What should AI be doing in your business?
The work nobody enjoys and nobody should be paying a person to do. Tap what eats your week and we'll show you exactly what we'd take off it.
Takes about forty seconds. No pitch deck, no jargon, and we'll tell you where automation is a bad idea too.
What it actually looks like
Not a chatbot on your homepage.
Work that was already done by 6am.
Four we have built. Pick the one closest to yours.
- CollectSoftware
The filing archive is polled around the clock, not once a morning.
- ReadModel
Each new filing read for the meeting type and the date it announces.
- CheckSoftware
Validated against the rules. A misread filing never reaches an analyst.
- FileSoftware
Dated, categorised and queued in the dashboard.
- ReviewYou
An analyst logs in, checks the list and marks items off.
Five stages. One of them is the model. Tap another, or tap the panel to replay.
How it is built
Most of it is not AI at all.
AI is a component. It is not the architecture.
Hand a model an entire process and it will get part of it wrong without telling you which part. So the steps that have to be right every time are ordinary software: testable, repeatable, and logged. The model is used only where judgement is genuinely needed, and checked before anything acts on it.
- SoftwareSame result every run. Testable, logged, reversible.
- The modelOne narrow question, only where judgement is needed.
- YouThe last word, on the exceptions only.
Now yours
Which of these shapes is yours?
Shapes of work, not industries. Tap anything familiar. If yours is not here, the last box is for you.
What we would build
Nothing picked yet. Scroll back up and tap whatever looks familiar. This fills in as you go.
One minute left
Want it costed?
Why it holds up
Built to be checked.
The boring tool, when the boring tool is right
Forecasting on one platform runs on statistical models, not a language model, because forecasting is maths. Sentiment scoring there is a lexicon, not a chatbot. Using a model where a formula belongs is how you get answers nobody can reproduce.
Every AI step is on the record
On one client platform we keep a written register of every AI-touched feature: what it does, which model runs it, its risk classification, and what the user is told. It also records which parts are deliberately not AI.
Wrong answers get caught before you see them
Extraction sits behind a validation layer that catches misclassification before it reaches a person. Answers are grounded in your own documents and cite them, so a claim can always be traced back to its source.
The honest rule
If it has a method and a source, it can be built.
- 1
A method somebody can describe
If you can talk us through it (the real steps, including the awkward exceptions), it can be encoded. It need not be written down anywhere yet.
- 2
Something it can read
A system, an inbox, a folder, a portal, a bank feed, a recorded call. If it does not exist yet, building the thing that captures it is part of the job.
- 3
A person where being wrong is expensive
Hard cases get a review step shaped around how you already work. You keep the judgement and lose the typing.
Just as important
And what AI should not be doing.
Cheaper to hear it now than after the build.
Judgement calls that carry real risk
Pricing exceptions, hiring, anything legal or clinical. AI can prepare the decision; a person makes it.
A process that is already broken
Automating a bad process just produces bad outcomes faster. Fix it first, then automate what is left.
Customer-facing work without a safety net
If being wrong is expensive, the check goes inside the workflow, not around it.
Replacing your team
Give people back the hours they lose to admin, not remove them from the work that matters.