Most guides to AI development cost multiply an hourly rate by a team size and a number of months. The result is a range so wide that it says nothing. Our prices are no secret, so I prefer to start with them and then put them next to what the market charges, with sources and dates.
Short answer. At 4xxi, a software and AI company building since 2008 and registered in England and Wales, custom AI development starts with a 2-week audit for $5–10k and a deployable prototype built in 5 business days for $20k, both at a fixed price; everything after that is $200 per hour. On the market, a proof of concept is quoted at $15–50k over 3–6 weeks (Netguru, July 2026), a UK custom AI build at £50,000 to £1m+ over 3–12 months (We Call Shotgun, August 2026), and US senior AI contractors charge $130–200 per hour (A.Team, June 2026).
How much does AI development cost at each stage?
Each stage has its own price, and only the last one is open-ended. The first two are fixed because their scope can be bounded in advance; the third is billed in hours because a product that people use keeps changing.
| Stage | Time | Price | What you get |
|---|---|---|---|
| AI opportunity audit | 2 weeks | $5–10k, fixed | Where AI pays off in your product and operations, where it does not, your existing AI reviewed (if you have one), a measured report and a roadmap |
| Prototype sprint | 5 business days | $20k, fixed | A working, deployable prototype on your documents and systems, with the scope agreed up front |
| Build and ongoing engineering | As long as needed | $200/h | Senior engineers, the founder on the project, work inside your codebase and process if you have one |
| Fractional CTO or AI advisor | As needed | $200/h | The founder (me) as a part-time CTO or AI advisor |
| Edtek licence (our own product) | Depends on setup | On request | A non-exclusive licence to Edtek Chat, Draft and Cite |
The audit has a full money-back guarantee, and its fee is credited in full toward any project signed within 60 days. The sprint can include an evaluation set (a fixed list of real questions with expected answers); how big it is depends on the project, so we agree it together with the rest of the scope. Kickoff is typically within 3 business days of signing.
One rate covers everything after the sprint. Development, consulting and fractional CTO work are all $200/h; there is no separate price list for ‘strategy’ and ‘engineering’. Moreover, in custom work the code, the prompts and the evaluation sets belong to the client. Edtek is different: it is software we sell, so you get a licence, not ownership.
How long does the hourly part last? It varies a lot. In my practice there were projects of a few days, when a client only needed something deployed. A legal publisher we work with has a plan of roughly 6–8 months in several iterations for a team of three (I describe it in working alongside an in-house team). A payments company has had us on complex development, including AI, as ongoing work. And the American Arbitration Association has been a client since 2023, with the engagement running through 2028.
What are AI consultant hourly and day rates in 2026?
In the US, independent AI consultants charge roughly $75–350 per hour, Big 4 and similar firms $250–1,000+, and senior AI engineers $130–240, depending on how you hire them. UK consultants and consultancies charge £500–5,000 per day, depending on the tier. Here is the table I would use as a reference. Last checked: 29 October 2026.
| Who | Rate | Source |
|---|---|---|
| US independent AI consultant | $75–350/h | Groovyweb, October 2026, range across the guides it aggregates |
| US boutique AI consultancy | $125–500/h | Groovyweb, October 2026, range across the guides it aggregates |
| US Big 4, MBB and large firms | $250–1,000+/h | Groovyweb, October 2026, range across the guides it aggregates |
| US senior AI engineer, independent contractor | $130–200/h | A.Team, June 2026 |
| US senior AI engineer through a staffing agency (bill rate) | $170–240/h | KORE1, August 2026 |
| Eastern European senior AI engineer, contractor | $80–140/h | A.Team, June 2026 |
| Offshore AI studio | $29–49/h, fixed-price projects from $20k | RaftLabs, July 2026 |
| UK AI consultant contract, median day rate in job adverts | £564/day (10th–90th percentile £448–713) | ITJobsWatch, 70 rates, 6 months to 29 October 2026 |
| UK independent AI consultant | £500–2,000/day | We Call Shotgun, August 2026 |
| UK boutique or mid-tier firm | £900–1,600/day | We Call Shotgun, August 2026 |
| UK Big 4, large integrators, strategy houses | £2,000–5,000/day | We Call Shotgun, August 2026 |
| 4xxi | $200/h, i.e. $1,600 for an 8-hour day | This page |
Project prices follow the same tiers. The same We Call Shotgun guide puts a UK AI readiness audit at £3,500–15,000 (Big 4 versions at £50,000+), a strategy sprint at £12,000–40,000 and a custom AI build at £50,000 to £1m+. SpotDev’s shortlist of UK firms, checked on 12 September 2026, lists The AI Consultancy with a readiness sprint from £3,500, a pilot from £15,000 and a build from £40,000, and OpenKit’s AI audit from £10,000. For a prototype, Netguru’s 2026 budget guide (updated 21 July 2026) gives $15,000–50,000 over 3–6 weeks for a proof of concept. Our sprint is $20k for 5 business days, built on your own data and deployable.
