Data Rooms: software that does the reading before you ask

“And when did your LDL start climbing?”

You're across from a cardiologist who asks, almost in passing, “and when did your LDL start climbing?” You know the answer is in a lab report. You know you have six of them: PDFs from two different labs, a photo of a printout, one buried in an email from last spring. You own every number the doctor needs, and you cannot answer the question.

That folder is what I've come to think of as the shoebox. Nearly everyone has one, and it isn't only medical: claims, contracts, statements, the letter that proves you disputed the charge. Cloud drives solved storing all of this a decade ago. What they store, though, is files. A folder doesn't add up your numbers, notice that a result is stale, remember that an appointment needs a referral letter, or draft the appeal for a denied claim. Every one of those jobs still lands on a person who is usually stressed, often unwell, and rarely an expert in the paperwork.

The newer chat-with-your-documents tools are a real improvement on the folder, and they share one interaction model: you upload sources, then the model uses what you uploaded to answer the questions you ask. Google's own help for NotebookLM describes exactly that loop. The work of knowing what to ask, asking it well, and judging the answer still belongs to the stressed non-expert holding the folder. Data Rooms starts from a different premise: the software should do the reading before you ask, so the answer is already standing there, with every fact traced to the document it came from. To be plain about what it is: Data Rooms exists as interactive design-sprint screens with a sample room behind them, with no real users. First comes the reading, then where the room admits ignorance and stops short of acting, then the gap a folder can never show you.

The answer stands on the overview before anyone asks

In the prototype, a room is a bounded space for one area of your life: health, home, the business. You put documents in, and the room reads them into two things: structured data (the actual values, with units, linked to the document they came from) and a timeline of what happened when. Everything on screen is sample data from the demo room, built to make the interaction concrete; the numbers below are illustrative rather than measured.

The clearest artifact of that reading is the chart the demo room keeps on its overview. The room has already read every lab report it holds, so the chart stands there, captioned with its sources, before any question is typed.

Look at the source labels under the points. Two labs' PDFs, a photo of a printout, an email attachment. That mix is the honest shape of personal records, and it's also where a real version of this product would live or die. Pulling “LDL 142” reliably out of a crooked photo, and knowing it's LDL and not total cholesterol, in mg/dL and not mmol/L, is genuinely hard. The chart earns its place the way everything in the room does: every point traces to a document you can open, so checking the reading takes one click instead of a redo.

The room admits its backlog and holds its drafts

The design's answer to that difficulty is to admit what it hasn't digested. The demo room shows 81 of 84 documents structured, with the three unread ones named in plain view instead of quietly guessed at. A confident wrong number would be worse than an admitted gap, and the person supervising the room needs to see the backlog as plainly as the results.

Reading is the foundation, and agents in the room carry the work one step further. In the demo, an explanation of benefits shows a denied physiotherapy claim, so the room drafts the appeal: it pulls the denial code from the EOB, cites the visit note and the referral that support the claim, and leaves a draft for you to review and send. The lipid panel predates the statin, so the room proposes a follow-up in June and can book it when you agree. A reminder from this room knows why it exists, and shows you the documents behind it.

Just as important is where the work stops. The room reads, structures, reminds, and drafts; it does not diagnose, and it does not send. “Your last panel predates the statin” is a scheduling observation grounded in a document; whether the number is worrying is the doctor's call, and whether the appeal goes out is yours. In the sprint we kept returning to that line, because software that acts on your medical paperwork is only tolerable if every action stops one step short of a decision that belongs to a person.

A folder can't tell you what's missing

The feature I'd defend hardest is the one that sounds smallest. In the demo, the room flags that cardiology wants a referral letter before the April appointment, and the letter isn't in the room. Search can only find what exists. Finding a gap requires knowing what a complete record should contain and comparing the pile against it, which is exactly the judgment the reading makes possible.

The gaps are also where the money and the risk sit. A missing IP assignment surfaces in diligence, at the worst possible moment, priced accordingly. An audit with a missing filing stops being routine. The missing referral letter reschedules the specialist appointment six weeks out. In each case the document you don't have matters more than the hundred you do, and a folder is structurally incapable of telling you about it. A room that has read everything can hold your pile against the checklist your situation implies and name the one item between you and ready.

You supervise a clerk who already read the file

Put the pieces together and the room casts you differently than the chat tools do. The pile never sits inert waiting for a well-phrased question; the room has already gone through every lab report it holds, so by the time the cardiologist asks, the trend chart has been standing on the overview for months with each point captioned by its source document. You review a clerk's standing work: the facts, the timeline, the gaps. Asking a question becomes the exception, not the interface. What makes this possible is that reading itself got cheap: a model can take a crooked photo and an emailed PDF from two different labs and produce the same typed value, with units, linked to where it came from.

If you're building any agent-backed surface, the two habits worth stealing are the clerk's. Show ignorance as plainly as knowledge, the way the demo room names its three unread documents, because a supervisor needs to know the backlog. And attach provenance to every fact, so checking the work takes one click. Then stop where a clerk should: the room drafts the insurance appeal and leaves the sending, and the diagnosing, to the people those decisions belong to.

The same shape fits three more shoeboxes

None of this is health-specific. A startup's fundraising data room is the same pile in a suit: incorporation docs, the cap table, contracts, financials, and founders spending diligence answering questions whose answers are already in the room. A house purchase is the same pile with deadlines attached: disclosures, the inspection report, loan estimates, and a buyer who is the least experienced person in the transaction. Customer onboarding flips the direction: the customer drops in what they have, the room reads it against your intake checklist, and both sides can see what's processed and what's still needed.

One test to take back to your own shoebox

Data Rooms is a working example of the kind of software we describe in the bespoke SaaS post: designed from scratch around one real problem and what agents can now do with it, then run for the people who use it. What we'd validate next is the reading itself, on real piles: whether extraction from crooked photos and mixed lab formats holds up well enough to earn the standing chart.

And here's the test for any tool that promises to organize your shoebox, whichever one it is: the health folder, the deal room, the vendor-onboarding thread. Somewhere in your next appointment, audit, or closing, there's a document you'll be asked for. Would you know today if it's missing? Ten minutes with the folder open will tell you.