AI with receipts: how we use models on client work

Our rule for AI fits in three sentences. AI drafts. People decide. Everything leaves a trail.

Models are useful on real client work. They are also confidently wrong often enough that nothing they produce should reach the public without a person who knows the subject. So we use them where a draft saves time and a person can check it, and we keep a record of who checked what.

Where AI helps

Research synthesis

After a round of interviews or usability sessions, a model can group notes into themes and pull out candidate quotes. A researcher checks each theme against the transcripts before it becomes a finding. The model shortens the sorting; it doesn’t decide what the research says.

Content inventories and migration at volume

Moving a site means knowing what is on it: thousands of pages and documents to classify by topic, owner, age and duplication, and old addresses to map to new ones. Models are good at a first pass across that volume. People review the classification, and every redirect is tested before launch.

Plain-language drafts

A model can propose a plainer version of a dense paragraph. It can’t tell you whether the people you serve understand it. A content designer checks that the meaning survived, and testing with the people who use the service settles it.

Alt-text drafts

Alternative text is where accessibility work piles up. Ockem Proof, the platform we built for document remediation, drafts alt text with a self-hosted model, so documents are not sent to an outside AI service. A draft is a starting point. A person decides, because only someone who understands the document knows what an image is there to say.

Code assistance

Our engineers use AI assistants to write and refactor code. Every change goes through the same review and automated tests as any other. Much of the code behind this website was drafted that way; a person reviewed and merged every change, and each one had to pass an automated WCAG 2.2 AA check.

Accessibility checks

Automated rules catch part of what makes a page inaccessible. Models can help sort findings and flag likely problems that rules miss, such as link text that makes no sense out of context. They don’t certify conformance. An audit still means people testing with keyboards and screen readers.

Search and answers over your own content

A model grounded in your own pages and documents can answer questions in plain language and show the sources it used. When your content doesn’t answer the question, the right response is to say so and point to a person, not to guess.

The guardrails

• Local or governed models. Client material goes only to models running on infrastructure we or the client control, or to services under terms the client has approved.
• No training on client data. Client content is never used to train a model, ours or anyone else’s.
• Human sign-off on anything published. Text, images, code and alt text are approved by a named person before they reach the public.
• An audit trail. For work a model touched, we can show which model was used, what went in, what came out, who reviewed it and when.

On this site, the same rules are built into the tools. In the content management system behind base22.com, a model can propose a title or tags for an article, but the proposal lands in the editor’s unsaved draft; it never saves or publishes anything. Tag suggestions only match topics that already exist, because inventing a taxonomy is an editorial decision, not a model’s. Every model call goes through one gateway, where it is rate-limited and metered.

What we will not do

• Publish model output that no person has read and approved.
• Put client data into a model or service the client hasn’t approved.
• Use a model to make decisions about people, such as who qualifies for a service or a benefit.
• Invent quotes, figures or sources, or let a model invent them for us.
• Call an automated scan an accessibility audit, or ship AI-drafted alt text as final.
• Hide it. We tell clients where AI was used in the work we deliver.

Why receipts

Public agencies and large organizations answer for what they publish. When a model helped, the question “how do you know this is right?” deserves a better answer than “the AI said so.” A draft, the name of the person who reviewed it and a date is that answer.

None of this makes the work slower in the places that matter. It moves the time from typing to judgment, which is where a senior team should spend it.