Please bear in mind that almost all of these numbers are published by firms that sell the same services, us included. The exception is ITJobsWatch, which counts the day rates in job adverts: the median UK AI consultant contract pays £564 a day. That is what an employer offers a single contractor, not what a consultancy bills. All of them are a good reference for the order of magnitude, not a price list. Thus, $200/h sits inside the US independent and boutique consultant bands, at the top of the US senior contractor band, inside the UK boutique day-rate band, and well below Big 4. It is four to seven times the offshore studio rate. So why would anyone pay it?
Why can a $200/h senior team cost less in total?
Because you pay for hours × rate, and the hours are the part most cost guides ignore. A $200/h team is cheaper than a $40/h team only if it needs less than a fifth of the hours for the same result. That is a strong condition, so let me explain where the hours go.
Firstly, AI-assisted delivery. We build our own products (Preduzmi, Marginalia, WhatsMyArc, Individuate, Aporia) from idea to release in a few days each, and we use the same pipeline for client prototypes. That is why the sprint takes 5 business days, while the market quotes 3–6 weeks for a proof of concept. A team that writes everything by hand bills for the typing; a team that knows how to use coding agents on a real codebase bills for the decisions.
Secondly, seniority. Our engineers have 15–20 years of experience each, and there are no juniors and no handoffs. Most of the expensive mistakes in AI projects are not in the code but in the architecture: where the data lives, how retrieval works, what is checked deterministically and what is left to the model. When we audit systems that someone else built, the most common problems are exactly security and architecture, plus a ‘zoo’ of stacks and ready-made automation (for example, on GitHub) that nobody switched on. All of it is paid for twice: once to build, once to fix.
Thirdly, there is no sales layer. The engineers you meet on the first call build the project, and I work on it too. Hence, the hours you pay for are build hours, not coordination hours between an account manager, a project manager and a team that has never seen your business.
Is it always cheaper? No. If the work is routine and well specified (a set of standard screens, an integration with a well-documented API), a cheaper team may well cost less in total, and I would tell you so. The $200/h argument works where judgement matters: an LLM feature that must not make things up, an integration into an existing platform, a system that has to pass a security review.
Should an AI project be fixed price or time and materials?
Fix the price when the scope can be bounded; pay by the hour when it cannot. An audit and a 5-day prototype can be bounded, so we sell them at a fixed price. A product in production cannot, because models change and users ask questions nobody expected.
The problem with a fixed price on an open scope is simple. The vendor carries the risk, so the vendor adds a buffer for it, and you pay for the buffer whether the risk happens or not. Moreover, every change becomes a negotiation about whether it is ‘in scope’. On the contrary, time and materials only works if you can see what the hours buy. That is why we quote ongoing work in hours, not seats, and why baselines, regression suites and cost budgets are part of what we deliver: you can check that the system got better, not only that the team was busy.
I’ve been very happy with 4xxi. They deliver on time. So far, new and complex features are delivered according to our specifications.
My recommendation is a hybrid: fixed price for the first two steps, then hours with an agreed monthly ceiling and a review of results at the end of each iteration. A fixed quote for a whole AI product before anyone has looked at your data is a warning sign; I list the others in how to choose an AI development partner. It gives you a cheap way out after the audit or the sprint, and flexibility after that.
What drives the cost of an LLM feature, and what does one answer cost to run?
Five things drive the build cost, and one more drives the running cost that many budgets forget. The build cost depends on:
- The shape of your data. Clean text is cheap. Scanned PDFs, tables, footnotes and twenty years of inconsistent formats are not; the RAG implementation guide explains where ingestion and chunking go wrong.
- Integrations. Plugging into an existing platform, its authentication, its deployment and its devops usually takes longer than the AI part itself. In the legal publisher project above, integration with their platform and deployment automation were part of the work alongside the AI itself.
