AI Business Audit: Validate an AI Idea Before You Build It
What an AI business audit covers, the four tests an idea has to pass, and how UK mid-market teams kill bad AI projects in a week instead of six months.
By Ivan Pylypchuk, CEO of SoftBlues. Has led Claude and automation projects for finance, legal and healthcare teams across the UK and Ireland. Last updated: 29 July 2026.
Gartner expects more than 40% of agentic AI projects to be cancelled by the end of 2027, mostly for unclear business value rather than technical failure (Gartner, June 2025). Almost all of those were cancellable on day one. The information needed to kill them existed before anyone wrote code, and nobody went looking for it.
SoftBlues is an AI consultancy at softblues.io, a registered Anthropic Partner Network member and a Google Cloud Partner, working with regulated mid-market companies across the UK and Ireland. This is the assessment we run before we agree to build anything, including on our own internal projects.
An AI business audit is a short, structured check on whether a proposed AI project is worth doing: whether it is technically possible with the data you actually hold, whether the value is real and measurable, whether it fits how the business runs, and whether anyone will use it. It takes days rather than months, and its main job is to say no cheaply. A project that fails this check has cost you a week. The same project discovered in month five has cost you a budget.
Key facts
What does an AI business audit actually cover?
Four tests, in this order, because each one is cheaper to fail than the one after it.
1. Feasibility: can this be built with the data you have?
Not "can AI do this in principle", which is almost always yes and almost always irrelevant. The real question is whether the specific data needed exists, is accessible without a six-month integration project, is consistent enough to rely on, and is allowed to be used for this purpose under UK GDPR.
This is where most ideas actually die. The task is perfectly possible and the data sits in three systems, two of which have no API and one of which is a shared drive of scanned PDFs with inconsistent naming. We wrote about how to assess that honestly in AI data readiness.
2. Value: what is the number, and who owns it?
Take the process as it runs today. How many times a month? How long does each take? At what loaded cost? What does an error cost when it happens, and how often does it happen?
If you cannot answer those, that is the finding. You do not have a business case yet, you have an intuition, and intuitions are expensive to build software around. Name the number and name the person who is accountable for it moving.
3. Fit: does this match how the business actually runs?
A system that assumes an approval step nobody performs, or an ordering of work that stopped happening two reorganisations ago, will be technically correct and operationally useless. This test is mostly listening to the people who do the work rather than the people who describe it.
Also in scope: regulatory fit. For FCA, SRA or CQC-regulated processes, work out at this stage where the human sign-off sits, because retrofitting it later means redesigning the system. See human-in-the-loop AI workflows.
4. Adoption: will anybody use it?
The quietest killer. Teams evaluate models and forget that the output has to land in front of somebody who will act on it, in the tool they already have open, at the moment they need it. A brilliant answer in a system nobody logs into is worth nothing.
We wrote a whole piece on this failure because it is so common: why your company bought AI and nobody uses it.
What an audit should produce
Not a slide deck about the potential of artificial intelligence. Four things:
If what comes back is a roadmap with no rejections and no numbers, you have bought a sales document.
When you do not need one
If you already know the process, the data is in one system you control, and the value is obvious and small, skip the audit and build the thing. Assessment has a cost too, and spending three weeks validating a two-week build is its own kind of waste.
Audits earn their place when the idea is expensive, when the data situation is unclear, when a regulator will ask questions, or when several ideas are competing for one budget and somebody has to choose.
How this fits with the rest of the buying process
An audit is not the same as a proof of concept, and the two get conflated. The audit asks whether to build. The proof of concept asks whether the thing you decided to build actually works, and it involves writing code. Run them in that order, and only run the second if the first says yes. The practical detail is in how to run an AI proof of concept.
On cost, a focused audit is a small fraction of a build. For the wider picture on what AI work costs in the UK, see AI consulting costs.
Frequently asked questions
What is an AI business audit?
A short structured assessment of whether a proposed AI project is worth doing, covering technical feasibility against the data you actually hold, measurable business value, fit with how the organisation runs, and whether the people involved will adopt it. It usually takes days to two weeks and is designed to reject bad ideas cheaply.
How long does an AI business audit take?
For a single focused idea, a few days. For a portfolio of competing ideas across departments, one to two weeks. Anything longer is usually a discovery engagement rather than an audit, and you should be clear which one you are buying.
What is the difference between an AI audit and a proof of concept?
An audit decides whether to build and involves no code. A proof of concept tests whether the thing works and does involve code. Running a proof of concept before an audit is how companies end up with a working demo of something that was never worth building.
What usually causes AI projects to fail the audit?
Data access and adoption, far more often than technical difficulty. The model can nearly always do the task. Whether the data can be reached, trusted and legally used, and whether anyone will change how they work as a result, is where projects fall over.
Do we need an audit if we already know what we want to build?
Not necessarily. If the process is clear, the data is in one system and the value is obvious, go and build it. The audit earns its cost when the spend is significant, the data picture is murky, a regulator is involved, or several ideas are competing for one budget.
Can we run an AI business audit ourselves?
Yes, and the four tests above are the framework. What outside help adds is pattern recognition on the data question and a willingness to say no, which is harder internally when a senior person is attached to the idea. SoftBlues runs this as a structured conversation, and we do tell people when the answer is not to build.
What happens if the audit says no?
You have saved a build budget, which is the point. In practice the answer is rarely a flat no. It is more often "not this, but the adjacent thing that is smaller and has better data", or "not yet, fix this data problem first".
The cheapest AI project is the one you correctly decide not to start. If you want a straight read on whether an idea is worth building, and we will tell you when it is not, book a discovery call. SoftBlues is a registered Anthropic Partner Network member and a Google Cloud Partner. When the answer is yes, we put the system into production in 90 days at a fixed price, with a money-back guarantee if it fails.
See it in production
Systems we have built and run for clients, with the numbers that came out of them.
