The demo was the easy part
Someone at the firm pasted a set of past reports into ChatGPT or Claude, asked for a draft, and the draft looked useful. You already know what that proves. Any competitor with the same subscription can do the same thing this afternoon. This article covers the decisions between that demo and a product that clients pay for, which is where AI product development for a services firm starts. Most of them are business decisions, and the founder has to make them before hiring a builder.
MIT Sloan, reporting research from MIT CISR, says: "when competitors can also use many of the same foundation models and AI agents, simply adopting AI won’t necessarily set an organization apart." It puts the strategic question as "what others cannot build without access to its unique data or capabilities". The same piece adds a warning for any firm planning a launch: "Some of the effort saved during prototyping is deferred until organizations prepare AI systems for production." Most of what follows is about that deferred effort.
What turns your expertise into a product
A product applies the firm's judgement when no partner is in the room. Today that judgement lives in people: the consultant who knows which finding a board will act on, or the partner who can tell a defensible position from a risky one. For software to apply it, the judgement has to be written down in a form a builder can turn into instructions and checks, and the firm's own material (templates, past work, data) has to be there for it to work on. The firm also needs someone who can tell when the output is wrong.
Writing down the judgement
Raja Pabba, who spent 25 years at EY and Accenture and founded an AI company, wrote in California Management Review: "AI can automate process. AI cannot automate judgment without encoding." Writing down what your best people do, in enough detail that software can repeat it, is most of the work, and nobody outside the firm can do it for you. Two questions for the founder to put to the expert: could a colleague who has never met the client follow the expert's written account and reach the same call, and which steps in it could someone else check? A useful answer names the steps a second person could check. MIT Sloan makes the matching point about checking: "Organizations need domain experts with enough contextual judgment to recognize when AI output is wrong, incomplete, or inappropriate". A services firm already employs those people. They are the ones whose time it sells.
An example from law
The best-documented example comes from law. The Corporate Law Journal reported in January 2024 that Allen & Overy's ContractMatrix was built by the firm's in-house innovation group on its existing contract templates, such as non-disclosure agreements and M&A terms. More than a thousand of the firm's own lawyers used it before five clients began using it in January 2024. A&O is a large firm with its own innovation team, and a 40-person consultancy will not match that scale. What a smaller firm can copy is the sequence: its own material first, its own people next, and clients last.
The decision to take away: which of your material is good enough to build on, and which of your people would use the product before any client sees it?
Test a general model on your own work first
How two law firms tested general models
Before deciding what to build, find out what a general model already does with your kind of work. Linklaters gave a public example. In October 2023 it published the LinksAI English law benchmark: 50 questions from 10 practice areas, of the kind that "would require advice from a competent mid-level lawyer", with answers "marked by expert lawyers from each practice area". Each answer was scored for substance, citations and clarity. When the firm re-ran it on newer models in early 2025, it still recommended, as Legal Cheek reported, that they "should not be used for English law legal advice without expert human supervision". Allens used the same method for Australian law in May 2024.
Your own version of the test
Neither firm was building a product. Each was testing general models on its own kind of work, marked by its own people. That method suits any services firm. Your version is a set of questions that your own senior expert answers:
- Which past pieces of work show the job as clients buy it?
- What does a right answer look like, in the expert's own words?
- Which mistake would be unacceptable in front of a client?
- How good must the output be before a client sees it without the expert checking it?
Two more questions for the expert before the test. How many past cases would they need to see before trusting the result, and why that number? A useful answer ties the number to the range of work clients buy. What marks on those cases would show that the model already does the job? A useful answer is a standard another senior person could apply to the same results and reach the same verdict. Then run the general model on those cases and have the expert mark the results. Whatever it can't do on them is what your product has to add, and that becomes the brief for any builder. If the model already meets the expert's standard, the firm probably needs a better way of working with the tools it has, and a product would add little.
Decide which expert owns the AI's behaviour
One named senior person has to decide what right looks like and sign off any change to how the product behaves, both before launch and after it. If nobody holds that role, the developer makes those calls by default, guessing at standards they were never given.
MIT Sloan names the cost. Experts who spend much of their time checking AI output have less time for the complex work that keeps their expertise sharp. In a services firm that expert is usually also a senior fee earner, so the hours have a price, and the risk to their expertise builds slowly.
The questions for the founder: whose judgement is the product built on? How much of their week goes to it, and for how long? After launch, when a client points out a wrong answer, how does that person change the product's behaviour, and do they have to wait for a developer to do it?
Decide what client material the product may use
A services firm's best material is often client work. Using that work to build a product is a different use from the engagement it came from, and clients may not see it the way the firm does. The ICAEW wrote in November 2026: "Accountants’ confidentiality duties still apply when using AI." The same reasoning applies to other professions with confidentiality duties. Where the material contains personal data, the UK ICO's consultation on generative AI says purpose limitation means making sure the use is "in line with individuals’ reasonable expectations".
Allen & Overy addressed this in public. When it launched ContractMatrix, the firm said client data was not used to train the models underneath.
The questions for the founder: do your engagement letters allow you to reuse client work in a product, either anonymised or as it stands? Which clients would you ask first, and what would you offer them? Which material stays out of the product altogether? Take legal advice on the answers before any of that material reaches a builder.
Decide who answers when it's wrong
Once clients use the product, its answers are the firm's answers. The Guardian reported in October 2025 that Deloitte agreed to a partial refund to the Australian government on an AUD 440,000 report. Generative AI had been used to prepare the report, and its errors included "nonexistent references and citations". Deloitte stood by its findings, but the refund and the headlines happened anyway.
