Turn One Kind of Invoice or Form Into Fields
Extraction is one document type, one schema, and a person accepting low-confidence fields. Several document types and a write-back are the longer scope.
By Muhammad Saif and Aashan Khan · Techtribe co-founders
An inbox of PDFs is not a workflow. A workflow is a document of one type, a list of fields you already know how to book, and a person who accepts the fields before they become a number in the system you trust. The model can propose the fields. The model is fluent and occasionally wrong by one row, one digit, or one vendor. The accept button is how that stays a suggestion instead of a bill you paid twice.
This is the extraction feature named in the two-week MVP: a document or a message in, fields you can check out. It is also the document-processing path in add AI to an existing SaaS. Chat over a policy is a different feature, with citations and a refusal, written up in chat over your documents. Do not buy a chatbot because what you have is a pile of invoices. Buy the fields.
One document type
Pick the type by the layout, not by the department's name for the pile. "Invoices" from one vendor, with one template, is a type. "Invoices" from two hundred vendors is a research program. A form you designed, with the same boxes every time, is a type. A form customers photograph at an angle, with notes in the margin, is a harder type, and you should bring ten real photographs before anyone promises the boxes.
The schema is the fields a person on your team can already point at. Invoice number, date, total, currency, the vendor as you spell them in the ledger. Not "everything on the page." Everything is how a tax line gets booked as the total. Write the list. For each field, write where it sits on a good example, and what you do when it is missing. Missing is allowed to be a blank with a flag. Missing is not allowed to be a zero that looks like a real zero. Zeros pay the wrong amount.
A second document type is a second schema. A receipt is not an invoice with a friendlier name. A purchase order is not an invoice that has not been priced yet. If you need both, say so on the call. A short build does one. Sequencing is allowed. Stacking them into the first release because the inbox contains both is how the schema becomes a junk drawer.
The person and the confidence
Every field comes back with a value and a reason to look. Low confidence is a highlight, not a footnote in a log nobody opens. The person accepts the row, or they fix a field and then accept. The system of record updates from the accepted row, not from the model's first guess. On the first release, prefer that the update be a person's action in the existing tool, or an export they import, if a direct write is not ready. A direct write is an integration. It has failure modes: a duplicate, a vendor matched to the wrong vendor, a timeout that retries and posts twice.
Confidence from the model is not a probability you can take to an auditor. It is a hint the UI uses to decide what to highlight. Calibrate it on your examples. If the hint is always high, the highlight is useless and the person will stop looking. If the hint is always low, the feature is a slower form of typing. The eval is how you notice which one you built. Ten documents with the fields filled by a person who knows the job, compared with the model's fields. Two documents that should fail: a page that is not this document type, a total that is not on the page. A longer build includes this set when the fields will be depended on. A short build can start with the ten on a table and a human who is still the gate. Zero examples means you are judging the extractor by a demo page we chose. Do not do that.
The person needs a queue, not an email of JSON. The queue shows the document, the fields, the highlights, and a way to say "this is not an invoice." That last action is how the wrong type leaves the flow without becoming a vendor. A model that is forced to emit every field will emit them. The product has to allow a rejection.
Scans, tables, and the pages that lie
Some PDFs are text. The fields are almost a parse. Some are photographs of paper. Some are text layered on a scan that does not match the text. Some have a total in a table and a different total in a sentence above the table. The extractor has to prefer the place your team prefers, and that preference is a sentence you write down. "The amount due, not the subtotal, not the amount in the remittance advice if they disagree, and flag the disagreement." A model without that sentence will pick the more central number. Central is not owed.
Tables break. A column shifts, a currency symbol attaches to the wrong cell, a European decimal meets an American parser. The health of a batch matters: how many pages had no text, how many totals disagreed with a sum of lines, how many files were the wrong type. You are not commissioning a scanning bureau. You are refusing to book a soup. The same honesty shows up in document chat as citations and empty pages. Here the honesty is a highlight and a reject button.
Handwriting and stamps are a known ceiling. If your pile is mostly stamps over thermal paper, say so on the call. The two-week build will not become a custom vision system. It might still pull the printed total and flag the rest. If that is not the job, the job is a different quote. Bring the ugly examples first, not the clean PDF you wish vendors sent.
Where the accepted row goes
Name the system of record. An accounting tool, a database you run, a spreadsheet a specific person owns. "Our process" is not a system. If the accepted row's home is a spreadsheet, the first integration might be that sheet, and you should admit the sheet is the product. If the home is an API, that API is the one connection in a short build, or one of a few in a longer one. The call is where we count them.
The write is idempotent. Posting the same accepted invoice twice does not create two bills. The key is your invoice identity, not the model's enthusiasm. Failures are visible: the row stays accepted-but-not-booked, with the error, and a person can retry. A failure that looks like success is the silent-money class of bug. You already fear it in payments. Fear it in extraction the same way.
Do not let the model write the ledger "to save the person a click" on the first release. The click is the control. After a month of clicks where the person changes nothing, you have evidence for a narrower unattended path: this vendor, this field, above a threshold, still sampled. That evidence is a second scope. It is not a toggle you leave on because the demo was smoother without the queue.
