The question executives ask most is "what should we do with AI?" It is the wrong question, and it produces the wrong kind of list: long, undifferentiated, and ranked by enthusiasm. A better question is how much human judgement should stay in this piece of work — because the answer tells you what has to be true before AI can take any of it on.
In retail the cost of getting this wrong is not the failed pilot — it is the two years spent on the wrong use cases while a competitor automates the right ones. That reframing gives you three tiers. Keep human. Augment. Automate. Every candidate use case goes into one of them, and each tier has its own three-way split. The result is a nine-cell matrix that turns a wish list into a decision.
Keep human
AI stays out — and that should be a decision, not a default. Three reasons justify it.
AI is not capable yet. The task needs judgement, context or trust that current tools cannot deliver reliably. Write down what would have to change, and set a date to look again. Twelve months is a reasonable default; the tools move fast.
The risk is too high. Regulatory, safety, reputational or ethical exposure outweighs any plausible gain. A person owns the outcome. This cell matters most in regulated retail categories — pricing of essentials, anything touching vulnerable customers, anything a regulator will ask about by name.
It isn't worth it. AI could do it, but the value is too small to justify the change, the risk, or the attention it would consume. This cell is underused. Saying "not worth it" out loud frees budget and goodwill for the cells that are.
What this tier needs: a named owner, a review date, and a written reason. Without those, "keep human" drifts into "we never got round to it".
Augment
Human and AI work together. What changes — and how much — depends on where the person stands relative to the loop.
Human in the loop. AI drafts, suggests, scores; a person decides every single time. This needs enablement: good tools, usable prompts, short training. The process itself barely changes. Most first wins live here.
Human on the loop. AI acts within set bounds; a person supervises, samples and steps in. This is where it gets real, because it needs process redesign — new roles, escalation paths, thresholds that say when a human must look. Companies that skip this step discover it when something goes wrong.
Human over the loop. AI runs the work; a person owns the outcomes and the direction, not the individual cases. This needs governance: monitoring, audit trails, clear accountability, and a tested way to pull the plug. It is the most powerful form of augmentation and the one most organisations are least ready for.
What this tier needs: role clarity and decision rights, matched by training. Augmentation fails when nobody is sure who is accountable — so in the end nobody is.
Automate
AI runs it end to end. The useful insight here is that the more impact a task has, the more infrastructure has to already exist before you start.
Simple tasks. Repetitive, low-stakes, well-defined. Minimal change management. These are the quick wins that build confidence and buy you permission for the harder cells.
High-effort, lower-value tasks. Important but tedious work that eats skilled time. Automate it to free people for the judgement work in the Augment tier. The value is not the task — it is what the people do instead.
High-impact tasks. Decisions that move revenue or cost. These require the most: clean data, integrated systems, risk management, monitoring, and a rollback you have actually tested. Promising to add these afterwards is how automation projects become incident reports.
What this tier needs: data quality, integration, monitoring and ownership in place before go-live — not promised for afterwards.
The people question
Sorting use cases is only half the work. Every cell in the matrix assumes a person — deciding, supervising, owning, or choosing not to act. So the second half is knowing how ready those people are.
I use a single scale, which I call the AI Balance Scale, running from no knowledge, through readiness, to riskiness. The ends matter less than the middle.
No knowledge means someone has not used AI meaningfully. Low immediate risk, low immediate value, and a clear starting point. Readiness is the target: the person uses AI as an enabler, checks its outputs, knows its limits, and keeps ownership of the thinking. Riskiness is the state organisations underestimate: someone who lets AI do the thinking and acts on it unchecked. A very capable tool in hands that have not learned to doubt it. This is the highest exposure a company carries, and it is usually invisible until it isn't. The scale scores a pattern of use, not a person — the same individual can sit in different places for different kinds of work.
The assessment measures this five ways — a self-assessment, short realistic cases, situational tasks, observed behaviour on real tools, and guided exercises — because stated habits and actual habits diverge, and the gap is the finding. The point is not to grade people. It is to know who can be trusted with which tier of use case, and who needs support first.
Putting it together
Run the two halves together and something useful happens: the matrix tells you what to build, the scale tells you who can safely run it, and the overlap tells you where to start. A high-impact automation with no ready owners is a later project. A human-in-the-loop augmentation with a ready team is this quarter's.
The output is a ranked shortlist with the reasoning written down — what to do first, what it needs, who should run it, and what to leave alone. Discovery, assessment, roadmap: scope and timing follow the size of the organisation.
— Rihards Garančs, Data Lab · fifteen years in FMCG and retail before founding it