# Where AI Earns Its Place

> A decision lens for adopting AI on regulated teams: sort tasks by what being wrong costs, then act differently in each zone. The finale of Still Accountable.

This series has spent six posts on what AI doesn't change: accountability, data duty, the review burden, the audit trail, the real bottleneck. That could read as a case against using it. It isn't. It's the case for using it well, which requires being clear-eyed about the limits first, so that the adoption sits on something solid rather than on enthusiasm.

So here is the constructive other half. AI earns its place on a serious team in more situations than the sceptics allow and fewer than the hype promises, and the difference between the two isn't about the technology. It's about the stakes of the specific task. Once you see the pattern, most decisions become straightforward, and you stop arguing about AI in general and start deciding about uses in particular, which is the only level at which the question has an answer.

## Why "should we use AI" is the wrong question

Teams exhaust themselves debating AI as a single yes-or-no, as though the organisation has to take a position for or against. That framing guarantees a bad conversation, because the honest answer to "should we use AI" is "for what?", and until that's specified, every general argument is met by an equally general counter-argument, and nobody moves.

The useful question is always about a particular task: should we use AI *for this*. And that question almost always has a clear answer, because the thing that decides it is something you can assess directly: what does being wrong cost here, and how cheaply can we catch it when it happens. The general debate is unwinnable. The specific decision is usually easy. Move the conversation from the first to the second and most of the paralysis lifts.

## The principle: sort by the cost of being wrong

Here is the lens that ties the whole series together. For any task, ask two questions. What does a wrong answer cost: nothing, something, or someone. And how easily can a wrong answer be caught before it does harm: trivially, with effort, or not until it's too late. Those two answers place the task, and the placement tells you how to use AI on it.

Where being wrong is cheap and errors are easily caught, use AI freely and enthusiastically. Internal drafts, first passes, exploration, brainstorming, summarising things a human will read with their own eyes, the low-stakes volume work that clogs people's days. Here the speed is close to pure gain, the downside is bounded, and hesitation just leaves value on the table. A serious team should be *more* willing here, not less, because clearing this work away is exactly what frees its experts for the work that needs them.

Where being wrong is expensive but reliably catchable, use AI with a real, funded check. The model produces; an accountable, competent person verifies properly, with the time and information to do it; the saving is whatever's left after honest review, which is often still worth having. The discipline is to keep the check real rather than letting it decay into a formality, because the decay is the default.

Where being wrong is expensive and not catchable until the harm is done: the irreversible decision, the unexplainable judgement, the action with no human between the output and the consequence. Be very cautious, and often decline. Not because AI can never help there, but because that's exactly the territory where everything in this series bites hardest: the accountability that doesn't transfer, the rationale you can't defend, the confident error that looks like competence. If you use AI there at all, it assists a human who remains fully and visibly accountable; it does not decide.

## What this looks like in one organisation

A single regulated organisation will have tasks in all three zones, and the point is that the same team can be aggressive and cautious at once without contradiction. The marketing team drafting copy a human edits, the engineers generating boilerplate that goes through normal review, the support staff getting suggested responses they adapt. All clearly first zone, and the organisation should lean in, not hold back out of a generalised caution that costs it real value. The same organisation putting a model near a clinical decision, a claims outcome, a citizen's entitlement, a financial assessment. Clearly third zone, and there the same organisation should be slow, demanding, and willing to say not yet.

A team that understands this doesn't have a single posture toward AI. It has a different posture for each zone, applied deliberately. That is what mature adoption actually looks like: not a policy of yes, not a policy of no, but the judgement to tell the zones apart and act differently in each. The failure modes are the teams with one setting: all-in everywhere until something serious breaks, or all-out everywhere while competitors capture the easy, safe gains they were too cautious to take.

## What it means in practice, and where this series lands

Stop debating AI in the abstract. Take your real tasks, sort them by what being wrong costs and how easily it's caught, and let that sorting decide. You'll find you can move quickly and confidently on most of them, the cheap, catchable majority, which buys you the room to be properly, unapologetically careful on the few that warrant it. That combination, fast where it's safe and slow where it isn't, is both better business and better stewardship than either blanket setting.

The through-line of everything here has been one idea: AI doesn't get a pass on the things that already mattered. That isn't a reason to avoid it. It's the reason to adopt it with your eyes open: to take the real gains where they're real, and to keep holding the line exactly where you were always accountable for holding it. The teams that do both will get more from AI than the ones who only cheer or only worry, because they'll still be standing, and still trusted, when it counts.

I do fractional engineering advisory work with teams thinking through exactly these decisions; there's more on the [about page](/about/) if that's useful.

---

Source: https://kads.au/writing/where-ai-earns-its-place/
Author: Kads Aziz