AI automation

Most automation projects fail on the first step rather than the last: the wrong process gets automated, faithfully. We start by working out where the time and the errors actually go, and only then build.

Finding what to automate

A short piece of analysis with the people doing the work: map the process as it runs now, not as the manual describes it, and find the places worth changing. Usually they are the dull ones — a form re-keyed into a second system, a queue nobody can see the length of, an approval that waits on one person's inbox.

• Steps repeated by hand because two systems do not talk.
• Work that cannot be tracked, so nobody knows what is stuck or for how long.
• Judgement steps buried under clerical ones, so the expensive people spend their day on the cheap part.
• Rules that live in someone's head rather than anywhere inspectable.

What we build

• Workflow and approval chains with an audit trail — the same engine behind our eForms and case management work.
• Extraction from unstructured material: tables out of PDFs, values out of reports, records out of documents that were written for people rather than parsers.
• Triage and routing — classifying what has arrived and sending it to the right queue, with a confidence threshold below which a human decides.
• Drafting assistance where a person still signs: prepared responses, summaries and pre-filled forms.
• Integrations, so the automated step writes back into the systems of record instead of a spreadsheet.

Keeping it defensible

For regulators and government the question is not whether a model is impressive, it is whether a decision can be explained months later. So: every automated step is logged with its inputs and its output, thresholds are explicit and tunable, a human checkpoint sits wherever the consequence is real, and the process runs on your infrastructure with your data. If the model is wrong, the audit trail shows what it was given and what it did.

How we start

Usually a two to four week analysis producing a shortlist of candidates, each with the effort to build it and the time it gives back, so the first thing built is the one worth building. This sits alongside our business applications work, which is where the workflow, eForms and case management engines come from.

One of four things we do under AI integration — the others being search and answers over the material these processes produce.