Workflow

Tax preparation is not a toy process. It has messy documents, client notes, prior-year data, regulatory consequences, and a human who signs at the end. That is exactly why it is a good first workflow. The corrections are frequent, expensive, and measurable.

The loop

The breakthrough was not a smarter model. It was the loop around it. A practitioner corrects a draft return. The correction is captured with its full context, from source file to extracted field to final filing. Recurring corrections become evaluations. The evaluations drive fixes. Ambiguous cases route back to a human instead of being forced through automation. The system improves because the experts do.

Proof

Across the network, the system processed about 7,000 complex returns. It saved practitioners roughly one third of preparation time, drafted returns with up to 97 percent accuracy, and lifted throughput by about half. Field completion climbed fast. At launch only 25 percent of returns reached 75 percent correct fields, and within six weeks 86 percent did. One senior accountant who spent 180 hours on preparation last year spent 15 this year, and used the freed time to advise clients.

Ecosystem read

Five battlegrounds move at once. The customer is the firm and its client, who now gets advice, not just a filing. The developer battleground is the improvement loop itself. The software vendors that win are the ones that expose their traces. The service partners, the accountants, become the feedback supply line, the most underrated asset in the system. The infrastructure battleground is the secure, traceable base without which the loop collapses.

The Commander’s move

Pick one high-friction workflow where expert corrections are frequent and costly. Instrument it before you scale it. Measure throughput, accuracy, review reduction, and freed capacity, not demos.

Independence note

The lesson has nothing to do with one vendor. Any capable model can sit inside this loop. What matters is who owns the workflow context and the corrections. We would build the same loop on whatever tool fits your data, your risk, and your budget, and we would change the tool the day a better one appears.

Sources

OpenAI, “Building self-improving tax agents with Codex,” 2026. Original analysis by Dr. Alejandro Canonero, 10 June 2026.