Spent the last year digging into AI stacks inside founder businesses doing serious revenue - £30k to £300k+ a month. Same story every time.

Half-finished agents. Orphaned n8n flows. A Notion board full of 'in progress' that's really purgatory.

Found one founder's workspace with 14 half-built flows. Nine had no data destination. He was paying three contractors to keep building more. The ideas looked great on Monday morning, but those flows never shipped.

The audit always comes down to three things:

No owner.
No success metric.
No home inside an existing workflow.

Here's how it plays out: ChatGPT open. Claude open. n8n open. Zapier open. YouTube open. The course they bought in October open. The Loom from their ops person open. The agent they started six weeks ago open.

Twenty tabs. Zero systems in production.

Guilt kicks in around tab fourteen. 'I'm behind on AI.' So they buy another course. Hire another freelancer. Spend the weekend building something new. Same outcome, different disguise.

Problem isn't motivation - these founders have plenty of that. It's order of operations. Every tool feels equally urgent, so nothing ships.

An AI operating system isn't a stack of tools. It's architecture. And before building anything, one piece of paper needs to answer three questions.

Question one: What process?

Not 'what could AI do.' Name the specific process eating your time, your team's time, or your revenue right now. Lead routing. Sales call recap. Client weekly report. 'Automating marketing' is a category, not an answer. Categories don't ship.

Question two: What data?

Every agent eats data and produces data. If you can't name both on day one, the agent dies on day thirty.

Input: where does it live right now? Slack, Gmail, HubSpot, a Google Doc someone updates every Friday? Who owns it? Is the format consistent?

Output: where does it go? Back into the CRM, into a Slack channel someone actually reads, into a Loom summary? Agents don't die from bad prompts - they die from orphaned data. Input nobody maintains, output nobody reads.

Question three: What win condition?

One sentence. Measurable. Time-boxed.

'Follow-ups sent within 2 hours of every sales call, 95% of the time, measured weekly.'
'Top 5 deals summarised in my inbox every Monday by 7am.'

If the win condition is 'save time' or 'be more efficient', the project is already dead and you're paying for the funeral.

One process. One data path. One win condition.

Here's how I'd run it today: take those three questions and run them against every AI project in the business, live or half-dead. Kill anything that can't answer all three in one sentence each.

For the survivors, pick one - the one with highest revenue leverage, not the most interesting to build. Ship that one in 14 days. Everything else stays closed until it ships.

Would love to hear where you land after running this - especially which project you've been avoiding because the data work is ugly. That one's almost always the highest ROI.