Reconciled payments
Payments matched across your bank and payment apps, with every mismatch listed for a person to check.
Governed AI workflows · Paid per output
AI workflows that can't overstep. You pay only for what they deliver.
Give me one repetitive job your team does every week. I'll build an AI workflow that does it, with hard limits on anything it's allowed to change. You pay for each finished output you accept, not for hours, seats or experiments.
3 pilot spots Free to pilot Then pay per accepted output
Every action goes through a check
ExampleAn illustration of how decisions are made. Each real workflow gets its own written rules. Try the full playground
What you pay for
You buy finished work, not software.
A pilot starts with one job. These are the kinds of output a governed workflow can hand back to you.
Sorted enquiries
New enquiries read, tagged and ranked, each with the reason it was ranked that way.
Call summaries
Recorded calls turned into notes: what was asked, what was promised and what's due next.
First drafts
Replies, quotes and reports written from your own templates, waiting for someone to approve them.
These are examples, not client results. Your pilot is scoped around your own work.
How it works
From one job to paid output in four steps.
-
Pick one job
Something your team does every week, with a clear idea of what "done" looks like.
-
Agree the limits
We write down what the workflow can read, draft and change, and who approves what.
-
Run it and review
The workflow does the job. You check the output, accept what's right and send back the rest.
-
Pay per accepted output
After the free pilot, each accepted output has an agreed price. Work you send back costs nothing.
How the checks work
Thinking is free. Changing things needs permission.
Autonomy is what an agent may decide. Authority is what it may change.
The rule every workflow is built around
Most AI tools give an agent a tool and, with it, the power to change things. I keep those apart. The workflow can work things out freely, but the right to change anything sits outside it, in rules you can read.
-
Reading and drafting are safe by default.The workflow can look things up and prepare work without changing anything.
-
Changes need permission.Anything that writes, sends, pays or deletes must be allowed in the written rules for that job.
-
Risky changes wait for a person.You decide which actions need a named approver before they happen.
-
Everything is recorded.Every request and decision is logged, so you can see what happened and why.
From the AI Diary
How the work actually gets done
A green build is not permission to merge
Three public incidents from last year explain why the software workflow I built today opens pull requests but refuses to land them, even when every test passes.
A fix isn't finished until a test attacks it
How I track weaknesses in a workflow, and why a fix only counts once a test tries the failure and it no longer works.
Why I don't trust a green checkmark on its own
A passing check can hide more than it shows. What I now require before I call a result real, for code and for AI output.
Who you'd work with
Hi, I'm Dinesh.
I design AI workflows around one idea: they should do useful work, and nothing else. Each one starts as a written set of rules, and I test those rules by trying to break them.
I also run small businesses of my own, so I build for people who need the work done on a Monday morning, not for a demo.
Design-partner pilot
Three teams. One workflow each. Free.
You get the workflow and its output during the pilot at no cost. In return, I ask for honest feedback and permission to write up what we learned, anonymised if you prefer.