- How wrong the system is allowed to be. A demo that answers well most of the time is cheap. A system that must cite a source for every claim and refuse when it has none needs grounding enforced in code and an evaluation set to prove it. I describe how we do it in how to build an AI assistant that doesn’t hallucinate.
- Security and hosting. Private cloud or on-premises deployment, your own models instead of a public API, audit logs, access control.
- Regulation. Finance, legal and education each add their own requirements; see the vertical guides for fintech and edtech.
The running cost is the cost per answer, multiplied by the number of answers. For an LLM feature, cost per answer is roughly the input tokens (the question, the instructions and every retrieved passage) times the input price, plus the output tokens times the output price, plus retrieval and hosting. Prices per token are public (Anthropic, OpenAI) and change often, so one should recalculate rather than trust last year’s spreadsheet.
Here is the arithmetic with today’s list prices, as an illustration rather than a measurement. Say a grounded answer sends 8,000 input tokens (the instructions, six retrieved passages and the question) and gets 600 tokens back. On Claude Sonnet 5.5, at $2 per million input tokens and $10 per million output tokens (Anthropic, checked 29 October 2026), that is $0.016 + $0.006, about $0.022 per answer, or about $1,100 for 50,000 answers a month. On Claude Haiku 5.5, at $0.10 and $0.50, the same answer costs about $0.001, roughly twenty times less, if the smaller model is good enough for the task. Retrieval and hosting come on top.
There are several ways to reduce it, and none of them is exotic:
- Do not call the model when there is nothing to answer from. In Edtek Chat, if retrieval finds nothing relevant for a first question, the model is not called at all; the reader is told the library has no answer. It is honest, and it costs nothing.
- Use a smaller model where the task is narrow. We have instructed and fine-tuned small models for specific tasks, for example a local model that anonymises legal documents and one that finds citations in them; when the task is narrow enough for their size, they can be good enough and run on ordinary hardware, but only your own test cases will tell. Our benchmark of commercial, open-source and small local models shows where this works and where it does not.
- Bring your own model. Edtek can run on the customer’s infrastructure with the customer’s models; with a self-hosted model, that turns a per-token bill into a hosting bill.
- Measure it. A cost budget in the regression suite catches the release that silently doubles the price of an answer.
How should you budget the first 90 days?
Budget the first month in fixed prices and the next two in hours with a ceiling. Here is an illustration with our prices (an example, not a quote):
| Weeks | Step | Cost |
|---|---|---|
| 1–2 | AI opportunity audit | $5–10k, fixed |
| 3–4 | Prototype sprint on your data | $20k, fixed |
| 5–12 | Build: one senior engineer half-time, 8 weeks × 20 hours | 160 h × $200 = $32,000 |
| Total | $57,000–62,000 |
The audit fee is credited in full if you sign a project within 60 days, so the real number can be lower. The last row is the one to adjust: a deployment job might take a few days, while an integration into a large platform is a plan for several months. What matters is that by week 4 you have a measured report and a working prototype on your own data, so the decision about weeks 5–12 is made with evidence, not with a slide deck.
If you are a startup, the logic is the same but the first step can be the sprint rather than the audit, because the product question is clearer than the operational one; see how to validate an AI product hypothesis fast. If you already have an AI feature that misbehaves, start with an audit of the existing system instead.
When don’t you need this budget (or us)?
You do not need custom AI development if an off-the-shelf product already covers your case. A team that needs a better writing assistant should buy a subscription, not a project. Custom work starts to pay off when the AI has to work on your own data, inside your own product, with your own rules about what it may say.
You probably do not need us in three cases:
- A large volume of routine, well-specified work at the lowest rate. An offshore studio at $29–49/h will be cheaper, and that is fine.
- A governance programme for a large organisation (ISO 42001, an EU AI Act compliance programme across dozens of systems). That is Big 4 or specialist compliance work; we review and build systems.
- A permanent team of your own. If AI is the core of your product for years ahead, hire. We can work alongside the people you hire; I describe how in working alongside your in-house team.
If you are not sure which of these options fits you, I compared them in AI consultant vs agency vs fractional CTO.
To sum up: an AI project costs what its hours cost, and the hours depend on judgement much more than on the rate. That is why I recommend paying a fixed price for an answer to ‘is it worth it?’ before paying anyone by the hour.
Want a number for your own project before committing to a build? Start with the AI opportunity audit: 2 weeks, $5–10k fixed, money-back guarantee, credited toward a project within 60 days. If you already know what to build, plan a 5-day prototype sprint for $20k, or book a call to discuss ongoing work at $200/h.