work-management systems show what is actually underway. Policies, workflows, and delegation-of-authority structures define important boundaries. Decision records preserve consequential choices. ERP, CRM, finance, supply-chain, planning, and other operating systems describe the state of the business.¹³

The longer-term question for the CIO, CDO, or CAIO is:

Can our AI decision environment reason across the relevant parts of those systems without forcing people to reconstruct the organizational context every time?

Use the systems already in place first. Improve their currency and authority. Connect them where value justifies it. Fill the gaps deliberately.

5. Define organizational boundaries before giving AI more authority

As AI moves from answering questions to taking actions, the governance question changes.

It is no longer enough to ask:

Can the agent technically perform this action?

Leadership needs to ask:

Under what organizational conditions should it be allowed to act?

Which actions can AI take autonomously? Within what financial, operational, or customer thresholds? Which policies, commitments, and active decisions must it respect? What requires approval? What happens when priorities conflict or the relevant organizational state is unresolved? Who receives the decision when the system routes it?

The goal is not more approval gates.

It is to create a sufficiently clear decision environment that AI can act more autonomously where the objective, organizational state, constraints, and authority are resolved, and escalate when they are not.¹⁴

Together, these actions provide a practical path toward the corporate second brain introduced earlier: shared organizational context maintained well enough that people and AI do not have to reconstruct the company from scattered artifacts every time they work.

The organization should increasingly provide the common context.

People should add the context unique to their role, their work, and the decision in front of them.

The practical sequence is straightforward:

Make the strategy explicit. Preserve the decisions that matter. Ground AI in sources you trust. Then connect the broader decision environment and define the boundaries for autonomous action.

That improves how the organization operates today while preparing it for AI systems that will increasingly recommend, decide, and act on its behalf.

Conclusion: From knowing the business to understanding the organization

Organizations have spent decades making the state of the business increasingly visible, connected, and machine-readable. That work remains essential.

But as AI moves deeper into consequential decisions, reliable business data will not always be sufficient. Organizations also make commitments, establish priorities, accept trade-offs, launch initiatives, and delegate authority. Those choices can change the appropriate action even when the underlying business facts do not.

That is why AI can be technically right and organizationally wrong.

The opportunity is not to surround AI with more approval gates or to give it every artifact the company has produced. It is to build a better decision environment in which AI can reason over reliable business state, relevant organizational state, and the material dependencies between them.

As that environment improves, the result should be greater autonomy, not less.

We spent years making the business more legible to machines.

The next step is making more of the organization legible too.

The companies that do both well will be better positioned to turn increasingly capable AI into better decisions, better actions, and measurable business outcomes.

Evidence and methodology note

This Perspective began with operating observations from enterprise data and AI work and was then tested against research in decision science, organization theory, coordination, automation, and current enterprise AI. The two business examples labeled Illustrative scenario are constructed examples designed to test the argument; they are not presented as documented production-AI incidents or client case studies.

The evidence base is intentionally mixed. Established research is used to support durable propositions such as the distinction between facts and preferences, the separation of information from decision authority, organizational interdependence, and graduated autonomy. Current AI research and product documentation are used more narrowly to test how those propositions intersect with today’s technology. Direct empirical evidence of production AI causing measurable enterprise harm specifically because it lacked organizational state remains limited; where the paper extends established organizational phenomena to increasingly autonomous AI, it should be read as a reasoned operating implication rather than a measured prevalence claim.

Current enterprise-platform capabilities referenced in the recommendations were verified against official product documentation in August 2026. Product details will change; the underlying recommendation is architecture-neutral.

