AI Is Operating, Not Deciding

Let me start with how I actually use these tools, because I distrust essays about AI written at a distance from it. When I prepare a board discussion or assess an expansion case, the machine does extraordinary work: it reads more than I can read, holds more variables than I can hold, and produces a structured view of a complex situation in minutes rather than weeks. I would not want to work without it, and I say that as someone who has spent decades building judgements the slow way. But the moment the analysis lands on my desk, something changes hands. The machine has given me a picture. I must decide what the picture means, what it does not show, and what we will do about it. No part of that second step has ever been improved by pretending a machine can take it.

I draw the line at consequence. Where a decision’s consequence is reversible and modest, I am content for AI to operate almost unattended: scheduling, drafting first versions, reconciling figures, summarising documents. Where the consequence is material and hard to reverse, the machine prepares and the human decides, every time, without exception. This is not caution for its own sake. It is an honest account of what the tool is. A probabilistic system produces its most confident output precisely where the training data is thickest, which is precisely where everyone else’s system produces the same confident output. Competitive advantage has never lived in the thickest part of the data. It lives in the judgement applied at the end of it.

As an illustrative workshop, list recurring AI-assisted workflows and sort them by consequence, reversibility and review needs. The same tool may draft routine communications and a market-entry recommendation, but those uses deserve different controls. Making the distinctions visible helps leadership assign review effort and accountability where they matter.

In our advisory work we sort every recurring decision into three tiers, and I will share the method because it has served us well. Tier one is operational: high frequency, low individual consequence, fully reversible. Here AI operates, humans spot-check, and the organisation gains its speed. Tier two is analytical: medium frequency, meaningful consequence, partially reversible. Here AI prepares the analysis, a named human reviews it against the specific facts the model cannot know, and the decision carries a human signature. Tier three is consequential: low frequency, high consequence, hard to reverse. Investment commitments, leadership appointments, market entry, anything that would embarrass this firm if it were wrong. Here AI researches and challenges, and the decision belongs to people who will answer for it in person.

The method is deliberately boring, and that is its strength. The failure mode we see in client organisations is not that they use AI too much or too little. It is that no one has sorted the decisions, so the same tool that drafts the office newsletter also drafts the market-entry rationale, and the review effort, never designed, distributes itself randomly. Sorting the tiers takes a leadership team one honest afternoon. Operating inside them takes discipline, but the ordinary kind, the kind every organisation already knows how to apply to money. Apply the same discipline to decisions, and the AI question largely answers itself.

I want to be equally direct about the other side, because scepticism about AI deciding can curdle into refusal to use it, and that is its own kind of failure. The machine’s proper role in decision-making is not to decide but to discipline the decision process. It plays the opposing counsel, arguing the case against your preference with a stamina no colleague can match. It finds the counterexample you would have missed because you were tired. It challenges the consensus in the room, and as I argued in The Danger of Agreement earlier in this series, the danger of agreement is greater than the danger of error. A tool that disagrees with you on demand is the cheapest governance instrument ever invented, and most organisations use it almost exclusively to agree faster.

This is where the cognitive division of labour I have described becomes concrete. The machine holds the corpus, tests the pattern, and argues both sides. The human holds the accountability, the context the model cannot have, and the willingness to be wrong in a way a machine never is. Remove the machine and you are slower and less informed than your competitors. Remove the human and you are faster and more confident on the way to a decision that was never actually yours. The firms that understand this division will outperform both the refusers and the abdicators, and the margin will look, from the outside, like wisdom.

When boards ask me where to start, I give them one question to ask of every AI-assisted workflow in the building: if this output is wrong, who answers for it, and how quickly would we know. The question is diagnostic because it exposes, in one sentence, the workflows where AI is operating safely and the ones where it has quietly become deciding. Where the answer is a named person with a review habit, the line is intact. Where the answer is a shrug, or a junior employee assumed to have checked, the machine has already crossed it, and the organisation has not noticed.

Asking the question honestly is uncomfortable, because it usually reveals that the highest-stakes decisions in the building carry the least review. The market entry case was assembled at speed under deadline and read once. The analysis that justified the last acquisition went unchallenged because it arrived beautifully formatted. AI did not create this weakness. It industrialised it. Which is, in the end, the strongest argument for the posture I am describing: not that machines must not decide, but that organisations which never learned to own their decisions now get to learn at machine speed. The line between operating and deciding is not a limitation of the technology. It is the definition of the firm.

Sort your decisions into tiers. Let the machine operate the first tier with light human oversight, prepare the second with named human review, and serve the third as researcher and challenger to human deciders who accept the accountability. Use it, in every tier, to argue against you. And measure your adoption not by how much the machine produces but by how clearly your organisation can answer, for every consequential output, who decided and why.

This is the closing note of the series, and I will close it the way I close it in the boardroom. The technology will keep improving, and most of what is written this year about its limits will look timid in three years’ time. None of that changes the assignment. The machine is getting faster. The decision is still yours. Firms that keep those two facts in the same sentence will be fine. Firms that let either one displace the other will discover, at their leisure or their expense, which one was load-bearing.

The series began with a question about what AI can genuinely research about a market, the question its opening article put. It ends here, with the answer that matters for the decade: a very great deal, provided someone remains in the room who knows what research is for.

If your organisation wants the decision-tiering method applied to its actual workflow portfolio, our team runs the sorting workshop with leadership teams regularly.

Let us talk.

TAPiO Management Advisory is an international advisory firm headquartered in Kuala Lumpur, supporting companies, government agencies and institutions with market entry, investment facilitation and strategic projects across Malaysia and ASEAN.