The default failure mode

Most AI projects we get called into started somewhere else. A vendor demoed something impressive, a steering committee was assembled, a six-figure pilot was scoped, twelve months later there’s a slide deck and no production traffic.

The failure isn’t the AI. The failure is the shape of the engagement.

What changed in our practice

We’ve been writing production software since March 2006. Most of those twenty years were spent on systems that nothing exciting was happening to — just code that had to be correct, fast, and on-call-able at 3 a.m. Then in the last three years, we started threading AI into those same systems for our own internal use first, then for clients.

The lesson from that arc is uncomfortable for the AI consulting industry: AI is not the hard part. The hard part is integrating a probabilistic component into a deterministic business workflow without breaking the workflow.

If you can already keep a CRM workflow healthy in production, you can add an AI scoring layer to it. If you can’t, no AI integration is going to fix the broken workflow underneath.

The 30-day rule

So we changed how we sell. Three weeks of building, one week of measuring. One workflow. One AI integration. Production traffic. Either the metric we agreed on moves, or it doesn’t.

This is unappealing to people running long, sticky, billable engagements. It’s the right shape for the buyer.

Week 1 — observation

We sit with the operators. The CS rep, the SDR, the AE, the support agent — whoever does the work. We watch them do their job. We map where the friction is. We propose three places AI could help, ranked.

The client picks one. We write down the metric we’re going to move (response time, lead-to-meeting rate, ticket resolution, manual-touches-per-record, whatever maps to money). We write down the cost ceiling (token spend per month). We commit to a number.

Weeks 2–3 — build

We ship the integration into the system the team already uses. HubSpot, Salesforce, Zendesk, Mailchimp, Klaviyo, Zoho — whatever it is, the AI runs underneath. The team’s muscle memory doesn’t change. The interface they touch every day doesn’t change. Only the work that’s happening underneath changes.

Production observability goes in from day one. Token cost tracking goes in from day one. The rollback path is tested before any real user sees the AI.

Week 4 — measure

We compare the metric against baseline. We write an honest report. We say whether the integration is doing what we said it would. If it isn’t, we don’t propose a second engagement — we walk away with what was spent.

Why this matters

AI projects fail expensively. Most of the failure happens in scoping. The longer the scoping, the bigger the gap between what the steering committee thinks the AI will do and what the system can actually do once it hits real data.

The 30-day rule isn’t a marketing format. It’s the right amount of time to find out whether a specific AI integration is the right move for a specific business. After that, you scale what works.

If you’re thinking about AI for your business

Three questions you should be able to answer before any AI engagement:

• What metric will move? Not “efficiency”. A specific, measurable number that someone in the company already cares about.
• What is the cost ceiling? Token spend can spiral fast. You should know the monthly budget before, not after.
• What does the rollback path look like? If the AI does something embarrassing, how does it stop in the next ten minutes — not the next three weeks?

If your prospective AI partner can’t answer those three questions before they send a proposal, they probably haven’t shipped enough AI in production to be useful.

If you want to start a conversation, the AI Readiness Assessment is the fastest way. A senior engineer reads your stack and replies with a 3-integration plan in 48 hours. No deck.