What is an AI opportunity assessment?

How it works

Start with a business frame, not a tool list

The first step is to define what the organisation is actually trying to improve. That sounds obvious, but many AI discussions start in the wrong place. They start with capability. They should start with pressure. That pressure may be slower service, rising administrative load, compliance friction, margin squeeze, staff shortages, inconsistent quality, weak management information, or too much expert time being spent on routine preparation.

A sensible frame includes a few plain criteria that reflect how the business creates value and where it must stay careful. For a founder led firm, those criteria may be cash generation, speed to first gain, low implementation burden and low operational risk. For a regulated service provider, they may include data sensitivity, reviewability and evidential quality. For a group with multiple departments, they may include reusable capability across functions and access to a reliable sponsor.

This is also the point to set guardrails. Leaders should decide in advance what is out of scope, what requires extra governance and what must always retain human authority. If the organisation serves the public sector, health, financial services, employment or other higher scrutiny environments, that screening matters from day one.

Build a broad candidate list from real work

The strongest candidate lists do not come from a blank sheet and an innovation workshop alone. They come from several sources at once. One source is strategic pressure. Another is process evidence, such as queues, wait times, rework, exceptions and error hotspots. Another is frontline frustration, because staff often know exactly where time is lost. Another is customer pain, especially where response times, handovers or document churn cause delay. Another is management bottlenecks, where senior people spend large amounts of time reviewing, drafting, checking or routing work.

At this stage, the aim is breadth before judgement. Most organisations benefit from collecting more ideas than they expect to pursue. The point is not to promise delivery. It is to avoid narrowing too early. A weak list often reflects the imagination of one function. A stronger list looks across commercial work, internal operations, compliance, knowledge management and service delivery.

Good sourcing questions are simple. Where do people repeatedly search for the same information? Where is work delayed by reading, drafting, classifying or routing? Where do staff spend expert time preparing material rather than making the decision that matters? Where is there heavy demand variation? Where is judgement supported by large amounts of text, images or records? Where does rework occur because information arrives late or in the wrong form?

Turn each idea into a use case card

Raw ideas are hard to compare. A short use case card makes them comparable. It does not need to be long. One page is often enough. The discipline lies in the fields you include.

A practical card describes the business problem, who experiences it, the workflow affected, the task or tasks involved, the likely AI role, the current baseline, the volume of work, the dependence on data, the degree of judgement involved, and the main risks. It should also say what would need to remain human and who would own the work if it moved forward.

The most useful cards describe work at the right level. They are not vague, such as "use AI in finance". They are not overly technical either. They capture a real slice of work, such as draft first responses to routine supplier queries, classify incoming maintenance requests, prepare an approval pack from standard records, or summarise contract changes for a human reviewer.

A card should also state the counterfactual. What happens today if nothing changes? How long does the work take, who does it, what quality concerns arise, what delays does it create downstream, and what is the practical cost of leaving it alone? Without that context, it is almost impossible to judge whether a candidate is genuinely worth attention.

Estimate value across several dimensions

Value is broader than labour saving. Leaders should resist the trap of reducing every candidate to a single speculative cash figure too early. A better approach is to look at several value dimensions together and only convert to a detailed business case once a candidate has survived the shortlist.

One dimension is time and capacity. Could the use case reduce turnaround time, queue time or the burden of repetitive preparation? Another is quality and consistency. Could it reduce missed steps, improve first pass accuracy, or support more consistent handling? Another is commercial impact. Could it speed revenue work, improve conversion, reduce leakage, or make a service more responsive? Another is resilience. Could it reduce dependency on a small number of specialists or make work less fragile when demand spikes? Another is strategic leverage. Could the same data, prompt pattern, governance method or integration be reused elsewhere?

Leaders should also distinguish between direct and indirect value. Direct value appears in obvious places, such as reduced manual effort or faster response. Indirect value appears in second order effects, such as better management visibility, improved staff satisfaction because tedious work drops, or stronger compliance records because decisions are better documented. These effects are real, but they should be treated with discipline rather than optimism.

