AI agents for opticians, explained in plain terms

What is an AI agent, and how is it different from a chatbot?

A chatbot answers a question and stops. An agent is given a job with edges: it can create a task, move an order on, or reply to a patient, but only inside rules that are set once and never bent under pressure. The rule that matters most is what it does when it isn't sure. Crewley's agents are built to say "I'm not sure" and stop, rather than guess and carry on. If two patients share a surname, it asks which one you mean. If it can't be certain a booking was actually saved, it says so rather than claiming success.

That same discipline applies before anything is created, not just when something goes wrong. A dictated order is read back in full, patient, lens type, supplier, frame, and waits for a typed confirmation before it exists anywhere. Nothing is written on the strength of what the software thinks it heard.

What does "the operational layer" mean, and what stays in your PMS?

Say this plainly, because it is the first thing any practice owner should ask: your practice management system stays exactly where it is, and clinical records never leave it. Crewley does not touch the clinical record. It runs the layer above it, the operational work that happens around a patient rather than inside their notes: messages, bookings, orders, recalls, reporting. That boundary is stated to every prospective client early in the sales conversation, and it is the same boundary that governs what the software is legally allowed to do without becoming a regulated medical device (more on that below).

In practice this means the assistant answering your phone or texting a patient about their glasses is working from your diary, your order board and your recall list, not from a prescription file. Where a clinical judgement would be needed, a person makes it.

Which agents does Crewley run today?

The build behind Crewley is not a demo. It is the live automation estate running Winstanley & Son, a bespoke independent practice in Liverpool, and as of the end of August 2026 it had nine agents live in daily use. Here is what each one actually does, in plain terms:

• Reception agent: the one staff talk to in Slack. Creates and completes tasks from a spoken sentence, takes dictated orders and reads them back before creating anything, records notes against a patient or an order, and answers questions like "what's at the lab?" or "what's booked in tomorrow?" across all the practice's diaries.
• Button handler: the tap-to-confirm layer sitting behind every card the reception agent posts, so a job can be closed with a thumb as easily as a sentence.
• Customer agent: replies to patients by SMS and WhatsApp, live around the clock. It books, reschedules and cancels appointments from real diary slots, and holds anything about price, stock or a specific product for a person to approve rather than answering from a guess.
• Send handler: the single route any message to a patient goes out through, so there is one place a stop can be applied.
• Master halt: the practice-wide stop switch. One word in the right channel and every patient message pauses; another word resumes it.
• Reception digest: turns anything that looks urgent into a flagged task within the hour, and posts three summaries a day so nothing sits unread.
• Document intake: reads delivery notes as they arrive and moves the matching order on the lab board, so a scan does the job a phone call used to.
• Meeting capture: turns a recorded conversation into a written record the practice can search later.
• Voice receptionist: answers the practice phone line and, since 27 August 2026, produces a written summary and a task for every call, so the team can see what the phone did all day. In August it answered 91 calls, at a peak of around 25 a day, with 96% positive sentiment.

Three further agents are built and certified but deliberately switched off pending a cutover decision: see the spectacle collections page for what they do and why they are paused rather than abandoned.

Read more on how each piece works: the AI receptionist, appointment booking, patient recall, and lab order tracking pages each go into one part of this in detail.

How does an install actually get put into a practice?

The whole point of building this at a real, trading opticians first is that the faults have already happened once, on Winstanley & Son's own patients, and been fixed at the root rather than patched over. A second practice does not get a rebuild from scratch. It gets the Nth copy of a certified, frozen build, installed against that practice's own diary, order board and staff channel, then brought up gradually: preflight checks, an inactive import, every credential probed live, a fixture test, staged activation, and an automated health check running from day one.

What changes practice to practice is one configuration sheet: account IDs, pipelines, channels, tokens, staff names, and the practice's own knowledge-base facts such as opening hours. The behaviour underneath does not change per client; it is fixed, tested, and paid for once.

Why only one practice per area?

Crewley takes one practice per area on a renewing six-month exclusivity. The positioning is closer to a franchise than a software licence, except run in reverse: the head-office capability, the build, the fixes, the guardrails, is supplied, but ownership of the client relationship and the profit from it stay with the practice. A neighbouring competitor cannot buy the same install while yours is live.

What happens when something goes wrong?

Software that touches patient messages and orders will occasionally fail, and the honest answer is that this one already has, in the open, on a real practice. What matters is what each failure changed permanently. A hand-assembled change once took reception down for around ten minutes; the fix was to stop hand-assembling changes and clone only from certified copies. An unhandled write error once disabled reception for two and a half hours; the fix was a rule that every model-fed write gets a handler. A supervising watchdog now runs a read-only health check every thirty minutes across every channel and queue, to catch a fault before a member of staff does.

Every fault in the build's own incident log falls into one of five classes, and only two of them, missing guardrails and platform quirks, would ever reach a second practice; both are now fixed once and shipped as standard rather than re-discovered. A fault caught here does not travel; the fix does.

Questions practice owners ask