Document Processing
A 40-page feeder scan lands in the shared mailbox as scan0001.pdf, and it's actually nine documents. This agent splits it, rotates the pages that fed in upside down, renames each file to your convention, and pulls the fields you actually track — carrier, policy number, expiration, member ID, PO line — into the record they belong to. Every field it writes points back at the exact spot on the page it read.
Yours outright — no required subscription.
Every agent is built for your business — your systems, your approval chain, your way of working. We scope it on the call.
If this sounds too technical, don't worry. We take care of everything for you.
- Private VPS deployment. This agent runs on a server that belongs to you, not a shared cloud tenant. Your patient packets, your contracts, and your scans stay on your box, and you own the server, the code, and the keys.
- A technician agent alongside it. Every hire ships as a pair: the Document Processing Agent reading and routing the pile, and a VPS-technician agent on the same box keeping it patched, backed up, and monitored. You're not hiring one thing, you're hiring a small team of two.
- Document intake connection. Wired into wherever documents actually arrive — a shared intake mailbox, the copier that emails whatever's on the glass, the fax number, the drop folder people use — so nothing has to be forwarded to it by hand.
- Your field list, not a generic one. We configure it around the documents you actually process and the fields you actually track, including your file naming convention and the folder each document type belongs in. If a layout is one it hasn't seen, it goes to exceptions instead of getting a best guess.
- An exceptions queue as the front page. Every field carries a status: read clean, low confidence, unreadable, or missing from the document entirely. The ones it wasn't sure about are the easiest thing in the system to find, not something buried in a report nobody opens.
- Security hardening. Private VPN, firewall, and encryption configured on your server before a single document touches it.
- 14 days of priority support after launch, plus a live walkthrough of both agents — so you see what it reads, where it stops, and what your real mix of scans and faxes actually does to it.
Not sure it fits? Check fit in 90 seconds in the free assessment chat.
Who this is for
- You're the practice manager at a clinic where new patients fill out a six-page packet on a clipboard, and the front desk retypes carrier, member ID, and group number between check-ins. A transposed digit comes back as a denial six weeks later.
- You run a contracting shop where every sub emails a certificate of insurance, and somebody keys the carrier, the limits, the expiration, and whether you're actually named additional insured into a spreadsheet tab nobody has worked in weeks.
- You have signed contracts sitting in a folder where the term, the auto-renew, the notice window, and the rate card exist only inside the PDF — and the person who knew which client renews in November left in March.
- You have a shared folder with four thousand files named scan0001.pdf, a fax line that still receives real referrals, and no way to answer "did we ever get that one back" without opening things until you find it.
How it earns trust
The failure mode of AI on documents isn't that it misreads a character — people misread characters too. It's that a wrong member ID arrives looking exactly like a right one, and nothing looks wrong until the denial shows up. This agent is built on the opposite bet: every field it writes can be checked in seconds.
Every field points back at the page
Click the member ID, the policy number, the PO line, and you land on the page and the region of the original file it came from, with the raw crop shown next to the value it produced. If it can't show you what it read, it doesn't write the field.
Every match carries its reason
Not just "attached to Vendor #218" but "matched on EIN plus street address; legal name differs from the DBA on file." Deciding which patient, which vendor, which job a document belongs to is the judgment half, so a wrong match is something you can trace and correct, not a mystery you find months later.
The counts have to tie
Documents that arrived in a window, documents processed, exceptions still open. Any difference is either something it flagged or a bug — nothing gets to quietly vanish between the fax line and the record, and nothing enters your system of record without a named person approving it.
How much of your pile lands clean versus in the exceptions queue depends entirely on what your documents look like, and we'd rather size that with you on the call than promise a number up front.
Pairs well with
These share a workflow with this role. Tick any to add them to your setup.
I have a folder called Scans with four thousand files named scan0001.pdf. Can it do anything about that?
Going forward, yes — it splits multi-document scans, rotates pages that fed in sideways, renames each file to your convention, and files it against the right record. Your existing backlog is a separate conversation: it can be run through, but it's scoped, because four thousand files with no naming will produce a real exceptions queue on the first pass. We scope both on the 15-minute call, and a working setup is typically about seven days from kickoff.
It comes in as a fax. A picture of a fax, honestly. Will it read that?
Often, and it will tell you when it can't. A fax that's been printed and rescanned twice is exactly where character-level misreads live — 0 for O, 1 for I, 5 for S, 8 for B — so those fields land in the exceptions queue with the raw crop shown next to whatever it thinks it read. Handwriting is the hardest case and gets flagged more often than it gets read. Nothing it can't read gets filled in with a guess.
I don't need it perfect. I need to know which ones it wasn't sure about.
That's the design. Every field arrives with a status, and low-confidence and unreadable fields go to a queue instead of into your system. The honest limit: confidence is a heuristic, not a guarantee. A clean-looking scan with one ambiguous character can read as high confidence and still be wrong, which is why every field is traceable back to the pixels it came from rather than just labeled with a score.
We found out the sub's COI expired in March. In July. Does this fix that?
It fixes the part where the date only exists inside a PDF. The expiration, the limits, the policy number, and whether you're named additional insured get pulled onto the record with the source crop attached, so the date lives somewhere you can sort by. It does not decide whether that sub is cleared to be on site, and it does not chase the renewal — pulling a date is extraction, judging an obligation is not. A person still owns that call.
What is it not allowed to do?
It never signs, countersigns, e-signs, or sends a document back to the other side. It never overwrites an existing record — a new value lands as a proposed change with the match evidence attached, waiting on a named person, so it can't silently update a patient's insurance or a contract's terms. It never alters, renames over, or deletes the original file; the source stays exactly as it arrived. It never files anything on your behalf with a payer, a government body, or a customer portal, and never invents a missing field to make a record look complete — a blank stays blank, named and assigned.
What's covered under the $1,997 setup, and what isn't?
One company, one system of record, one approval chain, and the document types we scope together on the call. Your existing approval chain stays yours — the agent adds a step, it never removes one. Several separate systems of record, or genuinely different handling rules per entity, is a scoped conversation rather than a price adjustment, and we'll figure out the right shape on the call.
Hiring more than one? A department on tap — the subscription puts a build team behind every request, agent after agent.