A two-week AI audit that recommended against most of it

The practice knew a large share of its work was repetitive and suspected AI could help, but had no way to tell which part. In immigration work the value and the risk are distributed very unevenly across what looks like one process: some steps are pure clerical drudgery, others are regulated advice that an unlicensed system must not touch.

In Canada, providing immigration advice for a fee is restricted to licensed representatives. Any automation in the enquiry path has to be architecturally incapable of giving advice, not merely instructed not to — which rules out the obvious build and changes what the right one looks like.

Two weeks. We reconstructed the workflow as it actually runs rather than as documented, then scored every step on two axes at once — hours released against regulatory and reputational exposure. A single automation-potential score would have ranked the highest-volume step first, which was exactly the wrong answer.

A workflow map with volumes and time costs, a two-axis scored assessment of every step, a sequenced and priced roadmap, and an explicit list of what we recommended they not automate, with reasons.

Our initial scoring used a single automation-potential score, which ranked first-contact eligibility enquiries top on volume alone — directly against the licensing boundary. We rebuilt the scoring on two axes with risk as a first-class dimension, which reordered the roadmap substantially and changed what we recommended building first.