How to build a private practice's management reports with AI

TLDR: A practice runs on one page a month: the schedule block, the economic block from a ledger like Xero, and the patient experience block. Claude drafts the reading and Gamma turns it into slides when someone else is in the room, but the report is worthless until it produces three decisions with names on them. None of this measures medicine. It measures the business around it.

The question nobody can answer

Someone asks, on the way out of the break room, whether the Tuesday afternoon block is still worth keeping open. Three people give three answers. One remembers it being dead all January, one remembers a run of new patients, and the third thinks the problem is that the referrals stopped. Everybody is sincere and nobody is citing anything.

That conversation repeats in every practice, and it is rarely a data problem in the way people assume. The numbers exist, spread across the scheduling system, the ledger and a survey inbox. What does not exist is one page where they sit next to each other once a month, with last month beside them, so the argument becomes short.

The reason most practices never build that page is that the first version takes an afternoon and the second one takes another afternoon, and by the third month it quietly stops happening. The build that survives is the one where nothing is assembled by hand.

Pick the ten numbers that decide something

Ten is not a magic number, and the exact count matters far less than being willing to take things off. A page nobody finishes reading decides nothing, whatever is on it. The test for each candidate is narrow: name the decision that would change if this number moved. If nothing comes to mind in a few seconds, it is a number you would maintain forever and use never.

Every figure carries its comparison beside it, previous month and the same month a year ago. A lone number invites a story, and the story is usually the one the person telling it already believed. Two numbers side by side end the argument faster than any amount of discussion.

The schedule block belongs to the appointments system and is defined there, so take those numbers as given rather than recalculating them here. The economic block comes out of the billing circuit, and the experience block out of the patient surveys you already send.

Give every number a system to come from

This is the step that decides whether the page still exists next spring. Each number needs a written recipe: where it comes from, what filter produces it, and who owns it when it looks wrong. That recipe is a paragraph in a shared document, not a project.

Xero starts at $25 a month and produces the whole economic block on its own, provided the tagging was done during the month-end close rather than invented at reporting time. That is the dependency worth naming out loud, because a reporting problem is almost always a bookkeeping problem wearing a different hat.

Any number you calculate by hand each month is a signal, not a chore. It means a system upstream was never finished, and building that system beats decorating the report. The profitability analysis and the cash flow forecast both depend on the same tagging, so the work pays three times.

Get a reading, then check it

Numbers without a reading decide nothing. Claude has a free tier and paid plans from $17 a month, and the useful prompt is narrow: here is the page with comparisons, tell me the three things that changed enough to matter, what the data itself suggests as an explanation, and what looks like it needs a decision this month.

What comes back is a draft, and the draft is where your knowledge does its work. The model can see that new patient bookings fell and that the ratings dipped in the same month. It cannot know that a hygienist was out for three weeks, and it will happily propose a plausible explanation that happens to be wrong. Correcting that is the job, and it takes minutes because the arithmetic is already done.

Gamma has a free tier and paid plans from $10 a month, and earns its place only when the page has an audience beyond you, such as partners, a practice manager or a lender who wants to see a trend rather than hear one.

The four ways this goes wrong

Receivables read as revenue. Billed and collected are different numbers, and a practice with slow payers can post a strong month while the bank balance tightens. Put both on the page, side by side, permanently. Practices discover this the hard way roughly once.

The occupancy number that flatters. Counting booked slots instead of kept ones hides no-shows and same-day cancellations, which is exactly the thing the number was supposed to reveal. Fix the denominator once, write it in the recipe, and never quietly change it, because a redefined metric erases your own history.

The number that survives its usefulness. A figure added during a bad quarter stays on the page for two years after the problem is solved. Review the page itself twice a year and take things off. A report only ever grows unless somebody is allowed to prune it.

The decision that comes back every month. The same item discussed in March, April and May is not a decision anyone is failing to make. It is a structural problem that the monthly meeting is the wrong venue for. After the second repeat, move it somewhere it can actually be resolved.

What to count to know it is working

Three numbers, and none of them are about the practice. They are about the report.

Count the minutes it takes to assemble the page, from opening the first system to having the numbers down. If that figure is not falling by the third month, something is still being done by hand. Count how many numbers on the page have no system behind them, and drive it to zero. Count how many of last month’s three decisions were closed by the time you met again, which is the only measure that separates managing from reading about managing.

Where the boundary sits

The software collects and the model drafts. A person signs, and signing means being able to say where each number came from and standing behind the reading attached to it.

That matters most when the page leaves the building. A number going to a lender, a partner or a prospective associate needs to be traceable back to the ledger entry that produced it, and “the AI wrote it” is not an answer anyone will accept. Keep the recipe document current for exactly this reason.

The meeting itself is thirty minutes with the page read beforehand. It opens by going back over last month’s three decisions, which is quick when they were written down with names on them, and what remains splits between what changed this month and the three decisions coming out of it, each with an owner and a date. That opening review is the whole difference between management and paperwork about management. The wider financial picture is in financial reports, and the rest of the back office in AI for health and wellness.

Frequently asked questions

Is this worth it for a solo practice?

In a reduced form, yes, and the build is easier because there are fewer systems to connect. Half a page, six numbers, and a meeting with yourself that you actually put in the calendar. Where it stops being worth it is a practice with one chair, no payers and no staff, because there the owner already knows the answer to every question the page would ask.

How is this different from what my accountant sends me?

An accountant’s statements answer what happened to the money, correctly and after the fact. This page answers whether the schedule, the payer mix and the patient experience are moving in the direction you want, in time to do something about it. They are complementary, and the accountant’s numbers are one of the three blocks rather than the whole page.

What if a number has no system producing it?

Leave it off until it does. An artisanal number takes the whole habit down with it when it stops being maintained, which is usually month three. Build the upstream system first and add the number when it comes out for free.

Should the AI reading go straight to my partners?

No. Send the version you have corrected, and be prepared to say which parts you changed. A reading that turns out to be invented costs more credibility than the report was ever going to earn, and it only has to happen once.

We tried a dashboard and nobody looked at it. Why is this different?

Mostly because it is not a dashboard. A dashboard is always available and therefore never urgent, and it demands that someone decide what to look at. One page arriving on a fixed day, with a meeting attached and three decisions expected, creates the obligation a dashboard never does. If you skip the meeting, this fails the same way, and there is no version of it that works without that half hour.