Endnotes

Herbert A. Simon, Administrative Behavior: A Study of Decision-Making Processes in Administrative Organizations, 4th ed. (New York: Free Press, 1997; originally published 1947). Simon's treatment of factual and value premises is the foundational support for the narrower claim used here: facts can constrain and inform a decision without themselves supplying the objective or preference that selects among outcomes.
Philippe Aghion and Jean Tirole, “Formal and Real Authority in Organizations,” Journal of Political Economy 105, no. 1 (1997): 1–29; James P. Walsh and Gerardo R. Ungson, “Organizational Memory,” Academy of Management Review 16, no. 1 (1991): 57–91. Together these support two elements of the argument: authority is distinct from possession of information, and organizational knowledge is distributed across people, structures, routines, and archives rather than residing in a single artifact.
Richard M. Cyert and James G. March, A Behavioral Theory of the Firm (Englewood Cliffs, NJ: Prentice-Hall, 1963); Michael D. Cohen, James G. March, and Johan P. Olsen, “A Garbage Can Model of Organizational Choice,” Administrative Science Quarterly 17, no. 1 (1972): 1–25. These works support the distinction between statements, preferences, and organizational decisions produced through institutional processes. They are not evidence that every decision must be formally recorded; they support the narrower point that an authoritative organizational decision is not simply a property of the underlying facts or text.
James D. Thompson, Organizations in Action: Social Science Bases of Administrative Theory (New York: McGraw-Hill, 1967); Thomas W. Malone and Kevin Crowston, “The Interdisciplinary Study of Coordination,” ACM Computing Surveys 26, no. 1 (1994): 87–119; Herbert A. Simon, “The Architecture of Complexity,” Proceedings of the American Philosophical Society 106, no. 6 (1962): 467–482. Thompson and Malone/Crowston establish the importance of interdependence and dependency management; Simon's near-decomposability helps bound the argument by explaining why most activity can remain local even though selected cross-boundary couplings still matter.
Hengyu Liu, Tianyi Li, Zhihong Cui, Yushuai Li, Zhangkai Wu, Torben Bach Pedersen, Kristian Torp, and Christian S. Jensen, “Reliable AI Needs to Externalize Implicit Knowledge: A Human–AI Collaboration Perspective,” Proceedings of the 43rd International Conference on Machine Learning, PMLR 306 (2026); Pei-Yun S. Hsueh and Johanna D. Moore, “Automatic Decision Detection in Meeting Speech,” in Machine Learning for Multimodal Interaction (2007), 168–179. Liu et al. is a position paper, not an empirical demonstration; it supports the argument for externalizing otherwise unverifiable knowledge. Decision-detection research supports the narrower current capability claim that AI can help identify candidate decisions from organizational communications. The stronger claim that AI can autonomously maintain authoritative organizational state remains emerging.
Hau L. Lee, V. Padmanabhan, and Seungjin Whang, “Information Distortion in a Supply Chain: The Bullwhip Effect,” Management Science 43, no. 4 (1997): 546–558; Thompson, Organizations in Action. The bullwhip literature is a canonical example of locally rational information processing producing system-level instability when interdependencies are not adequately coordinated.
Raja Parasuraman, Thomas B. Sheridan, and Christopher D. Wickens, “A Model for Types and Levels of Human Interaction with Automation,” IEEE Transactions on Systems, Man, and Cybernetics—Part A: Systems and Humans 30, no. 3 (2000): 286–297. The levels-of-automation literature supports selective and graduated allocation of functions between people and automated systems rather than a binary choice between full autonomy and universal human approval. The paper's specific organizational-state triggers are a reasoned application of that broader principle.
Deloitte, Tech Trends 2026, “The Agentic Reality Check” (2026), reporting that agentic AI adoption remains substantially concentrated in exploration and pilot stages rather than broad production deployment. This supports the paper's maturity framing that the organizational-state/control problem is an emerging constraint for many companies rather than today's dominant AI problem everywhere. The precise maturity of any individual organization will vary considerably by industry and use case.
