What an Agentic Workflow Audit Actually Looks Like

"We should be using AI" is where most automation projects start, and it's why most of them fail. The useful question is narrower: which specific steps in your operation should an agent own, which need a human, and which should stop existing? Answering that is what a workflow audit is for.

Step one: map how work actually moves

Not how the org chart says it moves. In the discovery session I walk through real recent work, an actual client engagement or an actual month-end close, and trace every handoff: the copy-paste from email to spreadsheet, the "quick check" that takes forty minutes, the report nobody reads.

The output is a current-state map. It's usually the first time anyone has seen their operation drawn end-to-end, and it's uncomfortable in a productive way.

Step two: the automation matrix

Every step lands in one of four boxes:

• Agent-owned: high-volume, rule-describable, verifiable output. Contact enrichment, triage, first-draft reports.
• Agent-drafted, human-approved: anything client-facing or judgment-heavy. The agent does the 80%, a human signs it.
• Human-only: relationships, negotiation, taste. Automating these is how you lose clients.
• Delete: the surprise category. Every audit finds work that exists only because it always has. No agent required, just stop.

Step three: projections you can hold me to

The audit ends with priority rankings and estimated hours saved per workflow, numbers written down so that a later build can be measured against them. If a £2,000 build can't credibly pay for itself, I'll say so in the document.

Why this comes before any build

Automating a broken workflow gets you a faster broken workflow. The audit is £750 precisely so the expensive decisions, the build or a full infrastructure install, happen with a map instead of a mood. Every system in my portfolio started as exactly this kind of map.