AI in the front office: what actually works

The pitch deck answer to "AI in dentistry" right now is a chatbot that answers patient calls. Some of those products are fine. None of them have moved a collections ratio or an AR-over-90 bucket for a practice we've audited. The version that moves real numbers is smaller, less glamorous, and looks more like a really opinionated assistant than a talking head. Here's how we designed Remy.

The rule we wrote on day one

Remy doesn't generate prose. It calls tools.

The single design decision that defines Remy is that the model isn't trusted to invent facts. Every answer it gives is the output of a tool call against your real data — a SQL query, an OD API read, a Stripe receipt, an EOB parse — and every citation it shows is a row you can click. If the model wants to say "your collections ratio dropped 4% last quarter," it has to fetch the rows, do the math, and pin the source. If it can't cite a row, it can't make the claim.

That single rule eliminates 95% of AI failure modes in this category. The model can't hallucinate a number because the number is read from your DB. It can't make up a procedure code because the code came from the ClaimProc row it just queried. It can summarize, surface, and explain — but the facts are pinned.

The tools Remy actually has

Sixteen of them. The smallest possible surface.

Remy is wired to a deliberately small toolbelt. We resisted the urge to ship an open-ended "ask anything" surface. Each tool is single-purpose and audited:

• · query_kpi — read a named KPI for a date range, optionally per provider / practice.
• · top_accounts_by_age — return AR rows older than N days, sorted by dollar amount.
• · explain_anomaly — given a flagged KPI shift, fetch the underlying rows and attribute the change.
• · get_eob_for_claim — render the EOB next to the posting it produced.
• · draft_appeal — given a denied claim ID, draft an appeal narrative citing the chart and the correct attachments.

Eleven more tools follow the same shape. Each is read-only by default; the small set that writes to OD (post a ClaimPayment, file an appeal, retract a posting) requires the user to click an explicit confirm. The model never writes on its own initiative.

Two things that didn't work

The dead-ends we hit so you don't have to.

Open-ended chat as the entry point. Our first prototype had a single chat field — "ask Remy anything." Owners didn't use it. They didn't know what to ask. The dashboard with surfaced anomalies and a "dig into this" button per card outperformed the chat field by a factor of about 6× in week-1 retention.

Free-text summaries of EOBs. The model was happy to write you a paragraph about an EOB. Nobody wanted a paragraph. They wanted the structured table next to the posting it produced, with the rule that fired highlighted. We replaced the prose summary with a structured side-by-side and the page got materially more useful.

What this looks like in practice

When Insights detects an anomaly — say, a 4-point drop in hygiene reappointment over the last two weeks — it surfaces a card. The card has a "why?" button. Click it and Remy runs three tool calls: pull the hygiene visits for the period, compare against the prior 8-week baseline, attribute the change by provider and weekday. Then it composes a one-paragraph explanation citing the rows it pulled. The whole interaction takes about four seconds.

That's the version of AI in the front office that actually works. Not a chatbot. A purposeful, narrow, tool-using assistant wired to your real data, with citations on every claim.

The model isn't the product. The tools are the product. The model is the interface.

Remy ships with the Bundle and with Insights Multi+. We retain enough of the conversation to power the "show your work" panel and nothing else; Anthropic doesn't train on Pulse customer data via our Claude API contract.