AI-Assisted Delivery & Requirements

ERP discovery generates a lot of words. A two-hour workshop with five stakeholders produces dozens of pages of transcript, hundreds of half-formed requirements and an exhausted business analyst trying to turn it into a backlog. The traditional path: a week of writing up, a round of client validation, a round of correction, then another week of rewriting. By the time the backlog landed, the workshop was a fortnight old and the stakeholders had moved on.

We rebuilt the front of the process around voice capture. Workshops were recorded with consent. AI transcription gave us clean, timestamped text within minutes of the session ending. That alone removed the longest delay in the old workflow, but the real leverage was in what came next.

Prompt-engineered structuring took the transcript and produced a first draft in Gherkin form: given, when, then. The structure forced clarity. If a requirement could not be expressed as a behaviour with a trigger and an outcome, it surfaced as a question rather than a vague paragraph. The business analyst became an editor rather than a transcriber.

Acceptance criteria were generated alongside each story, drawn from the same transcript with explicit citations back to the source. Traceability was the non-negotiable: every story carried a link to the workshop minute it came from, every acceptance criterion carried the stakeholder utterance that justified it. Client validation became a comparison exercise rather than a memory test.

The output landed with the client inside two working days. Validation rounds got shorter because the stakeholders could see exactly which sentence had produced which story. Corrections were precise. The backlog made it into the delivery tool with traceability intact, and the delivery team had context the old process never preserved.

The lesson worth keeping: AI here did not replace the analyst. It moved the analyst up the value chain, from transcription to judgement. The structuring was repetitive enough that a model could do it well; the editing was nuanced enough that it still needed a human. Practical AI is the kind that gets used on the next project without anyone calling it AI.