When AI Can Produce the Answer, Your Value Is Knowing What to Question

Future Compatible AI Series | Article 6

A recommendation can be ready for approval before anyone has agreed what a good decision would look like.

The analysis is coherent. The presentation is persuasive. The meeting is nearly over. Someone asks whether there are any objections, and silence does the work of agreement. An answer has acquired authority without anyone examining the question that produced it.

For me, this is a central challenge of AI leadership: preserving the quality of judgement when producing a convincing recommendation becomes easier. AI can help us think, question and decide. The leadership task is to establish which conclusions deserve action, under what conditions, and with whose accountability.

In the original five articles of the Future Compatible AI Series, I explored how AI changes decision-making, workplace culture, professional capability, vendor economics and organisation design. I am extending that discussion with Article 6 because the practical question remains: how does a leader exercise the judgement those changes demand?

Microsoft’s India Work Trend Index findings, released on 3 September 2026, offer a timely signal. Among surveyed Indian AI users, 63% identified quality control of AI output as a skill becoming more important, and 59% identified critical thinking. These are perceptions within that sample, but they point towards a capability worth taking seriously.

Why a good answer can end the thinking too early

An answer brings relief. It gives an uncertain conversation somewhere to land and a busy manager something concrete to move forward. When deadlines are tight, questioning a polished recommendation can feel like reopening work everyone hoped was finished. The person raising a concern may also be challenging a senior colleague’s preferred direction.

The cost of a question is visible in the meeting; the cost of not asking it usually arrives later.

There is relevant evidence behind this concern. In a 2025 study of AI and critical thinking, Hao-Ping Lee and colleagues surveyed 319 knowledge workers. Greater confidence in generative AI was associated with less reported critical thinking. The study examined self-reported behaviour; it did not demonstrate that AI causes a lasting loss of thinking ability. It also described thinking shifting towards checking information and integrating responses into work.

My concern is where that saved effort goes. If it creates room to examine assumptions, the decision can improve. If it simply permits more recommendations to pass through the same crowded approval meeting, the organisation may have increased throughput while weakening scrutiny.

Two habits deserve attention. The first is approval without challenge: a named reviewer signs off but lacks the time, context or authority to alter the recommendation. The second is automated reassurance: the team asks AI to critique its own proposal, receives a tidy risk list, and treats the presence of that list as evidence that the risks have been examined.

AI-assisted challenge can be useful. But when it starts from the same incomplete information and inherited objective, it can leave the most consequential omission untouched. A human approval is only meaningful when the human can change the decision.

The 2023 experiment by Fabrizio Dell’Acqua and colleagues illustrates why scrutiny must depend on the task. Across 758 consultants, GPT-4 improved performance on tasks within its capabilities. On a task chosen to sit outside that boundary, AI users were 19 percentage points less likely to reach the correct solution. The study concerned that model and those tasks, rather than a fixed limit on today’s AI.

For leaders, the implication is to test reliability where the decision will be used. Impressive performance elsewhere provides insufficient grounds for handing over this particular judgement.

Following the question through the decision

I would use a simple Decision Questioning Loop: purpose, evidence, consequences and ownership. It is a practical synthesis for leadership conversations, with each stage testing what the previous stage might have missed.

Purpose: What outcome are we trying to achieve, and what must we protect while achieving it? The objective determines which evidence matters. A request to reduce service costs produces a different analysis from a request to reduce avoidable demand while maintaining access for customers who need help.

Evidence: What supports the recommendation, what is missing, and what finding would make us change our minds? Test the assumptions that could reverse the decision. Expertise matters here: recognising an implausible estimate or an absent variable requires knowledge of the work.

Consequences: If the recommendation succeeds on its chosen measure, what else could deteriorate? Trace the effects across teams, customers and time. A saving in one budget may become rework, delay or lost capability somewhere else. Those consequences may require us to revisit the original purpose.

Ownership: Who can authorise, challenge, pause and revise the decision? Specify what will be monitored and when the decision will be reviewed. Observed outcomes feed back into the next statement of purpose, so the organisation learns from what it actually did.

The loop should be proportionate. A reversible wording change needs little ceremony. A decision affecting people’s roles, customer access or a major commitment deserves deeper examination. Its value lies in what it changes about the decision.

A recommendation that solves the wrong problem

Consider a hypothetical service team under pressure to reduce costs. An AI-assisted analysis recommends moving more enquiries to self-service and reducing staffed support. It draws on ticket volumes, handling times and the cost of each channel. The projected saving looks attractive.

The manager starts with purpose. Is the organisation trying to reduce contact-centre expenditure, or reduce the total cost of helping customers resolve their problems? The two objectives could lead to different choices.

The evidence review then reveals a gap: the ticket data contains completed contacts but says little about customers who abandoned the process. It also misses work absorbed by account managers after an enquiry was supposedly resolved. These are gaps in this hypothetical analysis, not inherent limitations of every AI system.

Tracing consequences brings those account managers and frontline staff into the discussion. Their knowledge suggests testing self-service for a narrower class of enquiries first. The decision owner sets a review date and agrees to monitor repeat contacts, successful resolution and transferred workload alongside cost.

AI remains useful throughout: analysing cases, identifying patterns and helping design the pilot. The manager’s contribution is to change what counts as success, bring missing evidence into view and make the decision revisable. That is a tangible form of professional value.

Making questioning part of the work

For AI leadership to become an organisational capability, this discipline needs a place in ordinary routines. I would begin with four changes.

Begin with a decision brief. Replace the default request for a finished recommendation with a short statement of the outcome, constraints and evidence that could change the choice. Use it before analysis begins. Look for fewer late disputes about objectives. Keep it short enough to use; a lengthy template can become another reporting obligation.

Give challenge a route into the decision. Instead of inviting objections at the end, ask someone close to the consequences to review the assumptions before approval, and record the response to material concerns. Progress means useful challenges change plans. If concerns are acknowledged but routinely overruled without explanation, the ritual will teach people to stay quiet.

Preserve some independent reasoning. For selected consequential decisions, ask the responsible person to write a provisional view before seeing the AI recommendation, then explain any change. Look for clearer reasoning about disagreement. Avoid turning this into an AI-free purity test: the purpose is to examine reliance and develop judgement, not duplicate every task.

Review assumptions after action. Move beyond recording who approved the proposal. Name an owner, a review point and a condition that would trigger reconsideration. Look for decisions revised when evidence changes. If review becomes a search for someone to blame, people will learn to defend the original choice and conceal uncertainty.

These practices give the middle manager a contribution that can be described in business terms: a misleading measure corrected, a costly dependency exposed, a customer consequence recognised before rollout. They also make leadership development part of delivery, because people explain choices and learn what followed.

That is where I would take the argument from Part 3. Developing judgement requires opportunities to use it, permission to challenge and feedback about consequences. A leadership development programme can begin the work; the surrounding management system determines whether the capability gets exercised.

There is a personal adjustment here too. A leader accustomed to being sought out for answers may have to become comfortable revealing uncertainty. The useful contribution might be an admission that the evidence is incomplete, followed by a precise way to resolve what matters. Endless doubt and performative questioning add little.

Your value is in the difference your judgement makes to what happens next.

When the next compelling recommendation reaches your desk, which question are you willing to ask, whose knowledge will you bring into the room, and what are you prepared to change because of the answer?