Andrew Crossley

    How much does it cost to build an AI application?

    SHORT ANSWER

    An AI application costs roughly £10,000–£30,000 to build as a focused MVP on hosted models, and £50,000–£150,000 for a production system with retrieval, evaluation and integrations. Running cost matters more than build cost: expect £0.01–£0.30 of model spend per task. If a customer generates a hundred tasks a month, price accordingly — AI products fail on unit economics more often than on build budgets.

    Answered by Andrew Crossley, Fractional Chief Product Officer · Updated 2026-08-01

    Why it matters

    Unlike traditional software, AI products have a real marginal cost per use. Flat-rate pricing over unlimited usage can turn growth into losses.

    Teams underestimate evaluation and quality work, which is typically 20-30% of an AI build and is not optional.

    How it works in practice

    1. 1

      Cost the build in three parts

      Workflow and interface, model integration plus retrieval, and evaluation plus guardrails. The third is the one that gets dropped and then re-added expensively.

    2. 2

      Model the per-task cost

      Tokens in and out, model tier, retries and context size. Multiply by realistic monthly usage per customer.

    3. 3

      Design pricing around usage

      Usage-based, credit-based or tiered caps. Unlimited plans need a hard cost ceiling behind them.

    4. 4

      Optimise once, deliberately

      Smaller models for easy cases, caching, prompt compression and retrieval trimming — validated against your evaluation set.

    Common mistakes

    • Pricing per seat when cost scales per task.
    • Using the largest model everywhere because it is simplest.
    • Skipping evaluation and paying for it later in churn.
    • Ignoring retry and failure costs in the model — they are often 10-20% of spend.

    FROM EXPERIENCE

    Unit economics that decide viability

    A £49/month product where the average customer runs 400 tasks at £0.08 per task spends £32 on inference alone. Gross margin before support is 35% — workable, but only if usage is capped and the model tier is right.

    Same product at £0.25 per task is loss-making from the first customer. That calculation belongs in the business case, before the build starts.

    Frequently asked

    Is fine-tuning worth the cost?

    Rarely at MVP stage. Prompting plus retrieval covers most needs; fine-tune when you have volume, stable tasks and evaluation data.

    Do open models reduce cost?

    They can at scale, but add infrastructure and ops cost. Below serious volume, hosted frontier models are usually cheaper all-in.

    How do I forecast model spend?

    Measure cost per successful task in the pilot and multiply by expected active users. Assume real usage is 1.5x your estimate.

    IN SHORT

    • £10k–£30k focused MVP; £50k–£150k production system.
    • Running cost £0.01–£0.30 per task — price on usage, not seats.
    • Evaluation is 20-30% of the build and cannot be skipped.

    Idea to MVP in six weeks

    A validated, instrumented first product in front of real users.

    Idea to MVP in six weeks

    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

    MORE ANSWERS

    AI and product development

    Will AI replace product managers?

    No, but it is removing a large part of what product managers used to spend their week doing. AI already writes specs, summarises research, drafts tickets, analyses feedback and builds prototypes. What it cannot do is decide what a company should refuse to build, hold accountability for a metric, or persuade a room. Roles weighted towards documentation are shrinking; roles weighted towards judgement are becoming more valuable.

    How do startups use AI?

    Startups use AI in three places: inside the product as a feature that completes a user task, inside operations to remove repeated manual work, and inside go-to-market for research, content and support triage. The highest-return use in a small company is usually operational — it lowers headcount pressure immediately, with no dependency on customers changing their behaviour.

    Which AI tools should founders use?

    Founders need four categories, not forty tools: a frontier chat assistant for thinking and drafting, an AI-assisted build tool for prototypes and MVPs, a research and synthesis tool for customer and market work, and automation for repeated operational tasks. Pick one per category, use them daily for a month, and only add a fifth tool when a specific recurring task justifies it.

    How can AI reduce operating costs?

    AI reduces operating costs by removing repeated, rules-light manual work: support triage and first-draft responses, data entry and reconciliation, research and reporting, and QA. Realistic savings in a small company are 20-40% of the time spent on those specific tasks, not of total headcount cost. The saving is real only if the reclaimed hours are redeployed rather than absorbed.