How much does an AI opportunity assessment cost, and what should you get?

A proper AI opportunity assessment runs three to six weeks and should leave you with one of two outcomes. A prioritised three-year roadmap with a clear "do these three first" recommendation, or a written argument for why you shouldn't start yet and what to fix first. Either is a success. Both beat the third, unfortunately common outcome: a glossy deck with no ranked kill list.

Quotes for this kind of work vary wildly because "AI opportunity assessment" is a phrase with no industry definition. The same phrase covers a two-week desktop review and a six-month engagement that builds three pilots along the way. Before approving any quote, you want to know exactly what's in scope, how decisions will be made, and what the deliverable looks like on the last day.

What the fee actually buys

A legitimate assessment has five components. Each can be cut for cost, but cutting the wrong ones leaves you with a document you can't act on.

Discovery (week 1). Interviews with 8-15 senior stakeholders. Not just the CIO and Head of Data, but sales, ops, finance, product, and at least one frontline manager. You'd be surprised how many "AI opportunities" evaporate the first time an external party talks to the person who'd actually do the work. A good discovery phase produces a long list of 30-60 candidate use cases, most of which will be killed by the end of week two.

Data readiness (week 2). A hands-on look at the data foundation against the long list. Schema walkthrough, quality scoring, lineage mapping, access-control audit. This is the step cheap assessments skip, and it's the single most load-bearing one. If the data can't support the use case, the rest of the assessment is fiction. Budget at least a week with a senior data engineer inside the business's systems, not just reading documentation.

Value and effort sizing (week 2-3). For each candidate use case, an estimate of business value (in dollars, hours, or margin) and an estimate of build and run cost. Not precise (those would require a full scoping exercise per use case) but defensible and directionally right. The output is a 2x2 of value versus effort, with each use case plotted.

Workshops (week 3). Working sessions with the leadership team to pressure-test the ranking. Which use cases does the business already have capacity to run? Which depend on a platform that doesn't exist yet? Which are blocked by process or policy rather than technology? The workshops are where the long list collapses from 30-60 candidates to a short list of 8-15.

Synthesis and recommendation (week 4). The final deliverable is three documents. One: the ranked short list with rationale. Two: the phased roadmap, with dependencies mapped, expected value per phase, and a honest assessment of what needs to change in the organisation. Three: a written recommendation on the first three projects, with specific sponsor, specific team shape, specific success metric, specific kill criteria.

That five-step shape, with one to two senior consultants plus a senior data engineer for three to six weeks, is the scope you want. Anything less is a document with gaps; anything more is usually scope creep into doing the build.

The best AI opportunity assessment we've run killed 60% of the ideas the sponsor walked in with. The sponsor called it the most valuable engagement they'd run with us. The second-best just redirected the existing AI budget from the wrong three projects to the right three.

What a bad assessment looks like

Three red flags to watch for in a vendor's proposal.

• No hands-on data work. If the proposed scope is all interviews and no database access, the vendor is producing market research, not an assessment. Market research has its place but it costs $5k, not $45k.
• No ranked kill list. A deliverable that endorses every idea the sponsor brought in is a deliverable designed to win follow-on work. A legitimate assessment names the three projects the business shouldn't do, and why.
• Proof-of-concept deliverables bundled in. "As part of the assessment we'll build a quick AI prototype to demonstrate value" sounds attractive and is almost always a mistake. A prototype built in two weeks will succeed in the lab and tell you nothing about whether production is viable. Keep the assessment and the build cleanly separate.

What a good assessment looks like

Four things, specifically.

1. A ranked list of 8-15 use cases, each with a one-pager covering the business decision being changed, named owner, data readiness state, value estimate, cost estimate, risk register.
2. A data readiness scorecard against the short list, with the gaps costed.
3. A three-year phased roadmap showing which use cases depend on which platforms and which other use cases.
4. A written "first three" recommendation with specific sponsors, teams, and kill criteria.

If all four are in the statement of work, the quote is likely defensible and the scope is serious. If any are missing, the scope is wrong or the quote is wrong.

The cheap valid version

There is a real version of the assessment that costs meaningfully less. Two weeks, scoped to a single business domain. Example: "AI opportunities in our finance function" rather than "AI opportunities across the business." This is the right fit when:

• A specific executive has a specific budget ($250k-$2m typically) and wants to know whether to pitch it to the board.
• The business's data platform is already in reasonable shape and a readiness gap-analysis isn't the main question.
• The question is "which of these six AI ideas should we start with?" rather than "where does AI fit in our strategy?"

It is not the right fit when the question is the bigger, enterprise-wide "where should we invest?" That needs full scope, and trying to compress it into a two-week budget guarantees a shallow deck.

The expensive less-valid version

There's also the six-figure version, usually pitched by tier-one consultancies. Six months, a team of four, elaborate deliverables, framework branding. For a mid-market business this is almost always overpriced by a factor of three. What you're paying for is the logo, the insurance of a big-name report, and a lot of junior consultant hours that don't move the decision.

If the brand premium is worth it to you (sometimes it genuinely is, particularly for a regulated environment where the board needs external cover), go ahead. Just understand what you're buying, and that it's a different product from a well-run independent assessment.

How to scope yours

Before you send the RFP, write down the decision the assessment is meant to inform. "Should we allocate $1.5m in FY27 to AI initiatives, and if so, which ones?" is a decision. "We should understand AI better" is not a decision. If you can't name the decision, delay the assessment. You're not ready to buy it yet.

Then pick the scope.

• Enterprise-wide, pre-investment: four to six weeks, full scope, ranked roadmap with first-three recommendation.
• Enterprise-wide, mid-journey refresh: three to four weeks, focused on rebaseline rather than ground-up.
• Single domain, specific budget: two weeks, narrow scope, one business function.

Match the scope to the decision and the quote should be defensible. If a vendor's number feels misaligned, ask them what's in their scope that isn't in the alternatives, or vice versa.