Nelson F. Liu, Kevin Lin, John Hewitt, Ashwin Paranjape, Michele Bevilacqua, Fabio Petroni, and Percy Liang, “Lost in the Middle: How Language Models Use Long Contexts,” Transactions of the Association for Computational Linguistics 12 (2024): 157–173. The study demonstrates that larger context availability does not guarantee reliable use of relevant information. This supports the paper's narrow claim that “more context” is not itself the target; it does not establish the stronger institutional-state argument, which rests on the organizational research cited above.
Gary L. Neilson, Karla L. Martin, and Elizabeth Powers, “The Secrets to Successful Strategy Execution,” Harvard Business Review (June 2008); Donald Sull, Rebecca Homkes, and Charles Sull, “Why Strategy Execution Unravels—and What to Do About It,” Harvard Business Review (March 2015); Donald Sull, Charles Sull, and James Yoder, “No One Knows Your Strategy—Not Even Your Top Leaders,” MIT Sloan Management Review (2018). These studies and practitioner analyses document persistent problems with decision-rights clarity, information flow, cross-unit coordination, and propagation of strategy into execution. They support the recommendation to strengthen the management process before treating technology as the solution.
Microsoft, “Microsoft 365 Copilot Architecture and How It Works,” “Data, Privacy, and Security for Microsoft 365 Copilot,” and “Knowledge Sources Summary — Microsoft Copilot Studio” (official documentation, verified August 2026); Salesforce, “Data 360” and Agentforce/Trust Layer documentation (verified August 2026); ServiceNow, “AI Control Tower” and ServiceNow AI Platform documentation (verified August 2026); Glean, “Enterprise Graph” and connector documentation (verified August 2026). These sources establish that leading enterprise AI environments can ground assistants or agents in centrally connected organizational information while applying enterprise permissions and governance. They do not establish that those platforms automatically resolve authoritative organizational state such as ratification, supersession, decision rights, or current standing.
Michael Nygard, “Documenting Architecture Decisions” (2011), which popularized the lightweight Architecture Decision Record pattern; Pei-Yun S. Hsueh and Johanna D. Moore, “Automatic Decision Detection in Meeting Speech”; Matthew Purver, Patrick Ehlen, and John Niekrasz, “Detecting Action Items in Multi-Party Meetings: Annotation and Initial Experiments,” in Machine Learning for Multimodal Interaction (2006), 200–211. The ADR pattern is used here as a transferable example of low-burden decision capture, not as evidence that the same software-engineering template should be imposed on executive decisions. Meeting-analysis research supports AI-assisted drafting, not autonomous ratification of consequential decisions. The broader conclusion that consequential management decisions frequently lack a dedicated enterprise system of record is based on the author's August 2026 review of current decision, governance, strategy-execution, and portfolio-management practices.
Author's August 2026 review of current documentation for leading strategy-execution, OKR, portfolio, project-management, workflow, and governance systems, alongside Paul Rogers and Marcia Blenko, “Who Has the D? How Clear Decision Roles Enhance Organizational Performance,” Harvard Business Review (January 2006). The reviewed systems commonly represent different portions of organizational state — objectives and priorities, initiatives and status, policies and approvals, or delegated authority — while decision rationale, accepted trade-offs, supersession, and current standing are less consistently represented as first-class structured objects.
National Institute of Standards and Technology, Artificial Intelligence Risk Management Framework (AI RMF 1.0), NIST AI 100-1 (2023), especially GOVERN 2.1, GOVERN 2.3, and GOVERN 3.2 on documented responsibilities, executive accountability, and differentiated human-AI roles and oversight. NIST is voluntary guidance; it is cited here as authoritative governance practice, not as a legal requirement.

About the author

Chris Tambos is the founder of Rhiza Advisory, an operational AI leadership and advisory practice. Over three decades, he has built and led data, analytics, and AI capabilities across financial services, technology, and complex operating businesses, with a focus on turning data and AI into better decisions and measurable business outcomes. Trusted Reasoning is his body of work on how organizations build AI-enabled decision systems leaders can rely on.

rhizaadvisory.com/trusted-reasoning · linkedin.com/in/christambos

Suggested citation: Tambos, C. (2026). AI Can Be Technically Right and Organizationally Wrong. Trusted Reasoning.