The Honest AI position

Three tiers, stated as a trade-off, not a spec sheet

Voice AI is sold to a practice as one of three tiers, chosen by the client, with the trade-off said out loud rather than buried:

• Safety net. AI answers only what a human structurally cannot: after-hours, overflow, an engaged line. The counterfactual is voicemail, so this tier is recommendable without caveats. This is the Winstanley configuration.
• Filter. AI removes time-wasting contact so humans can spend calls on the relationships that need a person.
• Front door. AI answers everything. For practices that want the highest possible answer rate over warmth, sold honestly: callers will know, and some will mind.

The positioning underneath all three: Crewley does not sell "AI answers your phone." It sells "AI protects human conversations." Automation takes the repetitive contact off a team so the calls that need a person get one. As voice technology improves, tiers two and three improve with it. Nothing about tier one needs walking back.

No undisclosed AI calls, ever

This is a standing policy across every client, not a preference. Every voice agent identifies itself, and the recording announcement is said in the practice's own words, in the greeting, every time. Winstanley & Son runs tier one. A practice that competes on care should automate everything behind the front door, not the front door itself, and everything behind it.

What the agents refuse to guess

The system is built to refuse rather than invent an answer. In daily use at Winstanley & Son this shows up as concrete, specific behaviour rather than a general promise: two patients sharing a surname produce a question, not a guess, at which patient was meant; an order it cannot match is refused with "please retype the order in one message" rather than created against the wrong record; a reply given in the wrong place gets a plain "nothing has been actioned" rather than silence that looks like success; and if the system cannot be sure something was saved, it says so, "check GHL before telling the patient," rather than claiming it worked. The voice receptionist follows the same rule on the phone: it takes a message rather than guess at unlisted hours, quotes no price beyond the exam fee, and never implies it is human.

The data boundary

Clients keep their existing practice management software. Clinical records, prescriptions, sight-test data, medical notes, stay inside it and never leave. Crewley runs the operational layer only: messages, bookings, orders, recalls, reporting. Full detail is on how it works.

The medical-device line

Software that transcribes, drafts or summarises for a clinician who then approves it is generally not a medical device under UK rules. Software that triages symptoms, supports a diagnosis, or makes a clinical recommendation can be, which brings UKCA marking and MHRA registration into play. The voice agent's clinical rule is written to stay on the safe side of that line: if a call turns clinical, it stops, takes a number, and tells the caller to contact 111 or A&E rather than attempting to triage anything itself. Any future feature that starts interpreting symptoms would move the whole build into a regulated category, which is why that line is treated as fixed rather than a target to edge towards.

No UK body currently licenses or approves the use of AI on patient records generically, including for AI scribe products used across the NHS. The honest question is not "are we approved," it is "what evidence could we show, if asked, that this was done responsibly." That evidence, a signed DPIA, a documented sub-processor chain, a clear retention schedule, a named human in the loop, is being built as part of the product itself, not treated as paperwork sitting alongside it.

Why the demo must never outperform the product

A demo built to impress and a product built to be trusted are different things, and the moment they diverge is the moment a practice stops believing either one. Every refusal shown above, the ambiguity questions, the "nothing has been actioned," the willingness to say "I'm not sure," runs the same way in a demo as it does live at Winstanley & Son, because it is the same system. We publish our failures as well as our numbers, including the Oakley incident described in the case study, on the same principle: credibility built on a demo that quietly behaves better than the real thing collapses the first time a client notices the gap.