How to Introduce AI
to Your Business
Written for owners and managing directors who know AI matters, have been told so by everyone selling something, and still have nobody able to say where it fits their particular business.
Quick Take
Start from your own work, not from the technology. Write down where your people's hours actually go, find the handful of tasks that involve reading, sorting or drafting something, and test AI against one of them for a month with a number attached. Most businesses have two or three genuine uses and a long list of expensive distractions. The skill is telling them apart before you spend anything.
Why Nobody Has Given You a Straight Answer
Almost everyone offering you advice on this sells something adjacent to it. Software vendors describe the problem in a shape their product happens to fit. Consultancies price a transformation programme because that is what they staff. Your own team, if they have opinions, usually have them about tools rather than about your margins.
The government's own assessment of why smaller firms are not adopting technology names three barriers, and the third is the interesting one: capability, cost, and awareness - business owners not understanding what the benefit would actually be.1 That is not ignorance. It is a reasonable response to being sold outcomes nobody has connected to your business.
The failure rate supports the caution. Gartner expects more than 40% of agentic AI projects to be cancelled before the end of 2027, and puts the causes as escalating costs, unclear business value, and inadequate risk controls rather than the technology falling short.2 Every one of those is a decision made before any software was installed.
Start With Where the Hours Go
An afternoon with a notepad beats a month of demos.
Take your last month. For each part of the business, write down what your people spent time on and roughly how much. Do not think about software while you do it. You are looking for the tasks that repeat, that everyone finds tedious, and that nobody would miss doing.
Then look for a particular shape. The tasks AI handles well have most of these properties:
- ✔ They involve reading, sorting, drafting or summarising text rather than deciding anything final
- ✔ A competent person could do them with no special training, given enough hours
- ✔ A wrong answer is annoying rather than dangerous, and somebody would notice
- ✔ They happen often enough that saving ten minutes matters
- ✔ The information needed already exists somewhere you control
In a typical SME that points at a short list: drafting first versions of routine documents and replies, pulling structured data out of invoices, forms and PDFs, summarising long documents so a person can decide faster, answering internal questions from your own manuals and policies, sorting and routing incoming enquiries, and helping technical staff write and review code. We have broken that down properly, function by function, in what AI can actually do for your business.
None of that is glamorous. That is rather the point. The unglamorous uses are the ones with a number attached, and a number is what makes the decision defensible in twelve months when somebody asks what the spend achieved.
Where It Will Waste Your Money
Four patterns that show up repeatedly, and the tell for each.
Automating a process nobody has written down
If two people in the same role do the job differently and both are right, you do not have a process, you have a set of habits. AI applied to that produces confident output that is wrong in ways only the experienced person spots. Write the process down first. Sometimes that alone fixes the problem and you can stop there.
Anything that must be correct every single time
These systems produce a plausible answer, not a guaranteed one. That is fine when a human checks the work and the check is cheaper than the task. It is not fine for calculations that go straight to a customer, a regulator or a bank without anybody looking. If the check costs more than doing it by hand, the use case is dead.
Buying the platform before proving the use
The order matters more than the choice. A month of paid trials on one bounded task tells you what a year of platform commitment cannot, and costs a fraction of it. Any supplier unwilling to be tested that way is telling you something.
Doing it because competitors are
They are mostly doing it for the same reason. This is the one where spend continues long after everyone privately knows it is not working, because stopping looks like falling behind.
Ninety Days, Four Steps
Enough time to find out. Short enough that stopping is cheap.
Weeks 1-2: List the candidates
From the hours exercise, pick the three or four tasks with the right shape. For each, write one sentence on what good looks like and one number you could measure today - hours per week, cost per item, days to turn something around, error rate. If you cannot name the number, the candidate goes to the bottom of the list.
Weeks 3-4: Pick one and set the bar
One task, one team, one measure. Decide in advance what result would justify continuing and what result would end it. Put both in writing before anyone touches a tool, because the temptation to move the goalposts afterwards is considerable and entirely human.
Weeks 5-10: Run it small
Two or three people, real work, everything else unchanged. Somebody senior checks the output weekly and keeps a note of what had to be corrected. That note is the actual finding. Ignore how impressive the demos were and pay attention to how much rework the thing generates.
Weeks 11-12: Decide honestly
Compare to the number you wrote down. Expand, try a different task, or stop. Stopping after ninety days having learned where AI does not fit your business is a good outcome and a cheap one. Continuing because the licences are already bought is how the 40% end up cancelled two years later.
The Boring Parts You Cannot Skip
Two things tend to be discovered late and both are cheaper to handle at the start.
Your staff are probably already using it. Someone in your business has pasted a customer email, a contract or a spreadsheet into a free AI tool this month. That is not misconduct, it is initiative without a rule, and the fix is a short written policy about what may be sent where rather than a ban nobody follows. If personal data is involved anywhere, the Information Commissioner's guidance on AI and data protection is the reference that matters in the UK.3
Regulation is about what you sell, not what you use internally. Using an assistant to draft your own emails is not a regulated activity. Putting an AI feature in front of your customers can be. If you sell into the EU, the AI Act's transparency duties for systems that interact with people or generate content apply from 2 August 2026, while the heavier obligations for high-risk uses were deferred to December 2027 and August 2028.4 The UK has no equivalent statute and works through existing regulators instead. Larger customers may also start asking whether you follow ISO/IEC 42001 or the NIST risk framework, both voluntary, both increasingly quoted in procurement questionnaires.5
Questions Worth Asking Anyone Selling You AI
- • Which specific task in my business does this replace, and how many hours a week is that task?
- • What does it cost when it is wrong, and who notices?
- • Where does our data go, who can see it, and is it used to train anything?
- • Can we test this on one team for a month before committing to anything longer?
- • What would you expect the measurable difference to be after that month?
- • What happens to the price when usage goes up?
- • Who in my business has to change how they work for this to pay off?
Two of those deserve more than a bullet. What AI actually costs a small business covers the pricing shapes and the costs nobody itemises, and is it safe to put company data into AI covers where your material ends up.
The last one catches most of it. Nearly every disappointing AI project we have looked at was a technology that worked attached to a change in working practice nobody sponsored.
Sources
- Department for Business and Trade - SME Digital Adoption Taskforce: 2026 update.Names capability, cost and awareness as the three barriers to SME technology adoption.
- Gartner - Over 40% of agentic AI projects will be canceled by end of 2027.Causes given as escalating costs, unclear business value and inadequate risk controls.
- Information Commissioner's Office - Guidance on AI and data protection.
- European Commission - Regulatory framework for AI, with the full text of Regulation (EU) 2024/1689 on EUR-Lex.Dates checked 28 July 2026. The staged timeline has been amended once already, so confirm before relying on it.
- ISO - ISO/IEC 42001:2023 AI management systems, and NIST - AI Risk Management Framework.
Still Not Sure Where AI Fits?
Book a free half-hour. Describe the business and we will tell you the two or three places worth testing, or that there are none yet.