Grow2.ai turned down four pilots in 2025 and signed one we shouldn't have. This note is the operational lesson from those five conversations — the four-question gate that now precedes every contract, and why the cost of walking away is much smaller than the cost of shipping into a pilot you can't hit.
Conversation 1 · The B2B SaaS that wanted "everything"
A 30-person SaaS in Munich asked us to build "a single agent that handles inbound sales, support and renewals across email, Intercom, Slack, and SMS, in five languages." The contract KPI proposed by the founder was "measurably better than our current team's NPS".
That sentence has three failure modes. First, "single agent" across three workflows is an architectural mistake — sales and renewals have different reasoning loads. Second, NPS is a lagging indicator that takes ~12 weeks to stabilise; we cannot contract against it in 14 days. Third, "measurably better than our current team" means we are competing with a moving baseline. We declined.
Conversation 2 · The dental network that wouldn't commit a workflow
A 14-clinic network asked Grow2.ai for "an AI agent that books patients better". Three calls in, the chief admin still hadn't named the workflow. Was it inbound from new patients? Recovery from no-shows? Treatment-plan follow-up? "All of them, eventually."
Pilots that don't name one workflow before Day 0 do not ship a workable agent in 14 days. We offered to run an unpaid 90-minute scoping call to pick one — they declined that too. We declined the pilot.
Conversation 3 · The one we should have walked away from
A real-estate brokerage in Q1 2025 had a clear workflow (lead-to-viewing for residential sales), a clear KPI (viewings booked / week), and a CRM we knew (Bitrix24). We signed.
What we missed in the brief: the brokerage had two competing data sources for "is this listing under offer" — the portal feed and a manually-maintained spreadsheet — and they disagreed about 8% of the time. We did not catch this until Day 9, when the agent had already suggested two under-offer apartments to live leads. We patched it, but the agent went live with a manual override flag and never hit its full KPI. We finished the pilot honestly, the brokerage was happy with the partial result, but it was the worst engagement of 2025 by team morale.
The four-question gate
Every brief now answers four questions before Grow2.ai signs:
- Is there exactly one workflow? Not one initially with three to follow — one. We are happy to discuss the future, but the contract is for one.
- Is the KPI measurable in 14 days? First-reply latency: yes. NPS: no. Quarterly revenue: no. We will help re-shape an unmeasurable KPI into one we can contract against.
- Is the system of record consistent? If the answer to "is this customer/listing/slot still active" depends on which of two systems you ask, the pilot will fail in week two. We need to know upfront.
- Will the contract counter-party be in the workshop? The person who signs the cheque has to be in the Day 5–6 tone workshop. Otherwise the agent goes live in someone else's voice, and the cheque-signer hates it.
This gate has cost Grow2.ai about €60k in declined revenue in 2025. It has also kept the pilot success rate above 92% — and more importantly, it has kept the team able to look our existing clients in the eye.