The Air Canada case shows the same thing from the other side. A Canadian tribunal held the airline to a wrong answer its chatbot gave a customer, and rejected the argument that the chatbot was responsible for itself. As CBS News reported, the tribunal said: "It makes no difference whether the information comes from a static page or a chatbot." A firm whose name sits on the product can't separate itself from the product's answers.
The questions for the founder: who checks the output before it reaches a client, and for how long after launch does that checking continue? What do the product's terms say it does and does not do? Does your professional indemnity insurance cover advice delivered through software? Ask your insurer directly, before launch.
Decide how it is priced against your fees
If the product delivers what the firm now bills time for, it competes with the firm's own fees. Foundation Capital, describing what it calls service as software, sets out the difference: "In the services business, responsibility for achieving the desired outcome sits with the company selling the service." Its piece describes outcome-based pricing, where the price is tied to the value delivered rather than to seats or effort. Pabba, in California Management Review, warns firms that hesitate: "Protecting existing labor revenue will delay transformation".
A services firm has an advantage here that a software start-up lacks. It already sells outcomes, and its clients already trust it to deliver them. The risk is that the product replaces billable hours without any change to how the firm charges.
The questions for the founder: does the product replace hours you bill, or does it sit beside the service? Do you sell it to existing clients or to a new market? Do you price per seat, per piece of work or per outcome? Whichever model you choose, can your expert's checking time support it, given that outcome pricing makes the firm answerable for the result?
Budgeting for the build when nobody can quote it yet
No credible public figure exists for what a services firm's AI product costs or how long it takes to build, so the number has to come from builders' quotes. Quotes vary widely because the builders giving them are guessing at the decisions above.
Making builders' quotes comparable
You can make quotes comparable. Write down the answers from the earlier sections: the test cases and what right looks like, who owns behaviour, which client material may be used, who checks output and what the terms say, and how the product is priced. Then ask each builder for a fixed price against that document. Builders quoting against the same written decisions give you quotes you can compare, and the document becomes the reference point when the scope starts to grow.
Costs to get quoted separately
Ask each builder to list four costs separately: the build itself, the model and hosting charges once clients use the product, changes after launch, and support when something goes wrong. Then add the cost that builders' quotes usually leave out, which is your expert's time before launch and after it. MIT Sloan's point about prototyping effort deferred to production applies most to the running and changing costs. To find the guesses in a quote, ask the builder two questions about each line: which of your written decisions does it depend on, and what has the builder assumed where your document says nothing? A line the builder cannot tie to your document is an estimate to press on before you sign.
Decide who holds the technical decisions on your side
Somebody on the firm's side has to choose the architecture, the model and the vendors, and judge whether the builders are doing good work. That person needs to be in place before a build partner is chosen, because otherwise the builder makes those choices and also checks them.
Three routes
There are three routes. The first is an employee: according to the Robert Walters 2026 salary survey, a full-time CTO in the UK earns between £105,000 and £320,000 depending on region. The second is an adviser brought in for specific decisions. Agathon works this way, advising with independent judgement, and it does not build the product. The third is a fractional CTO, which Agathon also offers as its Fractional AI CTO. A chief product and technology officer quoted by Raconteur said fractional CTOs are "quick to onboard and can start to add value quickly", but "their attention is always going to be split between their engagements". Split attention has another side: a fractional CTO brings what they have learned in other industries and other engagements, and part-time leadership fits a smaller budget than a full-time hire. The same article says a full-time CTO usually comes when "the business begins to scale and it’s too much for the founder to manage on their own". For a firm launching its first product, that point tends to come after launch.
Before the first build conversation
Whichever route you take, write down the decisions from this article before the first build conversation: what the general model can't do on your own work, who owns the product's behaviour, which client material it may use, who answers for its mistakes, how it is priced and who holds the technical calls. When the product starts to grow, who to hire next becomes its own decision.
FAQ
How should a professional services firm approach AI product development?
Write down the judgement the firm applies to one recurring client problem, then test a general model on the firm's own past work with the firm's own expert marking the answers. Decide which client material may be used, who answers for mistakes, how the product is priced and who holds the technical decisions, all before any build starts. Agathon advises firms on those decisions before they hire a build partner.
Can a services firm use client work to build an AI product?
Only within the firm's confidentiality duties and what its clients have agreed to, so check your engagement letters and take legal advice before any client material reaches a builder. Where the material contains personal data, UK data protection rules on purpose limitation also apply.
Do I need a CTO to build an AI product?
You need someone on your side who holds the technical decisions and can judge the builders, and that can be an employee, an adviser or a fractional CTO. Agathon's Fractional AI CTO holds those decisions for the firm and does not build the product.
References
- MIT Sloan: 5 things to know about how companies can create value from AI
- Linklaters TechInsights: The LinksAI English law benchmark: can you rely on AI for legal advice?
- Legal Cheek: Linklaters gives AI exam on English law
- Allens: Allens publishes Australia's first AI legal benchmark
- The Corporate Law Journal: City firm Allen & Overy launches AI contract negotiation tool
- ICAEW: AI and client data: complying with the principle of confidentiality
- ICO: Generative AI second call for evidence: purpose limitation in the generative AI lifecycle
- The Guardian: Deloitte to pay money back to Albanese government after using AI in $440,000 report
- CBS News: Air Canada chatbot costs airline discount it wrongly offered customer
- Foundation Capital: AI leads a service as software paradigm shift
- Raconteur: Should you hire a fractional CTO?
- California Management Review (Raja Pabba): From rate cards to outcomes: consulting's fourth transformation
- Robert Walters: Chief technology officer salaries: UK averages 2026



