Skip to content

Writing

What an AI feature costs, with the arithmetic shown

Two numbers matter and most quotes only cover one. Here's the build cost, the running cost, and the arithmetic for working out yours.

By Yash Mittal4 min read

Building an AI feature into an existing product typically costs $3,000–8,000 and takes three to six weeks; a full AI-first product runs $8,000–25,000 over eight to sixteen weeks. The number most quotes omit is the running cost, which scales with usage and commonly lands between $5 and $60 per active user per month before optimisation — and between $2 and $18 after trimming context, caching prompts and routing simple requests to smaller models.

Two numbers decide whether an AI feature is a good idea, and most proposals only contain one of them.

The build cost is what everyone asks about and the smaller figure over three years. The running cost is what turns a well-received feature into a margin conversation eighteen months later. Both are estimable before you start.

Build cost, by shape of project

ProjectCostElapsed
Feasibility / discovery$600 – $1,8003–5 days
One feature into an existing product$3,000 – $8,0003–6 weeks
RAG system over one corpus$4,000 – $10,0004–8 weeks
Constrained agent with tool access$8,000 – $20,0006–12 weeks
Full AI-first product or MVP$8,000 – $25,0008–16 weeks
Small-studio pricing, mid-2026, ours included. Large consultancies run three to ten times these figures.

What actually moves a project within its band, roughly in order of impact:

  • The state of your data. Clean, accessible, well-permissioned data is the single biggest determinant. Data spread across four systems with no consistent identifier can double a project before any model work starts.
  • Whether it acts or only answers. Anything taking actions in your systems needs permission design, audit logging and confirmation flows. Roughly doubles the surface.
  • Permission complexity. "Everyone sees everything" and "each user sees only their own records" are weeks apart in a retrieval system.
  • Where the model runs. A hosted API is the cheap path. Self-hosting an open-weight model because data can't leave adds infrastructure work and ongoing operations.
  • How wrong it's allowed to be. A drafting assistant with a human in the loop is forgiving. Anything customer-facing and unsupervised needs a materially deeper eval suite and more guardrail work.

Running cost, and how to work out yours

The arithmetic is genuinely simple. What makes it feel unpredictable is that people estimate token counts badly — in our experience most teams underestimate their input tokens by three to five times, because they forget the retrieved context, the system prompt and the conversation history all count.

monthly cost =
    active users
  × requests per user per month
  × ( input tokens  × input price
    + output tokens × output price )

Worked example — a grounded Q&A feature:
  500 users × 40 requests
  × ( 6,000 input tokens  × $3 /1M
    + 700  output tokens  × $15 /1M )
  = 500 × 40 × ( $0.018 + $0.0105 )
  = $570 per month
Do this before you build. An estimate with the working shown beats a surprise on an invoice.

That's $1.14 per user per month, which is comfortable. Now hold everything constant and multiply the users by a hundred.

Active usersNaiveOptimised
500$570 / mo$180 / mo
5,000$5,700 / mo$1,800 / mo
50,000$57,000 / mo$18,000 / mo
The same feature at three scales, before and after optimisation. Frontier-model pricing, mid-2026.

At the top row the difference is a rounding error. At the bottom it's a headcount. This is why the cost model belongs in the scoping conversation and not in a review after launch.

The three levers that actually matter

In order of effect, and none of them involve a worse product.

1. Trim the context you send

The largest and most consistently ignored lever. Most retrieval systems pass ten passages when three would answer the question, because passing more feels safer. It isn't — beyond a point, extra context measurably degrades answer quality as well as costing money.

Re-ranking down to the three or four genuinely relevant passages typically cuts input tokens by half or more, and usually improves accuracy at the same time. It's the rare optimisation with no trade-off.

2. Cache the stable part of the prompt

Every major provider now offers prompt caching, and it's substantially discounted on cached input tokens. If your system prompt, tool definitions and few-shot examples are constant across calls — and they almost always are — that's a large fraction of your input tokens at a fraction of the price.

It requires structuring prompts so the stable prefix comes first and the variable part last. Doing that from the start is free; retrofitting it is a day.

3. Route by difficulty

The price gap between a small model and a frontier one is commonly ten to thirty times. Most production workloads are a long tail of easy requests with a minority of hard ones.

Classify cheaply, handle the easy majority with a small model, escalate the rest. The eval suite is what makes this safe — you can measure whether the small model is good enough per case type rather than guessing. Without evals, routing is a gamble; with them, it's the largest cost win available after context trimming.

Three years, not three months

Put the numbers on the same timeline before deciding anything.

Year 1Years 2–3Total
Build$6,000$6,000
Model usage$21,600$43,200$64,800
Retainer$30,000$60,000$90,000
Total$57,600$103,200$160,800
A single AI feature at ~5,000 active users, optimised, on a mid-tier retainer.

The build is under 4% of the three-year cost. Which is worth sitting with, because it reframes the decision: choosing a studio on build price is optimising the smallest line on the sheet, while the two large ones — running cost and ongoing ownership — are precisely what a good build makes cheaper.

What we'd tell you on a call

If your feature serves under a thousand users, don't spend a week optimising tokens. Ship it, measure real usage, optimise when the number is large enough to care about. Premature cost engineering on an AI feature is as wasteful as premature optimisation anywhere else.

If you're heading past ten thousand active users, model it before you build, because the architecture that's cheap at scale — write-time enrichment, aggressive caching, model routing — is not the same architecture as the obvious one, and retrofitting it is more expensive than choosing it.

Our full published ranges for every service are on the pricing page, and how we scope an AI project covers what discovery produces.

This is what we do. See our full pricing.

See our full pricing

FAQ

Related questions

How much does it cost to build an AI feature?

A single well-scoped feature into an existing product costs $3,000–8,000 at a small studio and ships in three to six weeks. A full AI-first product runs $8,000–25,000 over eight to sixteen weeks. Large consultancies typically quote three to ten times those figures for comparable scope. The biggest variable is the state of your data, not the complexity of the model work.

What causes AI project budget overruns?

In order: data that turns out to be less accessible than anyone said, permission requirements discovered during a security review rather than during scoping, and scope growth from a feature into a platform. Almost never the model work itself. A few paid days of discovery that actually touches the data removes most of this risk for a few hundred dollars.

How do we reduce our LLM running costs?

Three things, in order of effect: trim retrieved context to the passages that genuinely answer the question, which often halves input tokens and improves accuracy at the same time; cache the stable prefix of your prompts, which every major provider now discounts heavily; and route straightforward requests to a smaller model, since the price gap to a frontier model is often ten to thirty times. Together these routinely cut spend by two thirds.

Should we optimise cost before launching?

Only if you'll launch to more than a few thousand active users. Below that, the monthly bill is smaller than the engineering time spent reducing it — ship, measure real usage, then optimise. Above roughly ten thousand users, model it before you build, because the cheap-at-scale architecture is a different architecture and retrofitting is expensive.

Keep reading

More on this

Article

Why your AI proof of concept never shipped

The demo worked. Four months later it's still a demo. Here are the six things that actually block the gap, in the order they usually bite.

Article

In-house AI team vs agency: the arithmetic, including when we lose

Real salary numbers, real agency numbers, and the crossover point where hiring wins. Written by an agency, which should tell you how the last section goes.

Next step

Want this applied to your situation?

Articles generalise. A 45-minute call doesn't — tell us what you're actually dealing with and we'll be specific.

Start a projectRead more articles