Do we get any working software at the end of this?

No. You get a document, a costed ranking and a short deck for the board. That is deliberate. The whole point of spending two to four weeks on the question is to avoid spending eight months building the wrong answer, and every hour spent writing code during an audit is an hour spent defending that code instead of the finding.

Three vendors have already told us what to buy. How is this different?

They are quoting for the work they want to sell. The section of this report that nobody quoting for a build will ever hand you is the list of things you should not do, each with the reason written next to it. Some of those reasons are that the data does not exist, some are that a person is faster, and some are that the return is real but too small to be worth the disruption.

Who do you need to talk to inside the company?

Eight to fifteen people, spread across departments, and mostly not the leadership. Directors describe how the process is supposed to work. The person who actually runs it describes the workaround, the re-keying and the spreadsheet that holds it all together — which is where the recoverable time usually is.

What if the honest answer is that AI is not worth it for us?

Then that is the finding and it gets written down with the working behind it. It is a legitimate outcome and a cheap one at this price. In practice the more common result is narrower than the board expected rather than empty — one or two operational processes worth automating, and a much longer list of ideas that do not survive contact with the data.

Is the board presentation included?

Writing the report and standing in front of a board to defend it are separate pieces of work and are priced separately. Both are usually worth buying, and which matters more depends on how your board actually reaches decisions. A report that arrives by email gets skim-read; a report that gets argued over in the room gets decided on.

What happens after the audit?

You own the document and you are free to take the ranked projects to anyone, including your own team or a competitor of ours. If you want us to build the first one, the audit work carries into it — the interviews, the cost model and the success measures are the same artefacts, so nothing is paid for twice.

AI & Machine Learning

AI readiness and strategy audit

A ranked, costed answer to where AI belongs in your business — and a written list of the places it does not.

From: $1,000–$10,000
Typical time: 2–4 weeks
Usually bought by: Mid-size companies

What this actually solves

The question usually arrives from the board and it is rarely a technical one. What is our AI strategy? Underneath it sits a specific discomfort: every vendor in the sector is selling something, three people inside the company hold three different opinions, a competitor has put out a press release, and nobody at the top has the time to read enough to tell the difference between a tool that will pay for itself and one that demonstrates beautifully in a forty-minute meeting.

The expensive version of this problem is not inaction. It is a pilot that has already started. Someone has bought seats, a department has quietly signed up for its own tool, a proof of concept has been running for two quarters and cannot be killed because it has a sponsor. Finding out that a project was never viable costs a fortnight during an audit; it costs most of a year and a budget line if you find out during the build. That gap is most of what this service is worth.

What you get at the end is a document and a decision, not software. That is not a limitation of the engagement, it is the engagement. The most valuable pages in it are usually the ones listing what you should not build, with the reason written beside each item — which is a page nobody who is also quoting for the build has any incentive to write.

What we do, and in what order

Interviews, department by department. Eight to fifteen people, deliberately weighted away from the leadership. The questions are about last week rather than about the future — what took the most time, what had to be redone, what gets copied by hand from one screen into another. Asking someone whether AI would help them produces enthusiasm. Asking what they actually did on Tuesday produces evidence you can cost. Three to five days.
A task inventory, sorted into four buckets. Every task named in the interviews goes into exactly one: it can be done with today's models now; it can be done with real engineering effort; it is not a job for a model at all; or it is a job for a model but the data required does not exist in usable form. That fourth bucket is the one that saves the money, and it is the one that almost never appears in a vendor proposal. Two to three days.
A data reality check. Everything in the fourth bucket gets looked at rather than described — where the records actually sit, what condition they are in, whether the history is complete enough to be worth anything, and whether anyone is permitted to move it out of the system it lives in. A use case that scores well on a slide and depends on fifteen years of decisions locked inside a legacy platform and a shared drive of spreadsheets is not a candidate yet, and saying so early is the job. Two to three days.
Cost and return, on the whole cost model rather than the API price. Each surviving candidate gets costed properly. The model bill is one line among several: embeddings, reranking, a vector database, application hosting, logging and monitoring, and a realistic allowance for retries and failed calls all sit next to it. So do the hours a person will spend every month reading outputs, correcting the bad ones and re-indexing new documents — which on small systems is routinely the largest number in the table. Against that total goes the time measured in step one. Two to three days.
Ranking, the do-not-build list, and a roadmap. Three projects, in order, each carrying the number that justifies it, with the first one planned in enough detail that work could start on Monday. Everything else is named and explicitly deferred with its reason, which is what stops the same ideas coming back around next quarter as if they were new. Then a three, six and twelve month roadmap and a short deck for the board. Three to four days.

