AI agents do work. Humans exercise judgment. The dial records both choices for each workflow.
The test before live work
I put an AI workflow in front of a prospect this month. We ran four test runs before letting it touch one live, high-stakes document.
It passed every run. The hard decision came before the tool went live: which calls stay human?
The privacy-by-design guide and data inventory guide give you two places to start when the workflow handles company data.
Two settings for every workflow
Every workflow gets two settings:
- A dial from 0 to 100 for how much AI does.
- One marked point where human judgment enters.
Zero means a human does everything. Hundred means the machine runs alone.
Five working settings
- Meeting notes at 90. AI records and summarizes. A person confirms the decisions and owners. Judgment point: what was agreed.
- Invoice entry at 80. AI reads and books it. A person releases the payment. Judgment point: money leaving.
- Customer replies at 60. AI drafts everything. A person sends anything touching money or a complaint. Judgment point: what the company commits to.
- Proposals at 40. AI assembles the evidence. The founder writes the frame, the ask, and the price. Judgment point: the number.
- Hiring at 20. AI sorts and screens. Humans interview and choose. Judgment point: who joins.
The number changes with the decision the workflow can make.
Where the human stays
Research and first drafts now run near ninety, with a person sampling quality. The release call sits at zero: one named person, with no machine in the chain.
The risk lives in the decision, so the human lives there too.
For the surrounding DPDP work, keep the Data Fiduciary guide and Data Processing Agreement guide close to the workflow record.
The choice per workflow
Teams that debate AI models keep the argument at the model level. Teams that pick the number per workflow and name the human decision are doing engineering.
The dial ends the fight. Pick yours and let the agents execute.
Set your workflow dial
Write one workflow the team wants AI to run.
Set its AI level from 0 to 100.
Mark the point where a human decides.
Name the person who owns that decision.
Run four tests before live use on a high-stakes document.
Next step
Bring one workflow. Leave with its AI level and named human judgment point.
Book a 30-minute call →Related reading
Written by Sushant Pasumarty
DPDP consultant. Founder, Meridian Bridge Strategy. This page began as field notes from live client work.