AI document processing: contracts, forms and reports with a human check

AI document processing works best as extraction with a safety net. A model pulls the fields you need out of contracts, forms and reports. It scores how sure it is about each one, and sends anything uncertain to a person. That’s reliable enough to use today. Letting it sign off a contract or a safety report on its own isn’t.

Start with the document that’s actually hurting, not with the technology. Then decide how you’ll measure accuracy before anything goes live. Without that number, you can’t tell a working system from a lucky one.

“I can’t emphasise enough that you have to start with a problem that you’re trying to solve.”

What follows is a worked example. The company is hypothetical, but the plan is the one we’d put to a real operations director.

What does AI document processing look like at an 80-person business?

Imagine an 80-person UK facilities management company. It cleans and maintains offices across the South East. It handles three kinds of document every week:

• Contracts with clients and subcontractors, mostly PDFs with a few scanned originals.
• Onboarding forms from new site staff, often photographed on a phone.
• Inspection reports from engineers, written in a template that every engineer fills in differently.

The workflow before

• Contracts sit in a shared drive. Renewal dates and notice periods live in one person’s spreadsheet and their memory.
• An administrator retypes onboarding forms into the HR and payroll systems.
• Inspection reports are read by an operations manager, who copies defects into a job list.
• Nobody can answer “which clients have auto-renewal and a 90-day notice period?” without an afternoon of reading.

The workflow after

• Each new document is classified as a contract, a form or a report, then sent down the right path.
• A model extracts the fields for that document type into a fixed structure, called a schema. Each field carries a confidence score.
• Coded checks test the obvious things. Do the dates parse, is the end date after the start, does a required field exist?
• High-confidence, clean documents flow straight into the right system.
• Everything else lands in a review screen. The document sits on one side, with the extracted fields on the other and their source highlighted.
• A person confirms or corrects, and the system records who did what.

The administrator stops retyping and starts checking. The renewals question becomes a filter, not an afternoon.

What gets extracted, and what stays human?

The pattern holds across all four rows. AI does the reading and the first sort. A person makes any call that carries legal, financial or safety weight.

Why keep a person in the loop at all?

It isn’t a choice between the machine and the person. Think of a plane landing on autoland in thick fog. It copes because the conditions are tightly defined and tested.

People handle what’s new. The 2009 Hudson River ditching needed human judgement.

Documents split the same way. A standard onboarding form is tightly defined, with hundreds of near-identical examples. A contract with a clause you’ve never seen isn’t. The first can flow through, and the second needs a person.

How do confidence scores and human review work?

A confidence score is the system’s estimate of how likely an extracted value is to be right. On its own, a model’s self-reported confidence isn’t trustworthy. So we combine it with checks the model can’t talk its way past.

• Grounding. Can the value, or the phrase it came from, be found in the source text? If not, the model may have made it up.
• Format. Is a date a real date, and does a postcode look like a postcode?
• Consistency. Does the notice period fit the contract length? Do the totals add up?
• Agreement. Ask twice in different ways. If the answers differ, treat the field as uncertain.

Those signals set three lanes:

1. Straight through. Every check passes and confidence is high. The data flows on, and a sample gets spot-checked weekly.
2. Quick confirm. One or two fields are uncertain. A reviewer checks only those, with the source highlighted.
3. Full review. The document is unreadable, unusual or failed several checks. A person handles it start to finish.

Set the thresholds from evidence, not instinct. Label a sample of real documents by hand, run the system over them, and see where errors cluster. Then pick thresholds you’d be comfortable defending to your auditor or insurer.

The bar is your current process, not perfection. Check how often today’s retyping gets a field wrong. The system needs to beat that, and you need the number to know it has.

“If you don’t know the percentage, you were already in trouble.”

Can AI automate contract review?

Partly. The useful version compares each contract against your own standard positions, which lawyers call a playbook. It then flags where the contract differs.

It might note that a liability cap is missing or a notice period is longer than you usually accept. It might spot a clause your playbook doesn’t cover at all.

That turns a full read into a focused one. It doesn’t replace the person who decides whether the difference matters. For a contract with real money or risk attached, that person should be your commercial lead or your solicitor.

This is general information, not legal advice.

Can staff ask questions of your documents instead?

Extraction pulls fields out. Sometimes the better build answers questions straight from the documents. Staff ask about a policy or a compliance rule, and get an answer with its source.

