Workflow AI Strategy and Design
For the workflows that pay: how they get designed, built or bought, and what they return.
Deciding what to automate matters more than deciding how. This is where the money leaks: tools stacked on processes nobody fixed, and output that is polished, persuasive and wrong. For the few workflows that survived Sprint 1, we design the transformation at a level a vendor can build against and a CFO can underwrite.
The technology is commoditising. The design is not. A workflow automated at the wrong step, or built where it should have been bought, is money spent moving nothing. We design the what and the why with enough precision that implementation becomes execution, not interpretation.
- 01
Every step allocated: run it, assist it, or leave it to a person
Run it automatically, run it with a human holding the decision, or leave it alone. Where the call is genuinely unclear, it gets a small, measured pilot with a stopping rule, not a guess.
- 02
The operating model the redesign implies: roles, accountabilities, review points
Which roles change, what they become accountable for, and where the review points sit. Executing this is Sprint 3. Designing it here is what makes Sprint 3 possible.
- 03
Requirements written technology-agnostic. The design is yours
Any competent vendor can execute it. No lock-in, and no architecture chosen to suit someone’s licence.
- 04
The controls that catch the failure nobody sees
AI output reads more persuasive precisely when it is wrong. We put friction exactly where a reviewer would otherwise wave it through, and review the logic separately from the presentation.
- 05
A costed P&L per use case, assumptions visible
What it costs to stand up, what it returns, and when. In time or money, stated as a range, with the assumptions on the table.
- 06
A 90-day proof, and a date to reassess as the frontier moves
So the first use case reaches production, not pilot purgatory. And because the frontier shifts, a date to look again: what sits outside AI’s reach this year may sit inside it next.
The jagged frontier
AI capability is not a smooth line. Inside one workflow it is brilliant at a task and confidently wrong at the next that looks identical, and people cannot reliably tell which is which. We map it task by task, so automation lands only where it holds.
Build, buy, partner, or wait
Most workflows do not need a bespoke model. Some need a vendor, some a partner, some simply are not ready. We take no referral fees, so the recommendation is the honest one.
The persuasion trap
The most dangerous AI error is the one that arrives beautifully written. Ordinary review, “does this look right”, passes it straight through. So the review layer gets built before the automation layer, not after.
The workflow that should not have been automated
A client wanted to automate customer quotes end to end. On the frontier, the templated portions sat well inside it. The commercial terms and the risk read did not.
The design automated the first and kept a human on the second, with a review point built in. The version that automated everything would have looked flawless and mispriced risk on every third quote.
- Technology-agnostic, so anyone can implement.
- Vendor partners, so it gets done. No referral fees.
- A 90-day proof, so it actually happens.
- Build, buy, partner or wait, and we say which.
- A workflow-by-workflow redesign for your top three to five use cases
- An automate / assist / keep-human map for every step
- Technology-agnostic requirements any vendor can build
- A costed P&L model per use case, and a 90-day plan to prove the first
A blueprint precise enough to build. And to underwrite.