Our view on AI
AI should take on the work. People should keep the decisions.
Every Orbixa business is shaped by one view of how AI should be used, which we call AI loops. It comes from what building and running these businesses has taught us.
How most businesses use AI
Linear. Every step ends at a person.
AI makes producing things nearly free. Checking them still costs human time. So every AI step adds to a pile waiting for someone, and growth is capped by how many hours your best people can spend reviewing.
Scales with headcount.
How it should work
Looped. People at the edges, deciding.
AI goes round: watching, spotting what is wrong, working out what it is allowed to do, doing it, then checking it actually worked. People set direction, hold authority and make the calls with real consequences.
Scales with judgement.
Principle one
Intelligence is not authority.
Most people assume that as AI gets smarter, you give it more freedom. That is backwards, and it is dangerous.
How clever a system is and how much it is allowed to do are two separate dials. A model can keep getting better at spotting a fraud pattern, a churn risk or a pathway that has gone quiet, just by being a newer model. That should never automatically buy it more say over what happens next.
AI can keep getting better at understanding. Only a person can give it more permission.
Four levels, no exceptions
Every action an AI takes sits on one of four levels.
Reading, watching, analysing. Nothing has changed in the outside world yet.
Flag suppliers past their payment terms
Low risk, reversible, and inside rules you have already set.
Chase an overdue invoice with an approved template
Money, customers, brand, or anything hard to undo.
Discount a customer's bill
Judgement that stays with a person, whatever the system thinks it has found.
Letting someone go. A clinical decision.
The AI can suggest moving an action between levels. It can never move itself.
Principle two
Judgement, made once, keeps paying out.
Every yes at level 3 is data. Approve the same small invoice chase four weeks running and it stops being a decision. It is a pattern, so it becomes a rule and drops to level 2.
The loop takes on more each month, on purpose. Your attention moves up to the decisions that are still genuinely hard.
Businesses will not win by having the most AI tools. They will win by turning good judgement into clear rules faster than their competitors.
Example
- Week 1Chase overdue invoice, £120Approved
- Week 2Chase overdue invoice, £85Approved
- Week 3Chase overdue invoice, £160Approved
- Week 4Chase overdue invoice, £140Approved
Principle three
Loops fail politely.
They rarely crash. They quietly do something easier, report success and wait for the next instruction. So every loop we build has four parts, and three of them are business decisions, not technical ones.
Yours
GoalA done condition a stranger could check in ten seconds.
Without it: hard work on the wrong thing.
Improves without you
WorkThe AI doing the task. Already good enough for most jobs.
Not where results are won.
Yours
CheckTests the outcome, not just that something finished.
Aimed at completion: bad work waved through.
Yours
StopBrakes decided in advance, including the right to say "I'm not sure".
Without it: runs all night and bills you for it.
We hold our own systems to this. An audit of Leva's planning step found it had taken a fallback route 76 times out of 76. Every check passed, because every check measured completion. None asked whether the plan was right for that business.
Nothing crashed and no alert fired. It surfaced only because the work ran as a loop, with named steps and a record of what ran. Now every loop records which path it took, not just whether it finished.
In our businesses
How the view shows up in practice.
Each business applies it differently, depending on its customers and what is at stake.
Every action the AI team takes is set to one of the four authority levels before it runs. Founders approve what matters.
Diagnoses first, then gives teams a clear system to run, so owners stop being the link between every step.
AI only where it earns its place, tested on the client's own data, with named people signing off each output.
The AI prepares the admin. Every clinical decision stays with a clinician. Proven on synthetic data first.
Lessons from building
Our view on AI, from building it.
Practical lessons on AI loops and putting AI to work in a business: what works, what breaks and what we would do differently.
Why most AI adoption turns good people into reviewers, and how to move them back to the decisions that need them.
Learned building LevaOur own audits, including the ones that went badly, and the checks that now catch problems before customers do.
Learned across the groupHow to decide what an AI can do on its own, and why that has to be a human decision made in advance.
Learned building Leva Health
Put AI to work