Research preview

Turn your team’s habits into your agents’ second nature.

Your agents shouldn’t reflect on everything. gives them learned reflexes for the work you repeat, and keeps reflection for anything new.

A pianist playing alone on a dark stage under a single spotlight

Route a refund requestReflex

RequestRefund over £50
Your team usuallySends it to Finance
Answered byA learned reflex
Decided in about a second
The problem

Agents keep asking the same question

1Agents ask a large, general model the same kind of question thousands of times.
2Every answer is slow and paid for every time, and the model does not know how your team decides.
3Your people correct it, and nothing remembers.

Not a mask on someone else’s model.

A person holding a plain white mask over their face
BorrowedMost agents put a new face on someone else’s model. Every decision goes back to it, and nothing is learned.
A potter finishing a clay bowl on a wheel
LearnedFor the work you repeat, gives your agents reflexes of their own, learned from your team’s approvals.
Reflex and reflection

Two ways of deciding

ReflectionYour AI model

  • Slow and careful
  • For anything new
  • For anything unclear

ReflexLearned from your team

  • Learned from your team’s approvals
  • For the decisions you repeat
  • Answers in about a second, and only when sure

Reflexes have to earn their place: they practise in shadow first and step back when corrected.

How it works

Your decisions become its lessons

Habits form the same way in the brain: repetition, plus a signal of what was right. Here the signal is your team’s approvals.

Your work goes to a learned reflex if it is sure, which answers in about a second. If not, your AI model answers, your team approves or edits, and the reflex learns from that decision before the next piece of work. Yes No Back to the start Your work Is a learned reflex sure? Answers inabout a second Your AI model answers Your team approves or edits The reflex learns
What you get

Learned reflexes, made your way

Each one is a small model of your own, trained on how your team decides.

1. Accuracy

It decides like your best people, every time.

2. Privacy

It learns only with your consent, from your work, with secrets removed first.

3. Cost

Decisions it has learned cost no credits on Pro and above, and need no large model.

Early results

One replay of real work

On a replay of one business’s five months of work (32,581 requests):

99.7%right on repeated decisions (vs 96.4%)
37%answered without the large model
About 1 sper answer

One replay, not a promise. The pilot measures it on real customers’ work.

What it does and doesn’t

Small, repeated choices only

It does

  • Sorting
  • Routing
  • Labelling
  • Picking a tool
  • Filling in a small form

Short choices your team makes again and again.

It doesn’t

  • Write
  • Plan
  • Reason through a new problem

Those stay with your AI model.

You stay in control

Nothing changes until you say so

1Off until you switch it onIt stays off until an admin switches it on and agrees to training.
2It has to prove itselfNothing goes live until it has agreed with your team on your own work.
3Hold or roll backAn admin can hold any learned reflex, or roll it back to an earlier one.
4Clear what it learnedYou can clear what it has learned at any time.
Coming next

What we are working on after the pilot

Planned, not available today.

1Starter reflexes in use-case packagesNew teams start with learned reflexes for the common decisions in their use case, instead of from zero.
2On your own infrastructureLearned reflexes that are trained and answer inside self-hosted and air-gapped installations.
3A monthly reportHow accurate your learned reflexes were each month, and what they saved.
FAQ

Questions about

Does it train anyone else’s model?

No. Your small models learn only from your organisation’s work, with your admin’s consent.

What if it gets something wrong?

It answers only when it is sure; otherwise your AI model answers. Corrections from your team count against it, and too many send it back to practising.

Which plans can join?

Pro, Agency and Enterprise, during the research preview.

Can I turn it off?

Yes, at any time. Your AI model takes over every decision again.

When is it generally available?

After the pilot.

Join the pilot

We’re working with a small number of teams on Pro and above.

Join the pilotSee how it works