When AI Develops a Gut Feeling, Who Knows When to Doubt It?

Fast decisions are becoming cheaper. Knowing when to trust them may become more valuable.

An experienced manager looks at a proposal and says, “Something doesn’t feel right.”

An AI examines the same information and returns a decision with 94% confidence.

One offers experience. The other offers a number. Neither, by itself, tells us whether to act.

We tend to give these judgements different treatment. Human intuition can acquire authority through seniority, reputation or familiarity. A machine’s judgement can acquire authority because it arrives quickly, consistently and with a reassuring decimal point.

Both deserve a harder question: under what conditions has this judgement earned our trust?

That question is becoming more practical as models such as Jev bring fast, structured decisions into software. TypeSafe describes Jev as a “System One” model: instead of producing an essay, it returns choices, scores and probabilities that another system can use directly. Classification, routing and checking become potential components of an automated workflow. TypeSafe’s introduction

The comparison with human intuition is useful, provided we treat it as an analogy. A fast model and an experienced person do not necessarily arrive at their answers through equivalent mechanisms.

But they create a similar organisational problem: a judgement can arrive before we have examined its basis.

Experience needs feedback

Human gut feeling is sometimes compressed expertise.

Someone who has repeatedly encountered a situation, acted and seen the consequences may recognise a pattern before they can explain every step. That recognition can be valuable.

It can also be misplaced.

Kahneman and Klein’s research offers a useful distinction. Skilled intuition depends on an environment with learnable regularities and sufficient opportunities to learn them through feedback. Confidence and experience alone are not reliable evidence of expertise. Conditions for Intuitive Expertise

That should make organisations examine their own decision-making habits.

Did the manager receive clear feedback on earlier decisions? Were unsuccessful outcomes investigated? Did anyone record occasions when the confident judgement was wrong?

Or has the organisation simply become accustomed to accepting it?

The same questions belong beside a machine’s confidence score. What was it tested against? Which mistakes did the evaluation reveal? Does performance hold for this kind of work?

A valid answer can still be wrong

There is promising evidence around fast decision models, alongside reasons for caution.

A September 2026 preprint evaluating Jev across 37 datasets reports strong performance on many tasks, but also weaknesses involving binary decision thresholds and rubric-based quality judgements. Those findings support investigation; they do not establish fitness for every operational decision. Benchmark study

Another recent preprint examines how typed decision models respond when option names are reassigned to different definitions. It reports substantial changes in decisions under some conditions, despite outputs continuing to conform to the required format. Type-Safe Is Not Error-Free

These studies address different questions. A model can perform well on a benchmark and remain sensitive to how a particular decision is framed.

That distinction matters when software acts on the answer.

A correctly formatted “approve” can still be the wrong decision. A high probability can be poorly calibrated for an unfamiliar task. And an accurate prediction does not confer permission to act.

We should test all three separately: the output’s form, the judgement’s reliability and the authority for the resulting action.

Who decides when to slow down?

The interesting design question is where fast judgement belongs.

Consider an agent handling an incoming customer request. A fast model might suggest a category or identify the likely team. A more involved assessment might be needed when the evidence conflicts, the request is unfamiliar or the consequences of a mistake are substantial.

A person may need to decide when the request involves an exception, a disputed obligation or a choice the organisation has not authorised the agent to make.

However, adding a human approval button does not automatically improve the decision. The reviewer needs relevant knowledge, sufficient time and the authority to challenge what is presented.

Nor should we assume that a slower model necessarily gives the better answer.

The escalation itself needs evaluation. Does the additional step catch meaningful errors? Does it introduce different ones? Does it reduce uncertainty, or merely produce a longer explanation?

The valuable capability is knowing when a fast answer has earned the right to become an action.

The economics of hesitation

This brings us back to the cost of digital labour.

A fast decision might cost very little. But if it sends work to the wrong team, creates unnecessary review or triggers an action that must be reversed, its low inference cost tells us little about its economic value.

Equally, requiring extensive reasoning and human review for every routine decision can consume the saving we hoped to achieve.

For Agent Control’s Digital Labour Exchange, I would frame this as an allocation problem: how much decision effort does this job justify?

The work order should establish the intended outcome, budget, permitted actions and acceptance conditions. The execution record should connect the initial decision to subsequent checks, escalation, correction and eventual acceptance.

Then we can compare fast, slow and combined approaches on equivalent work.

The measures should include total cost per accepted outcome, completion rate, elapsed time, reviewer effort and consequential errors. Otherwise, an apparently economical system may simply be leaving difficult work unfinished.

That is a practical design direction, not a claim that our work has established Jev’s advantage over human judgement.

Where will the next generation acquire its intuition?

There is a workforce question here too.

If agents handle routine cases, junior colleagues may lose some of the repeated exposure through which expertise develops. We could gradually remove opportunities to learn while expecting people to remain capable of supervising exceptional decisions.

That would be a fragile arrangement.

An organisation introducing digital labour should plan how people will practise judgement: working through representative cases, comparing decisions, investigating mistakes and receiving feedback on what happened afterwards.

The aim should be to develop people who can challenge an answer intelligently, including when they agree with it.

Judgement needs a memory

Neither human nor machine expertise should be treated as permanently valid.

Policies change. Customers behave differently. A model is updated. A decision that was sensible six weeks ago may no longer meet today’s requirements.

We need to preserve the evidence, applicable rules, model or reviewer involved, uncertainty and acceptance decision. Later reassessment should add a new judgement rather than silently overwrite the old one.

That memory lets us distinguish a poor original decision from a reasonable decision whose conditions have changed.

Fast models make this discussion timely. Human intuition makes it familiar.

For either, I would ask: how reliably does it decide in these conditions, what happens when it is wrong, and does it recognise when those conditions no longer apply?

A cheap answer is useful. A dependable decision is worth more.

Knowing when to pause may be the most valuable capability of all.

Also published on Medium ↗.

This article reflects the evidence and development status at its original publication date.