AI-driven data maintenance with Notion Custom Agents
Coverflex
2026
Coverflex is an insurance broker, and its health insurance catalogue is only useful if it’s right. Every month, Coverflex collects fresh quotes from insurers, and two AI agents compare them with the catalogue. They fix the small, low-risk differences on their own, bring the insurance specialists in for anything that matters, and learn from what those specialists decide. Designing it was mostly about deciding what AI gets to decide.

The problem
Health insurers don't offer APIs or public price tables. To get a price, insurance specialists fill in a form on each insurer's website, upload a spreadsheet with the list of people to cover, and get a quote back as a PDF. In a previous initiative, Coverflex automated that with scrapers, and turned those plans into a catalogue managers can compare.
But insurers change their plans: a price goes up, a waiting period changes, a limit moves from €20,000 to €25,000. Nobody sends a memo. And stale data is worse than no data. A manager comparing plans trusts that what they see is what they'll get.
The PDFs were already there
Coverflex already requested a benchmark quote for every plan, always for the same imaginary company, so prices could be compared. Each one came back with the insurer’s official PDF, and new ones could be requested every month. Could AI compare those PDFs with the catalogue, and flag what’s missing or wrong?
That became the sanity check: a monthly comparison between what insurers say today and what the catalogue says, with a person deciding anything that matters.
Turning those quotes into a monthly sanity check
Every month, an app on a server (Asclepius) requests the benchmark quotes and saves every price and PDF. About 30 minutes.
Then two custom AI agents in Notion, where the catalogue lives, take over. The Reviewer reads each PDF, compares it with the catalogue and proposes changes, split into low stakes and high stakes. About 40 minutes.
The Editor applies the low-stakes changes on its own. About a minute.
When a check needs a person, its status changes to “Needs review”, and a Notion automation sends the insurance specialists a Slack message. They open the table, decide on each proposed change, then click “Apply decisions”, which hands the check back to the Editor.
The Editor applies their decisions, and turns repeated ones into rules for next time. About 10 minutes.

Teaching AI to read carefully, ask when it matters, and remember
Reading every page, not skimming it. Left alone, AI skims. And insurance PDFs are tables, which is where AI gets confident and wrong: pull the text out and empty cells disappear, so a real number lands in the wrong plan's column. So the Reviewer looks at every page as an image, and every difference gets a second, sceptical check. It also had to learn that a value missing from a PDF isn't proof it should be deleted. Some insurers just don't print everything.
Letting AI change only what’s low stakes. The tempting answer is “everything it’s sure about”. But AI always sounds sure. So it changes only two things on its own: an empty number the PDF clearly shows, or a price that moved 5% or less. Everything else, like a price that rose 8% or a waiting period that changed, goes to the insurance specialists. Every proposal changes one thing, explains why in plain words, and can be undone. The scary part was never AI making a change. It was not being able to take it back.
Learning from the decisions people make. Each review starts fresh: the agents don’t remember past runs, so a past “no” never quietly hides a real problem. Their only memory is a page of plain-language rules. When people make the same call for the same reason three times, the Editor writes a new rule, so nobody has to make that decision again.
Designing a process, not an interface
Most of what I designed here wasn't an interface. It was a process: who does what, in which order, and when it's a person's turn.
Getting agents to hand work to each other, and to a person only when it’s their turn, was new to me. So was designing the Notion databases the whole thing runs on: one for checks, one for quotes and one for proposed changes, all linked to the plans they touch. Is that engineering work? Usually. But with agents doing the heavy lifting, a designer could build it.
What I took away
AI is great at the boring part: reading dozens of PDFs and noticing what moved. Designing this was mostly about drawing the line between what it does alone and what people decide, and making that line cheap to cross back.
A system that learns from people needs people to say why, not just yes or no. And an AI reviewer will never tell you everything’s fine: every run finds something slightly different to say. Part of the design is deciding when that’s worth a person’s time.
I think this is a small preview of how products will be built. Agents will read feedback, spot patterns and propose what to do next. When they've seen the same situation three or four times, they'll just act and let us know. When something is new or the stakes are high, they'll come back to a person with a plan, and the person will say "go ahead" or "go this other way". Teaching agents what matters, when to ask and how to learn from the answer is design work, whether the subject is a plan's waiting period or a company's next initiative.