The work, shown rather than described.

Before

An accounting firm where I hold the fractional CIO and CISO seat, filing 1,000 returns a year. The work started with the firm’s main workflow, mapped stage by stage to find the bottlenecks and the constraints. Three stood out: constant follow-up with clients to get the right information, staff updating the project tracker by hand as documents came in, were worked on, went on the return and waited on the client’s e-file form, and documents moved between systems by hand.

What we built

Custom agents in Microsoft Foundry, built in two months. Document requests are built from last year’s return, with a request for new accounts and tax changes. Agents watch documents arrive and update the tracker, move everything to the tax return once it is complete, and check the finished return against last year and this year’s documents before it goes to the reviewer. After approval, an agent puts the return in the client portal with instructions on how to review it. At every step, the tracker updates and the client is notified.

What changed

1. Clients know what to send, and where things stand. Requests come from last year’s return, and the client hears at every step.
2. Nobody updates the tracker by hand. Agents update it at every step.
3. Nobody moves documents by hand. Agents move them to the return, and the finished return to the portal.
4. Three hours saved on every return. 1,000 returns a year: 3,000 hours.

The tools came last. It started with the map.

Book a 20-minute call to map your workflow

Before

The forecast was assembled by hand. Each region exported booked orders from one system and pipeline from another, joined them in a workbook, and ran the formulas that split the result by country and by account manager. It ran for years. It was accurate enough.

What we built

An agent watching the folder where the exports land. It ingests each export as it arrives, applies weighting learned from each account manager’s own history, and publishes the forecast where the team already works. Nothing cleaned, nothing integrated, and no live connection to the source systems, which is the constraint people assume kills the idea. It doesn’t.

What changed, in order of how much it mattered

1. It stopped being a file. Nobody asks who has the latest version, and you can ask the forecast why a number moved.
2. Forecast accuracy moved a long way. One weighting ladder, set years ago and owned by nobody, gave way to each account manager’s own history.
3. The time saved mattered least. The expensive part had been handover: weeks of picking through formulas every time someone joined or left.

That spreadsheet is an AI accelerator.

The inputs, the rules and the guess are already written down, and an agent starts from there.

Book a 20-minute call to discuss your spreadsheet