Start where the work is repetitive, steady in volume and has a right answer that a person can check in seconds: supplier invoices going into your accounts, first drafts of replies and quotes, and notes from calls. Leave customer chatbots, forecasting tools and CV screening until your records and processes are in order. Whatever you pick, time the job before and after, and count the checking.
Updated 12 October 2026. We first published this in January 2025. We have removed the figures we could not source, including a chatbot resolution rate, monthly cost ranges and a 70/30 split for content, and replaced the tool list with checks you can run yourself.
Where UK firms are with AI
The Office for National Statistics reports that around 35% of UK businesses with 10 or more employees were using at least one AI technology in June 2026, up from around 12% in late 2023. Large language models, the technology behind ChatGPT and Claude, were the most widely used type, at 18%. Use is lower in industries such as construction, where the ONS puts it at 13%.
Most UK firms with 10 or more employees are therefore still not using AI, so you are not late. The question is not how to catch up. It is which jobs are worth the effort of trying it on.
How to pick the first job
Score each candidate on four questions.
- How often does it happen? Weekly or more often gives you plenty of chances to learn the tool and to measure it.
- Can a person check the answer in seconds? A figure against the PDF, a draft against your price list.
- Is the input already digital? An email, a PDF or a spreadsheet is easier than paper or a phone call.
- What does a mistake cost? A draft someone reads costs seconds if it is wrong. A wrong figure in your accounts costs more.
Here is the method applied to an imagined distribution firm of about 40 people. The jobs and scores are ours, to show how it works.
| Job | How often | Easy to check? | Cost of a mistake | Start? |
|---|---|---|---|---|
| Keying supplier invoices into the accounts | Daily | Yes, against the PDF | Low, if reviewed before posting | Yes |
| First draft of a customer reply | Daily | Yes, a person reads it | Low | Yes |
| Notes and actions from calls | Several a week | Yes | Low | Yes |
| Pricing a bespoke job | Weekly | Hard | High | Draft only, you set the number |
| A chatbot answering customers alone | Constant | Hard | High | Not yet |
Read the table by its shape. The three rows marked Yes are jobs where a person can see the right answer at a glance and a slip is caught before it costs anything. The bespoke price is a judgement, so AI can lay out the working while you set the number. The unattended chatbot is a mistake nobody checks before the customer sees it. The jobs worth starting with are usually the dull ones.
Start with what you already pay for
If you use Xero, Smart Document Capture is, per Xero’s page, included on its Ignite, Grow, Comprehensive and Ultimate plans. Xero says it reads receipts and invoices with AI from a photo, an email or an upload, and shows the original next to the extracted details so you can double-check them before a record is created. Try it on last week’s bills before you pay for anything else, and note which ones it reads cleanly and which it does not. Then write down two things: the rules you follow when you code a bill, and the exceptions that keep coming up. Those two documents are what any tool, off-the-shelf or custom, will need from you.
The same habit applies to your CRM, your accounting package and whatever office suite you use. Check what they have added recently before you buy another subscription.
A general tool will take you a long way on reading the document. Where the work is matching a bill to a purchase order and coding it to the right job by your own rules, you may find it does not go far enough, and that is where custom work starts. It is what we built for CGS Construction: we designed and rebuilt cgs-construction.com, then built an online quoting tool for the firm’s work and an invoice dashboard linked to Xero, where supplier bills are filed from the mailbox, matched to purchase orders and coded, instead of being keyed in by hand. We now work with them on a monthly retainer: ongoing technical support, and rolling out Claude with training for the team. The CGS Construction case study has the story in order. We have no figures to give you for it, and we are not going to invent any.
First drafts, replies and notes
ChatGPT and Claude on a business account can cover most of this: a first reply to an enquiry from your price list, a summary of a call with the actions listed, a tidy report from site notes. The rule is the same every time. A named person reads it before it leaves, and checks every figure, name and date. Tell people when a call is being recorded or transcribed.
Our guide to where AI saves time and where it does not has worked examples. If you are putting a tool in front of a whole team, rolling out ChatGPT or Claude to your team sets out a 30 day plan.
What we would leave for later
Customer chatbots that answer on their own. A wrong answer is yours, and nobody checks it before it goes. Some are also billed by usage, so the bill moves with the volume. We cover both in AI-powered CRM for small business. Our own first fix would be a quicker reply from a person, with a reminder when nobody has replied.
Forecasting and prediction. These tools learn from your past records. HubSpot’s own page says its AI lead scoring needs a Marketing Hub Enterprise subscription and a sample of at least 50 contacts, 25 converted and 25 not. That is a small sample, but it assumes someone recorded who converted. If your history sits in spreadsheets and inboxes, there is nothing for the tool to learn from yet. Get the records in one place first, and our note on signs your business has outgrown spreadsheets will help you judge when.
CV screening. The ICO audited AI recruitment tools and reported in November 2024 that some let recruiters filter out candidates with certain protected characteristics, and some guessed gender and ethnicity from names. It made almost 300 recommendations, including that candidates are told clearly how their information will be used. For a firm hiring a few people a year, that is a lot of care for little time saved.
How to know whether it worked
Pick one job and measure it before you start.
- Time it for two weeks. Hours a week, and who spends them.
- Run the AI version for four weeks. One person, one job, a business account.
- Count the checking. Include every minute spent reading, correcting and fixing, because that is where savings tend to disappear.
- Compare, and write it in one line. “Supplier invoices: from six hours a week to two and a half.” That sentence is your business case, and it is yours rather than a vendor’s.
- Keep it or drop it. Then pick the next job.
The six and two and a half hours there are invented to show the format. Your numbers will differ, and they are the only ones that count.
Where to start
If you would like a straight opinion on which one or two of these jobs are worth trying in your business, and which to leave alone, book a 30 minute call. We will tell you plainly. Any piece of work that follows is agreed in writing first.