Built by Full Score
Tablature
Nothing is trusted silently.
Tablature turns the documents a business receives, invoices, receipts, orders, contracts and statements, into records your systems can use. Every value comes with the exact quote and page it came from. Documents that clear every check go straight through. Anything it isn't sure of waits for a person, with the reason in plain English.
It isn't a product we sell. It's an example of what we build for clients: fast where it's safe, and honest about where it isn't. It's accountability built into software: every value can answer for where it came from.
Why the name
Tablature is the music notation that shows a player exactly which string and fret each note comes from, not just the note. Tablature does the same for data: every value comes with where in the document it was found. It's also a quiet pun: documents in, table rows out.
Reading was never the hard part
Any AI model can read an invoice. The trouble is that it's confidently wrong some of the time, and gives you no way to tell which times. A wrong value that looks right goes straight into the books, and nobody finds it until it costs something.
So teams check everything by hand anyway, and the automation saves nothing. Tablature is built the other way round: show the source of every value, and if it isn't certain, it doesn't go out.
Speed and truth, in the same system
Speed comes from letting the clean documents through untouched. Truth is what makes that safe.
Speed
- Clean documents go straight through
- A document that clears every check publishes with nobody in the loop. People only see the ones that were held, and each one arrives with its reason.
- Each review leaves less to review
- Teach it a new spelling of a supplier's name once, and the documents waiting on that name clear themselves.
- Records land where you work
- Published records flow to your Postgres database, S3 or Azure storage, or a signed webhook, or out as a spreadsheet or through a read-only API. Nobody copies them across.
- A new field is a question, not a project
- Ask "what are the payment terms?" It drafts the field, tries it on five documents with the quotes it found, then runs it on the rest.
Truth
- Every value shows where it came from
- Each value carries its quote, its page and a confidence. Click the quote and it lights up in the document's text, beside the original file.
- An invented value catches itself
- Tablature looks for every quote in the document. A value whose quote isn't there is flagged, not trusted.
- Old and new never mix quietly
- Every document stays pinned to the rules it was processed under. When the rules or the model change, older results are marked stale.
- Confident mistakes still get checked
- A random 5% of published documents goes to a blind audit, so a wrong answer that looked right still gets a second look.
How a document moves through it
Documents arrive the way they already do
By drag-and-drop upload, or by email to the workspace's own address, where every attachment becomes a document. Mail from unknown senders waits in quarantine. A byte-identical copy of an earlier file is caught as a duplicate, even under a new name.
Read the whole thing, keep the whole thing
PDFs, Word, Excel, PowerPoint, email, HTML, text and images. Scans and photos are read by a vision-capable AI model. The full document is kept as the record of truth.
Decide what it is, and show why
An AI model names the type, such as invoice, credit note or contract, against written guidelines, with a confidence and the quote that convinced it. The filename is a clue, never the decider. If the filename and the content disagree, the document is held.
Pull the fields, each with its source
Counterparty, dates, number, totals, currency, parties and line items, whatever matters for that type. Every value comes with its quote and page. Line items have to multiply out and add up to the subtotal.
Match to your list, never guess
The counterparty is matched to your own list of companies or people: exact, cleaned up, a known alias, a short name or a close spelling. Two possible matches means a person picks. No match means "not found", never a silent blank.
One gate decides publish or hold
Every document goes through the same checks. It publishes only when the type is certain and nothing else trips. Otherwise it waits with a reason anyone can read, like "Counterparty not found in your entity list" or "The filename and the content point to different types".
A person checks the quote, not the whole file
The reviewer sees the original beside the fields. They approve, fix the type, edit a value, pick the right company from ranked suggestions, or mark it "Not needed". Every correction is recorded and survives reprocessing.
The choices that make it work
- Nothing is trusted silently
- Every answer carries a confidence and a verbatim quote. If the quote isn't in the document, the answer isn't trusted.
- A gate nothing bypasses
- One set of checks decides publish or hold for every document. Unless it is certain what a document is, it doesn't go out.
- Rule changes are trialled first
- A new version of the rules is run as a trial that shows, document by document, what would change before anything is applied.
- Your own deployment
- Each client gets its own database, file storage and configuration. Nothing is shared with anyone else.
What it isn't
Not "fully automatic"
It holds documents on purpose. A few more reviews is the accepted price of never publishing a wrong value.
Not an OCR engine
Reading is one step. The value is everything around it: the gate, the citations, the review, the versioning and the delivery.
Not a model that trains on your data
"Learning" means recorded corrections and approved aliases. No model is retrained on your documents.
Not tied to one AI vendor
The model is a setting, not a dependency.
Still retyping documents, or checking the automated ones by hand?
Talk to us. We'll walk you through Tablature on sample documents, and what it would take to build the same discipline around yours.
Start a conversation →