Where to start with agentic AI in Tax
Most Tax functions approach AI the same way: a broad tool rollout, a few enthusiastic pilots, and a slide deck promising transformation. Six months later, usage has flattened and no one can point to a workflow that actually changed. The technology was never the problem. The starting point was.
Start with a workflow, not a tool
Agentic AI earns its keep when it owns an end-to-end piece of work — gathering inputs, applying rules, producing a reviewable output — not when it answers ad-hoc questions. So the first question isn't "which model?" but:
- Which recurring workflow consumes the most senior time?
- Where is the input structured enough to reason over reliably?
- Where is a human review step natural, so mistakes are caught cheaply?
Pick the workflow that scores well on all three. That's your first agent.
Design for control from day one
In Tax, an answer you can't defend is worse than no answer. The agents that survive contact with a real team are the ones built with control in mind:
- Traceability — every output links back to the source it used.
- Review — a person signs off before anything leaves the team.
- Boundaries — the agent operates on a defined scope, not "everything."
Measure the thing you said you'd improve
Before you build, write down the metric: cycle time, rework rate, or hours returned to the team. After you ship, measure it. An agent that doesn't move a number you cared about beforehand is a demo, not a capability.
That's the whole method: one workflow, controls first, a metric you committed to. It's deliberately unglamorous — and it's why it works.
← All insights