The test is not "can we imagine benefit?" The test is is there a credible path from this use case to measurable business improvement within a time frame we care about? If the answer is murky, the candidate usually belongs lower on the list.

Test feasibility, risk and organisational fit

This is where many inflated use cases fall away. A good candidate is not only valuable in theory. It must also be feasible enough to justify action.

Feasibility begins with the work itself. Are the tasks repeatable enough to learn from? Are digital inputs available? Is there a reasonable baseline? Are the success conditions clear enough to judge? Is the exception rate manageable? Does the workflow have enough structure that an AI supported step would be usable in practice?

Then comes data. Is the relevant data accessible, lawful to use, of acceptable quality, and linked to the workflow in a practical way? Are there enough examples, documents or transaction histories to support testing? Are there clear ownership and retention rules? If the work depends on fragmented files, undocumented judgement or missing records, the feasibility score should fall sharply.

Then comes risk. Does the use case touch personal data, confidential commercial material, regulated advice, employment decisions, or determinations that could materially affect an individual? Does it require explainability, auditability or human override? Would error be inconvenient, costly or unacceptable? What cyber and supply chain issues arise if an external model or provider is involved?

Then comes organisational fit. Is there a credible sponsor? Does the team have enough process stability to adopt change? Is there capacity for implementation, review and training? Would the use case depend on another capability not yet in place, such as document standards, a data inventory, or role clarity?

Score, rank and challenge without pretending certainty

Once cards are drafted, leaders need a ranking method. A simple method usually works better than a grand model. Most firms can score each candidate against a small set of criteria, such as value potential, speed to first gain, data readiness, workflow fit, implementation effort, governance difficulty and reusability. A red amber green scale or a one to five scale is usually enough.

The point of scoring is comparison, not mathematical theatre. If a team cannot explain why a candidate scored highly, the number is not helping. The most useful ranking sessions include challenge from more than one angle, commercial, operational, data, compliance and frontline. That cross functional view is important because a use case that looks attractive to one team may create hidden downstream burden elsewhere.

This is also the moment to ask whether an apparently strong candidate is only strong because it has been described too loosely. If scores differ wildly between reviewers, the use case often needs reformulation. The business problem may be real, but the proposed AI role may still be fuzzy.

Build a portfolio, not just a winner

A shortlist is better than a single bet. Portfolio thinking matters because one organisation should not load all its AI effort into the same risk profile. A healthy shortlist often includes a few near term use cases with modest complexity, plus one or two more ambitious bets that require stronger foundations.

That mix matters for pace and trust. Smaller, well chosen cases can create evidence, skill and confidence. More ambitious cases can stretch the organisation where there is enough sponsorship and patience. The balance will vary, but the principle is consistent: do not confuse the highest theoretical upside with the best next move.

Portfolio thinking also makes dependencies visible. Two attractive candidates may rely on the same scarce data engineer, the same legal review, or the same underlying document architecture. Another case may be slightly lower in immediate value but much better at building reusable capability. The shortlist should therefore show not only rank, but also sequencing logic.

Route each shortlisted candidate to the correct next step

The end product of an AI opportunity assessment is not a purchase decision. It is a decision map. Each candidate should land in one of a few clear categories: move forward now, hold for prerequisites, revisit later, or stop.

Candidates that move forward should usually pass into an AI workflow assessment. That next article is narrower by design. It asks whether one chosen workflow is genuinely suitable for AI, where the friction sits, what data and decision points matter, and what good looks like in practice. If the workflow proves suitable, the organisation often then needs workflow redesign so AI is fitted into a better process rather than laid over a bad one.

That handoff is where much selection work either pays off or unravels. A good opportunity assessment sets up that next stage cleanly because it has already exposed the business purpose, likely value, decision owner and main constraints.