An AI agent that can't be measured is a chatbot that can't be fired. Before we touch a keyboard, we agree on a number — first reply latency, no-show rate, qualified meetings per week, average review score — and we write it into the contract on Day 0.

If we miss it, you don't pay the ongoing fee. We've missed it three times in thirty-one pilots. We refunded one in 2023 in full because the recovery wasn't on a credible path. The other two we worked off-the-clock until the number landed.

Anything we can't define a measurement for, we say no to. There are good companies who do un-measurable creative work. We are not them.

Voice is not a font choice. The agent's first reply, refusal, escalation and apology are part of your brand the same way the receptionist's voice is. We refuse to ship in a generic AI tone — eager, em-dash-laden, bilingually robotic.

So before launch we run a tone workshop with the human who normally writes to your customers — the front-desk lead, the senior salesperson, the founder. Day 5–6 of every pilot. Iryna runs it. We come away with a voice document that sounds like one specific person on your team.

The cost: a couple of hours from someone you can barely spare. The benefit: nobody on the receiving end can tell, which is the only quality bar that matters.

▸ Atlas Rentals (Case 04) caught our first draft sounding "like a hotel chain". We rebuilt the voice with their ops team. Communication scores moved first; everything else followed.

No fake-human chatbots. Every agent we ship discloses, in the first message, in the user's own language, that it is an agent — not a person. Then it tries to be useful enough that the disclosure stops mattering.

This is non-negotiable for two reasons. First, regulatory: Ukraine, the EU and the UK are converging on disclosure mandates and we don't want to retro-fit them. Second, ethical: the trust dividend a customer gives a company that doesn't pretend is large, durable, and almost free.

If you need an undisclosed bot to make your numbers work, hire someone else. They exist. We won't compete with them.

Every conversation the agent has is logged, indexed and queryable by you. Not by us, not by an opaque vendor — by you, in your stack, owned in your audit trail.

We use Temporal for workflow durability and ship export hooks to your CRM, your data warehouse, your DPO's mailbox. If our model provider goes down, the conversation is replayable. If a regulator asks why a patient was offered a particular slot, we can show the chain of reasoning that produced the message in twelve seconds.

This costs us margin. We pay for the storage. We pay for the export tooling. We've baked it into the price.

Every pilot is staffed with two named engineers from grow2 — not a "team", not "an account manager and an offshore pool". One of them is in your Slack. The other is on call.

This rules us out of large enterprise procurement and we are at peace with that. The decision to keep the headcount low and the names visible is also a decision about who we sell to: businesses who want to know whose hand was on the keyboard when something broke at 02:00.

If you want a vendor with a logo wall and a 24-hour SOC2-typed call center, we are not your shop. We will introduce you to two we trust.

A two-week pilot is long enough to build something real and short enough to keep both of us honest about scope. We've run it thirty-one times now and the cadence is the same every time — Day 0 brief, Day 4 channels, Day 5 voice, Day 7–10 shadow, Day 11 calibration, Day 14 hand-off.

We will say no to clients who want a "discovery phase" before the pilot. We will say no to ones who want a "30-day extended pilot" because they want to push the KPI conversation off. The shape of the engagement is the work. If you can't commit to fourteen days you can't commit to a contract.

▸ Of thirty-one pilots since 2023, twenty-seven hit their KPI by Day 30. Three were recovered by Day 60. One we walked away from.

We cap ourselves at six pilots a quarter. Not because we couldn't sell more, but because two engineers can't ship more than that without one of the articles above breaking.

The waiting list is real and we don't apologise for it. If you call us in week 11 of the quarter, you'll start in the next one. The cap forces us to say no, which forces us to be honest about which problems we can credibly solve. Both are features.

If a year from now we're shipping twenty-four pilots a quarter, it'll be because we found two more engineers we'd both put our names next to. Not because we stretched the existing two thinner.

We are Ukrainian. We chose to keep the studio here, with the lights and the sirens and the coffee, because we believe the strongest case for the country right now is software made here that demonstrably works.

Our clients are in Kyiv, Lviv, Warsaw, Berlin, Vilnius and London. Every contract we sign keeps two engineers and one tone editor employed in a country that needs both. Our pricing reflects Western European rates because we deliver Western European outcomes; we are not a low-cost provider, and we don't want to be one.

▸ This document, like the agents we build, was written by hand. We will keep revising it. Every revision is dated.