The practice whose software had stopped telling it the truth

The problem that started it

Like most independents, this practice competes on care rather than price, which means its most valuable asset is the roughly 7,700 patients already on its books. The project did not start as a fascination with AI. It started with the honest maths of hiring: not that staff cost money, but that the work waiting for them was admin. I did not want to pay good people to chase lab orders and untangle lens pricing errors. I wanted everyone on the floor facing patients.

[FILL: years the practice has been established, staff count, one line of character]

The audit finding nobody expected

The build began as an automation project. It became an audit discovery. Reading the practice management software's underlying data directly, rather than trusting its dashboard, found recall flags that a routine housekeeping setting had switched off. No report surfaced it. No dashboard flagged it. The dashboard said everything was fine, and only the underlying data said otherwise.

The flags were reset, the affected patients went back into the normal recall cycle, and the check now runs automatically so it cannot drift again. The size of that gap is a number for the audit call, not for a web page.

[FILL: recall blitz outcome, once run: contacts made, bookings recovered, estimated revenue recovered]

The ad account that was blocking its own buyers

The same forensic pass ran through the practice's Google Ads account and found three stacked faults invisible from the surface: the conversion tag and the conversion action disagreed on an event name, so conversions were not being counted; the mobile booking page was missing its tracking tag entirely, a confirmed platform bug rather than a setup error; and the campaign carried 498 negative keywords, including "varifocal", on a varifocal campaign. The practice was paying to advertise and paying again to block its own buyers.

What runs now

As of August 2026, nine agents run live inside the practice, built on Make.com, GoHighLevel, Slack and Notion. A patient response agent answers routine enquiries, order status questions and booking queries at any hour, with anything clinical or complaint-shaped routed straight to a person. A reception agent lets staff manage orders, tasks and patient notes from Slack in plain English, with every action confirmed by a button press rather than free text. A voice receptionist has answered the practice phone since 23 August: 91 calls in its first nine days on the line, at a peak of around 25 a day, with 96% positive sentiment. A daily brief posts the day's open tasks and numbers to the team every morning.

[FILL: median patient response time before vs after, after-hours messages answered per week, admin hours saved per week, once the four-week baseline exists]

What the system never touches

Just as important is what it does not do. The practice management software stays. Clinical records, prescriptions, sight-test data, medical notes, stay guarded inside it, exactly where they have always lived. The recall audit read scheduling flags, never medical records. Crewley runs the operational layer only: messages, bookings, orders, recalls, reporting. Read the full detail on how it works and the Honest AI position.

What went wrong, and was fixed

Early on, the patient response agent was asked whether the practice stocks Oakley frames. It confidently said no. The practice does. The message reached a real patient before anyone approved it. That incident did not get a quiet prompt tweak. It became a standing rule: no agent may state a fact it cannot source, and no agent may claim an action happened unless the system has independently verified that it did. The fix ran a five-stage engineering gate, evidence, specification, roughly 350 logged test runs, independent adversarial review, isolated deployment, and went live, certified, within 48 hours. First day live: zero errors.

An estimated 50 to 80 hours of this project, a reconstructed estimate rather than logged time, went into testing, certification and investigation. That is what a practice is actually buying: You are not buying the automations. You are buying the faults that were found and fixed before you got them.

The uncomfortable maths

The multiples already run everything described here. They call it head office. Behind every large chain branch sits a team doing recall management, response handling and data hygiene that no independent could staff alone. Counting the roles this estate now covers, round-the-clock patient response, recall monitoring, document processing, daily reporting, comes to six to eight roles' worth of coverage, not six to eight salaries, running for less than the cost of one part-time hire.

For decades the only route to practice ownership with a head office behind it was the franchise model, paid for with a share of profits and a ceiling on independence. This estate is the head office. Ownership, independence and profits stay exactly where they belong.

In their own words

[FILL: quote from practice manager or front-desk staff, 2 to 3 sentences, on what changed day to day]