Vendor matching is the field that creates the worst quiet errors. "ACME" and "Acme Inc" and a legal entity you have never seen. The first build can propose a match and force a person to confirm a new vendor. Creating vendors automatically is how you get duplicates that split the spend report. Confirming is slower for a week and cheaper for a year.
Two weeks versus four to six
A short build is one document type. One schema. A queue a person accepts. One connection if the accepted row has a single obvious home, or no connection if the home is a download. Your repo, a live URL, and support on what shipped. Not a catalog of vendors, not a dashboard, not a second document type. Bring a document in and the fields out. Bring "accounts payable transformation," and the call will tell you that phrase is a costume.
A longer build is several layouts of the same type, if they are truly one schema. An eval you can rerun. A way to see how often people change a field, which is the quality number that matters. A write-back you have named. Admin basics so a lead can see the queue. A few connections, not every system in the company. Design in your brand if this screen is customer-facing. Most invoice queues are internal, and internal can look plain. Plain is the right cut. Spend the scope on the disagreement flag, not on a typeface. Which of these two shapes you are in is what the call is for.
A continuing stream of new vendors, new forms, and new destinations is a team that stays, or a separate project if what you have is a backlog of document types with no single flow. Hire a dedicated AI team fits a backlog that does not end. A single form you designed does end. Do not book a standing team for one form. Say which backlog you have when you book.
If the documents are sensitive, say where they may be sent. A model API is a place. A region is a decision. A key in your account, with a spend alert, is the default. We will not quietly route customer invoices through a personal key. Say the constraint on the call, before a file is uploaded anywhere.
What this is not
It is not a promise that every total will be right. A serious build includes the eval and the highlight so you can see the wrong ones. Anyone who sells perfect capture from messy scans is selling a brochure. Ask them what the screen does when two totals disagree. If the answer is "it picks the right one," ask how a person notices the pick.
It is not chat. A box where you ask "what did we spend on shipping" is a question over records you already trust, which is closer to document chat or to a report in the accounting tool. Extraction is how a PDF becomes the record. Do not skip the record and ask questions of the PDF pile. You will get a fluent average of a soup.
It is not a rebuild of finance. The ledger, the approval policy, and who is allowed to pay stay in the system that already implements them. The model proposes fields. The policy decides whether they can be paid. A team that lets the extractor become the approval will not be able to explain a payment. You already have an explanation. Keep it.
It is not an excuse to boil the archive. Years of unscanned cabinets are a scanning project. We can start from the files you already have as PDFs this month. The cabinets are a vendor, a timeline, and a different budget. Say if the cabinets are the actual problem. We would rather point at the problem than sell a queue with nothing in it.
A short operating picture
Someone receives the file the way they already do: an inbox, an upload, a folder. The product pulls one type into the queue. The person accepts or rejects. The accepted row reaches the one system, or waits as a file they import. Rejected rows are visible so a second type does not vanish. Once a week, someone looks at the fields people changed. Those corrections are the next examples. A pipeline nobody corrects and nobody checks is either perfect or unread.
If your team is the person, they should see the queue in the first week of the engagement, not in a handover video. The standard is an AI product your team can maintain. Changing a field label they own is a safe edit. Changing the write into the ledger is an engineer's edit. Split those on purpose.
What to bring
Ten real documents, including two that are ugly and one that is the wrong type. The field list, with the page location on a good example. The name of the system the accepted row must reach, or an honest "a CSV is enough." The person who will live in the queue. A note on which fields are allowed to be wrong for a day and which fields are not, so the highlight matches the risk.
Leave the other document types in a folder marked later. If you cannot bear to leave them, bring three types to the call and expect us to say the work is longer, not a short flow with a wink.
If this is your app, send the repo.
A 30-minute scope call. We read what is live, name what is dangerous, and quote a fixed scope.
How Techtribe builds extraction
We take one document type and one schema, and we put a person between the model's fields and the system you trust. Low-confidence fields are marked. The wrong type can be rejected. A write-back happens only to a system you named, and a repeated post does not create a second bill.
That queue can be a short build. Several layouts, an eval, and a named write-back are a longer one. A backlog that does not end is a team that stays. We say which when we have seen the documents. The service page for a feature added to software you already run is SaaS development, and the pattern is add AI to an existing SaaS.
The scope call is 30 minutes with a founder. Bring ten documents and the field list. That is the meeting where the schema gets agreed, and where we tell you if the pile is not one document type.
Questions
What does a first extraction build include?
One document type, a fixed set of fields, and a screen where a person accepts or fixes each field before it is trusted. Low-confidence fields are marked. The model does not write them into the ledger on its own.
Can we start with every vendor's invoice?
You can start with one vendor format, or one form you designed. Every vendor is a catalog of layouts. That catalog is a longer project than the first schema. Bring it to the call instead of hiding it inside a short flow.
When does the result write back into our system?
After a person accepts, in a later step, or in the production scope if the write is one integration you have named. An unattended write on the first release turns a bad field into a booked number. The accept button is the product.
How do we know if one document type is enough?
Book the scope call and bring ten real files plus the field list. One document type and a human check can be a short build. Multiple layouts and a write-back into the system of record are a longer one. A cleanup of a bad archive is a different conversation, and we will say so.