It’s counterintuitive, but saying no more often is the single thing that has improved our work the most in the last year.
We pass on roughly half the projects that reach a discovery call. Here are the three patterns behind most of those noes.
1. “We want to be more AI-driven” with no specific problem
The most common shape. A founder, often well-read, often genuinely excited about what’s now possible, books a call to explore how AI might transform their business.
We used to take these. We’d run a half-day workshop, map their processes, identify candidate automations, write up a report.
Three months later, none of it had been built. The report got circulated, admired, filed. The business kept operating exactly the same way.
The problem isn’t the client — they’re right that AI can help. The problem is that “be more AI-driven” isn’t a problem. It’s a vibe. And vibes don’t ship.
Now we ask on the first call: what’s the first automation you’d push live if we sat down today? If the answer is “I don’t know, you tell me”, we send them the articles, offer a paid scoping engagement instead of a free discovery, and let the market decide. Most don’t come back. That’s fine.
2. “Replace the humans entirely”
Every few months someone asks us to build an AI that handles customer support with no human in the loop. Or drafts and sends contracts without review. Or approves invoices automatically up to a threshold.
These builds are technically possible. They’re also the kind of thing that blows up publicly within the first 90 days. Customers screenshot the wrong answer. A contract gets sent with the wrong party name. An invoice gets approved to a phishing vendor.
We don’t mind ambitious automation. We mind automation that removes the only person who’d catch the mistake. The pattern we’ll build is “draft it, don’t send it”. The pattern we won’t build is “send it without human check”.
The clients who insist on the latter usually go elsewhere. The ones who accept the guardrails usually ship something that still works a year later.
3. Projects where the data isn’t there yet
A frequent third: a client wants an AI that classifies, routes, or recommends based on their data — but the data is in ten different systems, inconsistent formats, and no one has a clean source of truth.
The project the client is asking for is the AI build. The project they actually need is data consolidation. Those aren’t the same project, and we can’t pretend they are.
We’ll happily do the data consolidation work. We just can’t do the AI build on top of it in the same engagement without the timeline doubling and the budget doubling and everyone feeling misled.
What we do take
The projects we say yes to look like this:
- A specific, recurring process the business wants automated
- Data that already exists in a usable form (even if it’s in spreadsheets — spreadsheets are fine)
- A decision-maker who understands that “AI” means “probabilistic” and is comfortable with the 3-5% edge-case review rate
- A willingness to ship a small first version and iterate, not commission a complete system up front
If that sounds like you, let’s talk.