Andrew Crossley
    AI PRODUCT · UK

    AI product consultant

    An AI product consultant helps a company decide which parts of its product should use AI, designs those features so they behave predictably for real users, and puts the evaluation and cost controls in place that keep them working after launch. The work is product judgement applied to a probabilistic technology — scoping, evaluation design, failure handling and unit economics — not model training or infrastructure engineering.

    WHO THIS IS FOR

    The failure patterns I get called in for

    • The AI feature demos perfectly and breaks the first time a real customer phrases something differently.
    • Nobody can say whether the model output is getting better or worse between releases, because there is no eval set.
    • Your inference bill is growing faster than the revenue attached to the feature.
    • The board wants an AI strategy and engineering wants a spec, and the two documents do not connect.
    • You have shipped a chatbot because everyone shipped a chatbot, and usage is flat.
    • You are choosing between fine-tuning, RAG and prompt engineering and the arguments are being made on vibes.

    THE FRAMEWORK

    The Crossley Method: idea to first revenue in seven stages

    See the full method
    1. STAGE 1DiscoverWeek 1
    2. STAGE 2ValidateWeek 2
    3. STAGE 3PrototypeWeek 3
    4. STAGE 4Build MVPWeeks 3-4
    5. STAGE 5LaunchWeek 5
    6. STAGE 6First RevenueWeek 6
    7. STAGE 7ScaleOngoing

    WHAT THE WORK COVERS

    How AI product work maps onto the framework

    AI does not change what makes a product good. It changes the failure modes, the cost model and the definition of done. Every stage below is the standard stage with the AI-specific work called out.

    Discover

    Find the tasks in the user's workflow where being roughly right at speed beats being exactly right slowly. That test rules out most of the AI features companies plan and rules in the ones that stick.

    Validate

    Build the evaluation set before the feature. Fifty to two hundred real examples with agreed correct behaviour, including the awkward ones. If you cannot write the eval set, the feature is not specified.

    Prototype

    Cheapest possible version — prompt plus retrieval, no fine-tuning — measured against the eval set. Most teams discover their problem does not need the expensive architecture they assumed.

    Build & Launch

    Guardrails, fallbacks, confidence thresholds, human handoff, latency budgets, cost per action, and an interface that sets the right expectations about fallibility.

    First Revenue & Scale

    Unit economics per user. Pricing that survives token cost. Model routing so the cheap model handles the easy 80%. Ongoing eval so quality does not silently drift when a provider updates a model.

    HOW IT RUNS

    How an AI product engagement runs

    1. Week 1

      AI product audit

      Every AI feature shipped or planned, assessed on user value, technical risk, cost per action and evaluability. Output is a ranked list including the features to cancel.

    2. Week 2

      Opportunity mapping

      Workflow-level analysis of where AI genuinely compresses effort for your users, versus where it adds a layer they have to supervise.

    3. Weeks 3-4

      Evaluation and spec

      A written eval set for the priority feature, acceptance thresholds, failure-mode handling, and a spec engineering can build from without guessing.

    4. Weeks 5-8

      Build support

      Working with your engineers through implementation: prompt and retrieval design, model selection, threshold tuning, interface behaviour, instrumentation.

    5. Ongoing

      Quality and cost review

      Monthly review of eval scores, cost per action and user outcomes. Model providers change things silently; someone has to be watching.

    PRICING

    AI product consulting pricing

    Fixed-scope for the diagnostic work, retainer for anything ongoing. I do not bill hourly — it rewards slowness.

    AI product audit

    £3,500

    One week, fixed scope. Full review of AI features shipped and planned, with a ranked recommendation and a costed opportunity map.

    Feature definition sprint

    £6,000

    Three weeks. Eval set, acceptance criteria, failure handling, model recommendation and a build-ready spec for one priority feature.

    Ongoing advisory

    £4,000-£8,000/mo

    One to two days a week embedded with the team through build and launch, including monthly quality and cost review.

    Building rather than hiring? Get a number in 60 seconds with the MVP cost calculator.

    COMPARISON

    AI product consultant vs AI engineering agency

    Different problems. If you already know exactly what to build and it is well specified, an agency is the better purchase.

    FactorAI product consultantAI engineering agency
    Core question answeredShould this be an AI feature, and what does good look like?How do we implement this specification?
    Main deliverableEval sets, specs, roadmap, cost model, product decisionsWorking code and infrastructure
    Typical cost£3,500-£8,000 per month£20,000-£80,000 per project
    Works best whenYou have engineers but no AI product judgementYou have clarity but no build capacity
    Risk it addressesBuilding the wrong AI feature convincinglyBuilding the right feature badly
    After launchQuality drift, cost creep, pricing, iterationUsually a support contract or nothing

    PROOF

    AI products I have actually shipped

    CallFlow AI

    An AI product built end to end — problem framing, evaluation design, build and launch — using the same framework described on this page. Available as a live demo, not a case-study PDF.

    Journly and Vynl

    Two further AI-assisted products taken from idea to deployed and instrumented. Both were scoped down aggressively before build, which is the single decision that determines whether an AI MVP ever launches.

    Co-Ride

    Pre-seed community carpooling platform, founded November 2025, using AI for matching and operational automation rather than as a visible feature. Not every good AI feature looks like AI to the user.

    Wocal — £2.7M pre-money

    300+ venues, 2020-2025. Pre-dates the current AI wave, and that matters: the discipline of pricing, retention and unit economics is what stops an AI feature quietly destroying gross margin.