There is no exotic tooling behind any of this. It is a structured interview guide, a spreadsheet holding the cost model, and a document. What makes it worth buying is the arithmetic inside the spreadsheet and the willingness to write down a negative finding.

What moves the price

The range is $1,000–$10,000 across 2–4 weeks, and three things decide where a company lands inside it.

How many departments, and therefore how many interviews. A single operational team is a short piece of work. A company with sales, claims, underwriting, finance, compliance and customer service is six separate task inventories, six sets of workarounds and six different answers to the same question about where the data lives.

Whether the data can be inspected or only described. Being given read access to the systems is faster and produces a far better answer than being told about them in a meeting. When access is not possible, the audit still works, but the data section moves from assessment to assumption and the report says exactly which conclusions rest on that.

Whether we present it as well as write it. Defending the ranking in front of a board, taking the objections and adjusting in the room, is separate work and is quoted separately. It is usually the difference between a report that is read and a report that is acted on.

What usually goes wrong

The audit turns into a wishlist. Twenty-two possible projects, neatly categorised, every one of them plausible. You read it, agree with all of it, file it and do nothing — because after twenty-two options you still do not know where to start, which is the exact question you paid to have answered. Three projects in order, with the first one costed and planned, is worth more than twenty good ideas. One decision that gets made beats a comprehensive document that does not.

The interviews collect opinion instead of time. Would AI help your team? is a question that gets a yes from almost everyone, and the yes is worthless — it is a statement about mood, not about work. The whole return calculation downstream depends on knowing how many hours a week a task really consumes and how often it goes wrong, and that number only comes from asking about specific things that already happened.

The return is calculated on the token price alone. A project is judged against the published cost per million tokens, it looks extremely profitable, and the vector database, the reranking calls, the hosting, the retries and the four hours a month someone spends reviewing conversations are all missing from the table. The result is a business case that is wrong by a factor of two or more, and the first project ships at a loss — which is the fastest way to make an organisation decide that AI does not work for them.

See where the site stands before you brief anyone.

If the site is already live, start from what it scores rather than from a guess. The audit reports what it checked, what it found, and — the part most tools leave out — what it could not check, and why.

See the website audit Talk about this work

Questions people ask before they commit.

Do we get any working software at the end of this?: No. You get a document, a costed ranking and a short deck for the board. That is deliberate. The whole point of spending two to four weeks on the question is to avoid spending eight months building the wrong answer, and every hour spent writing code during an audit is an hour spent defending that code instead of the finding.
Three vendors have already told us what to buy. How is this different?: They are quoting for the work they want to sell. The section of this report that nobody quoting for a build will ever hand you is the list of things you should not do, each with the reason written next to it. Some of those reasons are that the data does not exist, some are that a person is faster, and some are that the return is real but too small to be worth the disruption.
Who do you need to talk to inside the company?: Eight to fifteen people, spread across departments, and mostly not the leadership. Directors describe how the process is supposed to work. The person who actually runs it describes the workaround, the re-keying and the spreadsheet that holds it all together — which is where the recoverable time usually is.
What if the honest answer is that AI is not worth it for us?: Then that is the finding and it gets written down with the working behind it. It is a legitimate outcome and a cheap one at this price. In practice the more common result is narrower than the board expected rather than empty — one or two operational processes worth automating, and a much longer list of ideas that do not survive contact with the data.
Is the board presentation included?: Writing the report and standing in front of a board to defend it are separate pieces of work and are priced separately. Both are usually worth buying, and which matters more depends on how your board actually reaches decisions. A report that arrives by email gets skim-read; a report that gets argued over in the room gets decided on.
What happens after the audit?: You own the document and you are free to take the ranked projects to anyone, including your own team or a competitor of ours. If you want us to build the first one, the audit work carries into it — the interviews, the cost model and the success measures are the same artefacts, so nothing is paid for twice.

Often taken together.

Start with the brief, not the invoice.

Send what you are trying to achieve. You get a written scope, a price and a date — or an honest “this is not the right service for that”.

Talk about AI readiness & strategy audit All AI & Machine Learning services