The problem, as you feel it
Documents arrive all day. Supplier invoices as PDFs, bills of lading as phone photos, client applications as scans, contracts as email attachments, receipts in a shoebox. Every one of them contains information that has to end up in a system, and the way it gets there is a person reading it and typing.
That person is expensive, gets interrupted constantly, and makes the occasional transposition error that takes three weeks to surface as a short payment or a mis-filed record. When they are on vacation, the pile grows. When the business grows, you hire another one.
Old-style scanning software promised to fix this and mostly did not, because it could read the characters but not understand the document. It needed a template per supplier and broke when the layout changed.
What has changed
AI models can now read a document the way a competent clerk does: find the vendor, the total, the line items, the due date, the PO number, regardless of layout, on a crooked phone photo, in a language the clerk does not speak. They can also do the second thing a good clerk does, which is notice when something is off. The total does not match the lines. This PO was already invoiced. The insurance certificate expires before the job ends. That is where the value is.
The pipeline I set up does this for every document that arrives:
- Capture from wherever it comes. An email inbox, a shared folder, a phone app, a scanner, a web upload.
- Classify. Invoice, credit note, application, contract, certificate, delivery note, junk.
- Extract the fields that document type needs, with a confidence score for each.
- Validate against your systems and your rules. Does this vendor exist. Does the PO match. Is the total within tolerance. Are the required signatures present.
- File and post. Into the accounting system, the client record, the job folder, the CRM, named consistently.
- Escalate exceptions only. A person sees a short queue of documents that failed a check, with the document and the reason side by side, and resolves each in seconds.
- Keep a trail. Every document, every extracted value, every decision, searchable.
Take a business like this one
The business in this section is a composite of the kind of company I talk to, not a named client. The numbers are the shape of the problem, not a case study.
Picture a small accounting and bookkeeping firm with six staff and about ninety small-business clients. Every month, clients send receipts and bills as photos, PDFs, and forwarded emails. Two staff members spend most of their time turning those into transactions in the clients’ books, chasing missing ones, and matching them at month end.
What I set up for a firm shaped like this:
- One intake address per client. Clients forward, snap, or drop documents and never think about it again.
- Classification and extraction of every document into vendor, date, amount, tax, category suggestion, and payment method, with the original attached.
- Matching against the bank feed in the client’s books, so a receipt that matches a transaction is attached and categorised automatically and one that does not is flagged.
- A month-end missing-documents list per client, generated automatically and sent as a friendly reminder with the exact transactions still undocumented.
- An exceptions queue for the two staff, showing only what needs a human: unreadable images, unusual vendors, amounts that do not match, category calls the rules could not make.
Now picture a second business, a freight brokerage with twelve people moving a few hundred loads a month. Carrier invoices, rate confirmations, proofs of delivery, and insurance certificates arrive by email and text all day, and a load cannot be billed until the paperwork is complete and consistent.
The same pipeline, tuned for them: every document classified and tied to its load number, proof of delivery checked for the right signature and date, carrier invoice checked against the rate confirmation, insurance certificates checked for expiry against the load date, and a load marked ready-to-bill the moment its paperwork is complete. The dispatcher’s morning becomes a short queue of the loads with a real problem.
In businesses shaped like these you should expect most documents to flow through with no human touch at all, the remainder to be handled from a queue in a fraction of the time, month-end or billing to happen days earlier, and the errors that used to surface weeks later to be caught on arrival.
What it costs to run
Reading a document costs a cent or two in AI usage. The pipeline runs on a small server or inside your existing automation tool for a few dollars a month. The setup is the work: defining your document types and rules, connecting your systems, tuning extraction on a few hundred of your real documents, and designing the exceptions queue so it is genuinely fast to clear.
What I need from you
A few hundred real documents of each type, including the ugly ones. Your rules, stated plainly: what makes an invoice approvable, what makes a load billable, what a complete application looks like. Access to the systems where the data lands. Then a person to work the exceptions queue for the first two weeks and tell me which exceptions should not have been exceptions.