Britain Bought the AI. Nobody Checked the List.
Thirty-five per cent of UK businesses with ten or more employees were using at least one AI technology in June 2026, up from around twelve per cent in late 2023, according to ONS figures reported by TNW. Among those adopters, roughly seventy-seven per cent saw no change in revenue at all, and about twelve per cent could attribute a revenue increase to it.
That is a lot of software and a very quiet till. The instinct is to blame the tools, or the vendor, or the people who signed the contract. The more boring explanation is that most of these systems were pointed at a contact list nobody had looked at in years.
Adoption is wide. It is not deep.
The same ONS data puts the average adopter at 1.6 AI technologies, with only ten per cent describing their use as extensive. That is the shape of a pilot, not an operating model. One tool in marketing, one in support, a licence somebody expensed, and a quarterly review where everyone agrees it is early days.
Sales is the exception, and that is exactly where the problem bites. Salesforce’s State of Sales report for 2026 puts AI use at eighty-seven per cent of sales organisations, with fifty-four per cent of sellers saying they have used agents and nearly nine in ten planning to by 2027. Sales did not dabble. Sales handed over the outbound.
What the agent is actually reading
An AI sales agent does not know anything about your market. It knows what is in your CRM. Every judgement it makes about who to contact, what to say and when to say it is a function of rows somebody typed in at some point, often at speed, often from a photograph of a name badge.
Validity’s State of CRM Data Management in 2026, a survey of 500 marketing professionals published in August, found that just twenty-one per cent say their CRM data is very well prepared to support AI. Sixty-two per cent said their organisation had lost revenue directly because of poor CRM data quality. Nearly a third of teams spend six or more hours a week fixing and reconciling records.
The number that should worry anyone running a team: seventy-eight per cent of C-suite respondents said they had acted on an AI recommendation they later suspected was wrong because the underlying data was bad. Not ignored it. Acted on it.
Automation does not audit
A human rep opening a contact record from eighteen months ago applies a small, unglamorous check. The job title looks senior for that company. The mobile number has the wrong country code. The email pattern does not match the other three people at that account. The rep pauses, checks LinkedIn, fixes it or bins it.
An agent does none of that. It sends. At volume, politely, on a schedule, to whoever is in the row. The gap between eighty-seven per cent adoption and twelve per cent revenue impact is partly this: automation multiplies whatever you already had, including the rubbish.
The error is made before the CRM ever sees it
Almost every bad contact record starts in the same place, and it is not the database. It is the last moments of a conversation, when two people try to exchange details and both are already half way to the next thing.
Someone reads a phone number aloud in a loud hall and someone else types it into a phone. A paper card goes into a jacket pocket and surfaces three weeks later with no memory attached to it. A rep spells out their surname twice and the prospect still saves it wrong, so the follow-up lands in a search result rather than a contacts app. A badge gets scanned by a stand system that captures the job title the delegate typed when they registered, in March.
Nothing downstream fixes this. Enrichment tools guess. Deduplication merges two wrong records into one wrong record. The agent inherits all of it and sounds confident.
Fixing intake is cheaper than cleaning a database, and it is the part you actually control. If you work events and meetings on your own, the practical version is carrying a card that cannot be mistyped: get the HeyDrop app and share by QR, NFC, AirDrop or link, with the details arriving on the other person’s phone exactly as you wrote them, no app needed at their end. Going the other way, the AI Contact Scanner reads the paper cards and conference badges you collect and turns them into structured contacts, so the pile in your bag becomes rows rather than a weekend of typing. Contacts export to CSV and Excel, which is the honest description: a clean, CRM-ready file, not a magic sync.
For a team, the failure mode is different and worse, because it is inconsistent rather than merely messy. Six reps at a trade show produce six formats, three spellings of the company name and two versions of the job title taxonomy. Standardising that at the point of capture is an admin job, not a data project: set your team up on HeyDrop, create the cards centrally, and every contact the team collects lands in the same shape before it ever reaches the pipeline.
A contact audit you can run this week
None of this requires a data governance programme. It requires an afternoon and a willingness to look.
Pull a sample, not a report
Take fifty contacts your agent or sequencing tool has emailed in the last month. Not a dashboard, the actual records. For each one, check three fields: is the person still at that company, is the job title current, and does the email domain match the company’s live domain. Count the failures. Most teams doing this for the first time find a rate that stops the meeting.
Trace the worst ones backwards
For every record that failed, find out how it entered the system. Event scan, webform, list purchase, manual entry after a call, imported spreadsheet from a rep who left. You will find one or two sources producing most of the damage. That is your fix, and it is usually a process, not a product.
Put a date on every contact
If a record does not carry the date and context of its capture, your agent cannot weight it and neither can you. A contact from a conversation last month and a contact from a list bought in 2023 should not be treated as the same asset. Most CRMs will do this with one custom field.
Decide what the agent is not allowed to touch
Records older than a set age, or from sources that failed the audit, should be excluded from automated outreach until a human has looked at them. This is the single cheapest control available and almost nobody sets it.
The uncomfortable bit
The ONS numbers will be read as a story about AI being overhyped in Britain. That reading is convenient, because it puts the fault outside the building. The Validity findings point somewhere less comfortable: the technology is mostly doing what it says, and it is being fed by a process that was already broken when a human was doing the sending. The difference is that a human sending the emails felt the wince, and eventually stopped. An agent does not wince, and does not stop.
Before the next licence renewal, it is worth knowing what percentage of your contact records would survive a phone call. If the answer is unknown, that is the answer.
This article and its accompanying image were generated with the assistance of AI.