    Why most AI features fail after the demo

    The demo is built by the person who knows exactly what to type. The product is used by someone who does not. That single gap accounts for most AI feature failures, and it is a product problem rather than a model problem.

    There is a second, subtler failure. Traditional software fails loudly — a 500 error, a blank screen, something obviously broken. AI features fail quietly and confidently. The output looks right, reads fluently and is wrong. Users cannot tell, so they either stop trusting the feature entirely or, worse, they trust it and act on it.

    Designing around that means accepting fallibility in the interface itself: showing sources, exposing confidence, making the correction path faster than the generation path, and defaulting to a human handoff whenever the stakes exceed the model's reliability. Products that do this feel trustworthy even when the model is imperfect. Products that hide it feel unreliable even when the model is good.

    Evaluation is the product spec

    If you cannot write down fifty real examples with the behaviour you would consider correct, you have not specified the feature. This is the single most useful discipline in AI product work and the one most consistently skipped, because it is slow and unglamorous and there is a demo to get out.

    A working eval set has three parts: the happy path examples that must always work; the edge cases that reveal whether the system understands the domain; and the adversarial cases where the correct behaviour is to refuse, escalate or say it does not know. That third category is where most of the commercial risk lives.

    Once the eval set exists, everything downstream gets easier. Model selection becomes a measurement instead of an argument. Prompt changes get tested instead of debated. And when a provider silently updates a model — which happens — you find out from a score drop rather than from a customer.

    The unit economics nobody models until it hurts

    Traditional SaaS has near-zero marginal cost per user. AI features do not. Every action has a token cost, and if a heavy user triggers forty generations a day on a frontier model, a £30-a-month subscription can go gross-margin negative without anyone noticing until the quarterly review.

    The work here is unremarkable but it has to be done: cost per action measured in production, cost per user modelled at the ninetieth percentile rather than the average, and pricing designed with that in mind — usage tiers, fair-use limits, or a higher price point defended by the value delivered.

    The engineering lever is model routing. Most requests do not need the most capable model. A cheap model handling the easy majority, with escalation to a stronger model when confidence is low, routinely cuts inference cost by half or more without a measurable quality difference on the eval set. That decision is a product decision, because the thresholds encode a trade-off between cost and user experience.

    • Measure cost per action in production, not from provider pricing pages.
    • Model the ninetieth-percentile user, not the average one.
    • Route cheap models first and escalate on low confidence.
    • Cache aggressively — a surprising share of requests are near-duplicates.
    • Set fair-use limits before launch; retrofitting them annoys your best customers.

    Choosing between prompting, RAG and fine-tuning

    The decision is simpler than the discourse suggests. If the model lacks knowledge, that is a retrieval problem — give it the right context at request time. If the model lacks a behaviour, format or tone, that is a prompting problem first and a fine-tuning problem only when prompting has demonstrably plateaued against the eval set.

    Fine-tuning is worth it when you have a narrow, repetitive task, a few thousand good examples, and a cost or latency requirement that a smaller tuned model can meet where a large general model cannot. That is a real set of circumstances, but it is a much smaller set than the number of teams currently planning to fine-tune.

    The practical sequence is always the same: prompt, measure, add retrieval, measure, then consider fine-tuning. Teams that skip to the end spend three months and a lot of money to discover that their retrieval was the weak link all along.

    FREQUENTLY ASKED

    AI product consulting questions

    What does an AI product consultant do?
    An AI product consultant decides which parts of a product should use AI, specifies those features so they behave predictably for real users, and puts evaluation and cost controls in place so they keep working. The work covers opportunity mapping, eval set design, failure-mode handling, model selection guidance and unit economics — not model training or infrastructure engineering.
    How much does an AI product consultant cost in the UK?
    A fixed-scope AI product audit is around £3,500 for a week. A feature definition sprint including eval sets and a build-ready spec is around £6,000 over three weeks. Ongoing advisory at one to two days a week runs £4,000 to £8,000 per month.
    Do I need an AI product consultant or an AI engineer?
    If you know exactly what to build and need it built, hire an engineer or an agency. If your team can build but keeps shipping AI features that do not get used or cannot be measured, that is a product problem and a consultant is the cheaper fix.
    How do you measure whether an AI feature is working?
    Two layers. Model quality against a fixed eval set of real examples, scored on every release so regressions are visible. And product outcomes: adoption, task completion, correction rate, time saved and retention among users who touch the feature versus those who do not.
    Should we use RAG or fine-tuning?
    Use retrieval when the model is missing knowledge, which is most cases. Use prompting when it is missing a behaviour or format. Consider fine-tuning only when prompting has plateaued against your eval set and you have a narrow repetitive task with a few thousand quality examples plus a real cost or latency constraint.
    How do you stop AI costs running away?
    Measure cost per action in production, model the ninetieth-percentile user rather than the average, route the easy majority of requests to a cheaper model with escalation on low confidence, cache near-duplicate requests, and set fair-use limits before launch rather than after.
    Can you work with our existing engineering team?
    Yes, and that is the normal arrangement. I work in your tools alongside your engineers on specs, eval sets, thresholds and interface behaviour. I am not there to write your production code or replace your technical leadership.
    How long before we see results?
    The audit produces actionable decisions within a week, usually including features to cancel, which pays for itself immediately. A defined feature reaching production with proper evaluation typically takes six to ten weeks depending on your team's capacity.

    GET IN TOUCH

    Tell me about the AI feature that is not landing

    What you built, who it is for, and what is going wrong. I reply within 48 hours.

    Your details are used only to reply to you and are never sold or shared. See the privacy policy.