Synchronise AI

Long story short

An AI that acts on your business has to ask permission. Nobody has designed that yet.

Synchronise watches a small business's connected tools and stages fully-drafted actions, emails, replies, ad changes, for the owner to approve. No agent holds a send button. Approval is the product.

I built it, shipped it to real owners, and watched what they did. Most proposals were never touched at all. And when one was approved, 61% of the time it had been rewritten first. Approval is not a yes-or-no. It is authorship.

01 — The research

I sold it, used it, and asked about it, because no single signal tells the truth.

Before any of those numbers existed the question was narrower: is this a real pain for a specific person, and can that person actually adopt it? I took it to the exact profile I believed it was for rather than to whoever would take a call, and ran more than 100 conversations across moderated user testing, sales calls and customer interviews.

A segment had to pass both tests. Plenty of owners agreed the idea sounded useful. Far fewer were already doing the work badly at 10pm with a workaround they were embarrassed by. And a segment can have the pain and still be unable to adopt, with nothing connected and no moment in the day the product fits. Those were the expensive ones to learn, because the enthusiasm in the call was real and it went nowhere.

I found the prospects and drafted the outreach using the product itself, so every sales call was also a user test. Interviews tell you whether someone likes an idea; they are much worse at telling you whether the value repeats. Working in it against a real deadline was the only thing that surfaced the second and third week failures, the mornings where the queue stopped being worth opening.

02 — The wall

The wall was not traffic. It was the first decision.

I instrumented the funnel end to end before trusting any of it. Acquisition was fine: 17% of visitors reached sign-up. The collapse happened after the product was in their hands.

Of 18 people who created an account, 3 ever approved or dismissed a single card. I had been designing a queue. The unsolved problem was the first act of trust.

03 — The pattern

Approval is not a choice. It is a rewrite.

Every agent product ships the same control: accept or reject. When I looked at what people actually did with an approval, 61% of approved drafts had been edited first.

So the binary was wrong. I made editing a first-class state, auto-saving on blur and distinct from execution, so someone can take authorship of a draft without firing it. A human edit also freezes the agent from re-proposing over the top of it.

04 — The control

Say what it will do. Then make them mean it.

Two things must be true before someone lets software act for them: they must know exactly what happens, and they must not be able to do it by accident.

The consequence line is derived from the same payload the executor runs, so the preview cannot drift from the effect. Anything with a wide blast radius arms before it fires, and the button never reports success before the server confirms.

The control opposite is the real interaction. Press it twice.

05 — The system

Consistency enforced by the compiler, not by a library.

Thirty-seven kinds of proposal, five surfaces, one designer. A component library would not have held, because surfaces drift when the rule lives in a document nobody reads.

So the visual language is a single projection module. Every card collapses into eight verbs, each with a fixed palette, icon and present-participle loading state, and every surface renders from that one contract. Risk is typed so that adding a new kind of action fails the build until someone declares its blast radius.

06 — Removed

Three things I shipped, measured, and deleted.

The queue's real failure mode is volume, not accuracy. Twenty-seven of the producers I built earned zero approvals across 158 proposals. Supply was never the constraint.

Each of these is in version control with its reason attached.