The problem Care exists for
Your prices move. A package changes what's in it. A test becomes available at one site and not another. Someone asks a question nobody ever wrote down.
The assistant announces none of this. It carries on sounding exactly as confident as it did in the demo, and starts being wrong to people who believe it. That failure doesn't arrive as a complaint — it arrives as an appointment that didn't happen.
Four things are included with every assistant I build. They aren't add-ons and they aren't a support plan. They're the loop that keeps it true.
The four steps
Every conversation is kept and read against what your clinic actually does. Alerts reach a named person at your clinic with the conversation attached, not a ticket number. You are never relying on a patient to complain, because most of them don’t — they just book somewhere else.
You can stop the assistant yourself, from the admin screen, at any hour of any day. It also stands down on its own when it recognises it is out of its depth, and I can stop it if something is wrong across the board. Containment never depends on me being awake — my response time is about repair, not about stopping.
When it gets something wrong, your staff correct the reply in plain English and it takes effect the same day. No developer, no redeploy, no editing prompts or flow diagrams. Self-service editors don’t fail because editing is hard; they fail because nobody sees the wrong answers, so nobody knows what to edit. Here the correction starts from the wrong answer itself.
Every week I work through the corrections. The ones that are genuinely your policy get written permanently into the assistant’s rulebook; a one-off answer stays as guidance rather than becoming standing policy. Each one becomes a test case that has to pass before anything is deployed.
Two of these run on their own. The other two are a person reading what your assistant told your patients, and deciding what to do about it. Staff onboarding and training happen once. This doesn't stop.
Take any one of the four out and the others stop compounding. Monitoring without a pause means watching it be wrong. A pause without correction means an assistant that just hands everything back to your front desk. Correction without testing means fixes that quietly regress. Testing without monitoring only ever tests for what you already knew about.
What the loop actually produced
On a live deployment covering four locations, June to August 2026:
A good share of those were facts the assistant already held and failed to use. Nothing broke, nothing was flagged, and the only reason anyone found out is that a person read the reply.
An assistant that generates no corrections is not a better assistant. It's an unread one. The question worth asking any vendor, including me, is: when it tells a patient something wrong, how will I find out?
Which tier
Care, S$300 per site per month. Daily monitoring with alerts to your named contact. You stop it yourself, instantly; I respond within four business hours. One update window a month, up to thirty minutes of changes. Every change tested before it ships.
Care Plus, S$900 per site per month. Everything above, plus a weekly read of what it actually told people with corrections applied, two hours of changes a month sent in any time, a one-business-hour response, and a monthly note of what changed and why.
Care suits a Chat build running on its own. Care Plus is required once Connect is live, or from three sites up — once the assistant writes into your records, a next-business-day answer is the wrong commitment. Both run twelve months, billed monthly in arrears, and both carry a break clause the other way: if I miss the response times twice in a quarter, you leave on thirty days' notice.