Most people use AI to write. I use it to work.

I run sales at a payment orchestration platform. A while ago I built an AI system that works my own pipeline: it drafts, it chases, it documents. It never sends anything to a client and it never decides anything. My signature is the last gate, every time.

This is what it actually fixed, what it didn’t, and the thing it changed my mind about — which of the assets in a business is actually worth anything.

The problem

A big, disorganised pipeline and not enough hours. That’s it. That’s the whole problem, and it’s more specific than it sounds.

A pipeline stops being manageable at the point where it exceeds one person’s attention. Not one person’s skill — attention. I knew what to do with every deal in there. I just couldn’t hold all of them in my head at once, and the ones I wasn’t holding were the ones going quiet.

So the work split into two piles. The first pile was selling: calls, pricing, objections, the actual judgement. The second pile was everything the first pile generates and nobody sees — reconstructing where a deal had got to before a call, remembering what I’d promised on Thursday, checking whether the record matched reality, working out which of the quiet ones had gone quiet for a reason and which had just been forgotten.

The second pile isn’t selling. It’s the tax you pay for selling. And it all lived in one place: my memory. That’s a single point of failure with a holiday calendar.

What was built

A system that takes the second pile.

It works the pipeline on a schedule rather than when I remember to look. It finds deals that have gone quiet and prepares the follow-up. After a call it takes the transcript and drafts what I said I’d send, while the conversation is still current. It reconstructs a deal’s full history into a brief I can read before a call instead of assembling it myself. It keeps the record straight: what was discussed, what was agreed, what’s outstanding.

What it deliberately does not do is more important.

It does not send. Every message it writes lands in drafts and waits for me. Not because the writing is bad — because the cost of one wrong message to a client is higher than the time saved on all the right ones.

It does not decide. Price, concession, walking away, who to push and who to leave alone — that stays with me. The system prepares; it doesn’t choose.

It does not assert facts it can’t source. A deal cannot be moved to “demo done” because the system infers a demo probably happened. Either there’s a transcript, a message, a calendar record — or the item is marked unverified and I look at it.

Why it’s built that way

Because the constraint was never writing. It was attention.

If you think the problem is “producing text faster,” you end up with a chatbot in a browser tab: you paste context in, you get something good out, you paste it back, you edit, you send. The output improves and your day doesn’t, because you’re still standing in every step. I did that for a while. Nothing about the economics of my week changed.

The shift was treating the model as a component inside the process rather than as something I consult. A component runs on a schedule, reads the systems where the work actually lives, and hands back something finished. That’s a different category of thing from a conversation.

And once you do that, the discipline becomes the whole design. A component that runs without me watching can quietly put wrong things into the record, and wrong things in a CRM are worse than gaps, because gaps are visible. Hence the two rules that shape everything: a human on the send button, and no fact without a source. Those aren’t safety theatre bolted on afterwards. They’re what makes the rest of it usable.

The result

I’m not going to give you a performance percentage.

I could. I have counts, and a set of them looks impressive. But a productivity statistic on a page like this is worth exactly as much as your willingness to trust the person who picked it, and picking flattering ones is the easiest thing in the world. The numbers I will give you further down are about the size of the job, not the size of my results — those you can sanity-check against your own operation without taking my word for anything.

What changed is that the second pile stopped depending on me remembering it.

Things surface now. That’s the difference I’d point to. Before, a problem became visible when it had already cost something — a deal had gone cold, a promise had aged past the point where sending it was still useful. Now it comes up while it’s still just a task. The system raises it; I deal with it. I’m not being modest about that division: it doesn’t find things because it’s clever, it finds them because it looks at everything, every day, which is precisely what I couldn’t do.

Follow-ups from calls go out because they’re drafted while I still remember the call, not two days later when I’m reconstructing it. Deal context exists outside my head, in a form someone else can pick up — which stopped being theoretical the first time a colleague had to take over deals they’d never worked on.

And I get my attention back for the first pile.

What surprised me

That the writing was the least of it.

Almost everyone I talk to is using these tools to produce text. That’s the obvious application and it’s the least valuable one. The real use is work: give it tasks. Fan out, delegate, come back to finished output. A large amount of work gets done in parallel, facts get found and cross-checked against each other, and the fog over a situation lifts.

I have an unusually good memory. I’ve relied on it my whole career. This beats it — not by being smarter, but by being complete and tireless where I’m selective and finite.

Here is the piece of work that made the point.

I had close to two years of recorded calls sitting in an archive — my own client conversations and internal meetings, in several languages, going back to autumn 2024. Roughly five hundred hours of audio. Alongside it, the written communication over the same period.

Almost none of it had ever been looked at again. That’s normal. Nobody re-listens to a call from fourteen months ago, because the cost of finding the useful thirty seconds is higher than the value of finding it. So a company accumulates a large, expensive record of exactly what its market said it wanted, and then doesn’t read it.

I had the whole corpus pulled in and worked through as one body of text: who these buyers actually are, what they keep asking for, which pains come up repeatedly, who we get compared against, and where the thing they need diverges from the thing we describe.

Do that by hand and the listening alone is five hundred hours. Twelve working weeks, at a desk, with headphones on, before you have written down a single observation. Nobody is ever going to authorise that, which is why the archive just sits there. The reading-and-noting on top of it is more time again.

That’s the compression I didn’t expect. Not “faster at drafting.” A category of work that was previously impossible on any realistic budget became a task you set up and come back to. That was the moment the tool stopped looking like a tool.

And it left me with a conclusion I wasn’t looking for.

Everyone I compete with can access the same models I can. That access is a commodity, it gets cheaper every few months, and it confers no advantage whatsoever. What none of them have is two years of my customers explaining, in their own words, what they actually needed. Neither did I, in any usable sense — the record existed, but it couldn’t be read.

So the data was never the missing piece. It was always there, accumulating, unread. What arrived was the ability to read it at a price that makes reading it worth doing.

That reverses how I would have ranked my own assets two years ago. The tooling isn’t the advantage; my competitors can buy the identical tooling this afternoon. The archive is the advantage, and it’s the only part that can’t be purchased, copied or subscribed to — it accumulates only if you were in the room for those conversations.

The practical version, for anyone running a company: your call recordings, your support tickets, your email, the notes your salespeople leave in the CRM. You have probably been treating that pile as waste — a by-product of doing the work. It isn’t waste. It’s the only proprietary input you own, and as of now it’s readable. That changes what it’s worth to keep it properly.

What I’d do differently

I’d put it on a schedule from day one.

For the first stretch, everything was built and almost nothing was running. The recurring jobs sat switched off — partly caution about what the runs would cost, partly a reluctance to let something operate while I wasn’t watching. So the system only worked when I dragged it into action by hand, which meant it worked in bursts: a lot of activity on the days I remembered, nothing on the days I didn’t.

That’s the whole failure, and it’s worth being blunt about it. A system that depends on you remembering to run it has reproduced the original problem inside the solution. The entire point was to stop relying on my memory. I built the thing, then gated it behind exactly the thing it was meant to replace.

The lesson generalises past me: if it isn’t on a schedule, you don’t have automation. You have a tool you’re fond of.

If you want to look at whether the same discipline would hold up on your own systems and your own data, the advisory page has the shape of that conversation and what it costs.