How I did in two weeks what the client had been paying for over four months
A story about a $380k AI platform and a $6k assistant. About how the blocker is usually not the technology, but the framing. And about the one process I picked.
Nikita called me on a Thursday evening. Nikita runs a company — 80 people, B2B SaaS, revenue growing for two years, and everything, on the face of it, is fine. Except that four months earlier Nikita had paid a startup three hundred and eighty thousand dollars for an "AI-powered automation platform" — and right now, Thursday evening, the platform still hadn't launched.
— Listen, — Nikita says, and I can hear in his voice he's not angry, he's tired. — We genuinely don't understand what stage we're in. They show a demo, something works in it. Then they say: "two more weeks." And we've had four of those "two more weeks" already.
— What are you actually automating? — I ask.
Pause.
— Well… all of it. Customer service, mailings, analytics. "A single AI contour for the enterprise."
That's when I understood the problem.
Nikita's problem isn't the technology. Technology in 2026 is roughly the same for everyone — the startup charging three percent of revenue and me on my laptop. Nikita's problem is that he paid for "a single AI contour for the enterprise" rather than for one specific process that stops eating the time of one specific person.
Those are two different tasks. And the first one literally can't have a correct spec, because "automate everything" isn't a spec, it's a wish. Wishes tend to be expensive.
(I later asked the startup whether anyone on their team had actually watched one of these processes end-to-end, with their own eyes. Turned out no one had. They were working off "business requirements" written by a product manager on Nikita's side, who had never done the processes himself, because he'd inherited them three years earlier from a previous product manager. And so on.)
So I offered Nikita a deal. Stop the platform. Temporarily. Two weeks. Let me pick one process, one user, and one measurable result. In two weeks we look at what came out. If nothing came out — I return the money, we resume the platform, conversation over.
Nikita agreed. Mostly, I think, because of the "I return the money" part, which you don't hear often in consulting.
The first two days I just watched how people work. That is, generally, the most valuable thing I do as an outside person — I watch how people work. Insiders can't really do this: they have their own picture, they can't disagree with it, because they live in it.
(In parallel, to save myself a day, I ran Nikita's site through inite.ai/en/analyze. It's our own tool — it shows in two minutes what AI currently "says" about a company, which of its processes and products the models know, and which it doesn't. Free, quick entry point: by the end of the first coffee I already know where they're empty and where they're overstuffed. I recommend it to anyone starting anything similar — saves at least half a day of "tell me about yourself.")
Nikita had three time-eating processes. First — a support operator sorting and reclassifying incoming tickets by hand. Second — a salesperson writing almost identical qualification emails to new leads. Third — a finance analyst pulling together a weekly report from four sources into a readable shape.
I picked the second. Not because it was the "most important" — the most important for the business was the first. I picked the second because it was the simplest, the most repeatable, and had the clearest measurable outcome. And because the salesperson was a star — her name was Lena — who hated writing those emails so much that she said so openly. My rule is to start where there's pain plus enthusiasm. Lena was perfect.
We wrote the spec in forty minutes. "After we ship this, Lena spends no more than 3 minutes per lead on qualification emails (down from 15), with template-fit accuracy no lower than 85%." That's it. No "platform," no "ecosystem," no "centralised contour."
Then three days in Claude. I loaded twenty real emails Lena had written (thanks, Lena), drafted the prompt, tested, reformulated, tested again. By day three Claude was producing emails Lena accepted without edits in four cases out of five. One more with one or two tweaks. One more — no. 80%. For now, enough.
Why I didn't jump straight to writing code — a separate conversation. Short version: if the model can't do it through the interface, it won't do it through the API. That's a simple truth I didn't invent; smart people told it to me several times before I started honouring it.
Days nine through eleven were integration. I'll say up front — this is the least creative part of the project. Nikita's CRM was HubSpot. HubSpot's API does everything needed. I wrote a thin wrapper in Hono, connected it to claude.ai through the Anthropic API, and the flow became this: Lena opens a lead card, presses one button, seven seconds later she has an email. Edits if needed, sends.
On day twelve I sat next to Lena and we walked through ten real leads. Lena stumbled twice: once on a lead from a sector the model had seen rarely (I added three examples to the prompt, fixed); a second time on a phrase Lena found "too salesy" (tightened the tone-of-voice in the prompt, fixed). This, by the way, was my favourite part of the work — sitting next to someone who actually uses what you've built. Without it the rest is meaningless.
On day thirteen Lena was flying solo. On day fourteen Nikita called to say that at the standup Lena had said something positive about her work for the first time in three months. That's a separate metric I hadn't put in the spec, but — it turned out — it counted too.
I took the second process — support tickets — three weeks later. The third — the finance report — two months later. Half a year in, Nikita has seven assistants running. Each one was built in two to three weeks on the same principle: one process, one user, one measurable metric.
Nikita later told this story on a podcast, and they asked him: "So what happened with the three-hundred-and-eighty-thousand-dollar platform?" He answered honestly: "We closed the platform, got part of the money back, wrote the rest off as education."
(Actually we counted the "rest" as a first lesson. He wouldn't want one or two more like it, and I don't think there will be.)
I'm not writing this to trash the startup or the platform. Maybe it works for their other clients. I don't know.
I'm writing this because four out of five of my clients come to me after they've already paid someone for an "AI platform," or "digital transformation," or "unified contour," and three-to-six months in, nothing has happened. Then they come to me. And I build them one thing, and it works. Not because I'm smarter than a three-hundred-and-eighty-thousand-dollar startup. Because I take one process.
If the first sentence of a project description contains the words "platform," "ecosystem," or "centralised" — it's not two weeks. It's a yearly budget with no guarantees. You might have that budget. Most don't.
Start with one process. You've got, I guarantee it, about twenty of them eating more time than they should.