Digitalise first, then AI

The order

Most conversations with us start with AI. The first thing we do is look at where their data lives. If a job's record is a paper form, a WhatsApp photo and a spreadsheet row, there is nothing for a model to work from.

Digitalising comes first because it produces the one thing every automation needs: a single, structured record of what happened, captured at the point it happened, by the person who did it. Once that record exists, automation is a smaller step than most people expect. Without it, AI is a demo.

The order also protects the budget. Digitalising one process is a bounded job with a visible result: staff use the phone app, and the office sees the records arrive. Automation built on top of it is a second bounded job with its own result. Doing both at once, or the second one first, produces a long project whose outcome nobody can see until the end, which is how most of these projects fail.

What digitalising produces

• Structured data. Every field on the old form becomes a named value, not a pixel in a photograph.
• Provenance. Who entered it, when, and where.
• Completeness. Required fields are required, so the record is whole before it moves on.
• A history. Every job, every check, every approval, in one place and searchable.

For an 800-person engineering consultancy on our system, that is twenty document types and more than sixty live operations across Sarawak and Peninsular Malaysia, captured on phones, offline where there is no signal.

What becomes possible next

With structured records in place, three things become practical.

• Documents. The final report, certificate or plan is generated from the record in the company's own format. The parts that are calculation and layout are done by the system. The parts that are prose can be drafted by a model and checked by the person who signs.
• Plans. When the history of jobs is in one place, next week's schedule can be proposed rather than compiled.
• Answers. Staff can ask a question of the company's own records and documents, with the answer limited to what that person is allowed to see. A company-wide assistant of this kind went from idea to live for its first users in seven days on our platform, because the records, sign-in and permissions already existed.

The caution

A model's output is only as good as the record it reads and the person who checks it. We keep computed figures in the system's own logic, where they are exact and traceable, and use models for prose, drafts and answers, where a person reviews before anything is signed. We also keep the company's documents in the company's own store, governed by the same permissions that decide who may open them by hand. Features that skip either step produce confident documents that nobody can vouch for.

What to do this quarter

Pick the process whose documents take longest to produce. Digitalise its capture first. When the records are flowing, automating the document is a short, well-defined job, and the first draft of it can be running within days.