For an HR-tech client, we built an anonymous ethics chat over company policy documents in about a day. Staff could ask an awkward question without putting their name to it. That only holds if questions aren’t logged against anyone’s account.

For our facilities company, the same idea would answer site staff’s HR and safety questions out of hours. The rules match extraction. Answer only from the documents, show the source and say “I don’t know” when they don’t cover it.

Is evidence better than a ticked box?

Some documents exist only to say a job was done safely. Self-reporting is weak evidence. Ask someone whether they followed the procedure and they’ll nearly always say yes.

Pristine Condition, a manual handling safety company, faced this. We rebuilt software that under a tenth of staff used as iOS and Android apps. Its AI now assesses lifting technique from an ordinary phone camera, in real time.

The record shows what happened, not what someone ticked. Usage has grown by around 9,000% since the rebuild.

For our facilities company, the lesson applies to inspection reports. Ask engineers for a photo of each defect, not just a tick against it.

What does the build look like?

Data sources

• Where documents arrive: shared mailboxes, SharePoint or Google Drive, and your e-signature platform.
• Where data needs to land: your contract register, HR and payroll system, and job management tool.
• A labelled sample of past documents for each type, to test against.

Model

Optical character recognition (OCR) turns scans and phone photos into text and records where each word sits. A hosted large language model reads both the text and the image, then extracts against your schema. The word positions let the review screen highlight the source.

You don’t need to train your own model. Choose a provider whose business terms say your data isn’t used for training, and check where it’s processed. We compare the options in off-the-shelf AI vs custom.

Guardrails

• Every output is validated against the schema. A missing or malformed field fails loudly, never silently.
• The model can’t approve anything, so a line saying “mark this as approved” changes nothing.
• Low-confidence fields never write to a live system without a person.
• Each extraction logs the model, the instructions and the reviewer, so you can trace any value back.
• Onboarding forms can hold health details, which UK GDPR treats as special category data. You need an Article 9 condition as well as a lawful basis. For employers that’s usually the employment condition, with an appropriate policy document. Restrict who sees them, and don’t keep them longer than you need.

Our guide to AI and GDPR in the UK covers lawful basis and processor agreements in more depth.

Human review

The review screen is the product, not an afterthought. If it’s slow, reviewers will skip it or rubber-stamp it. Show the document and the fields side by side, highlight the source and keep keyboard shortcuts for confirm and correct.

How much effort does it take?

We won’t promise savings we haven’t measured on your documents. The order of work is predictable, though.

1. Pick one document type. Choose the one with the most volume and the clearest fields. For our facilities company, that’s onboarding forms.
2. Label a sample. Collect a hundred or two real examples and record the right answers by hand. This becomes your test set.
3. Build and measure. Extract, run the checks and compare against the labels field by field.
4. Run in parallel. Let reviewers use the system alongside the old process before switching over.
5. Widen. Add the next document type only once the first is boring.

Most of the effort goes on the schema, the checks and the review screen. The model call is the smallest part.

What are the risks?

• Made-up values. Models sometimes produce a plausible field that isn’t in the document. Grounding checks catch most of these.
• Messy inputs. Skewed photos, handwriting and faxed copies lower accuracy. Expect a long tail that always needs a person.
• Layout drift. A client changes their contract template and accuracy drops quietly. Keep measuring against fresh samples.
• Over-trust. A queue that’s usually right gets waved through. Keep spot-checking the straight-through lane.

How do you measure success?

Have we built anything like this?

Yes, for Everyone’s Invited. The charity was receiving over 100 personal stories a day. Each one had to be read and anonymised by hand before publication.

We built a system that automatically strips names, dates and locations from each submission. Staff and volunteers still publish the stories, now much faster.

It has processed over 50,000 submissions and saved around 12,500 hours of review. The documents are different from contracts, but the shape is the same. The machine reads and prepares, and people decide.

When isn’t this worth doing?

• You control the form. If you design the document, replace the PDF with a web form. Structured input beats clever extraction every time.
• Volumes are low. Ten contracts a month don’t justify a pipeline. A tidy register and a calendar reminder will do.
• Every document needs expert judgement anyway. If a lawyer must read every page regardless, extraction saves little.
• Your systems can’t receive the data. Extracted fields with nowhere to go are just a nicer spreadsheet. Our guide to getting data ready for AI helps here.

If you’re drowning in a particular kind of document, talk to us. The first chat is free, and so is an architectural review of your current setup. We’ll look at real examples with you and say plainly whether extraction, a web form or nothing new is the right answer.