What “Agentic” Actually Delivers

July 2026 · 7 min read

Most AI in go-to-market produces a copilot — something that drafts an email a human still has to send. We built the other thing: a system that notices a merchant's deal has gone offline, contacts them, gets their consent, and makes the fix itself.

The quiet failure

In a marketplace, the loud failures get fixed — a checkout bug pages someone within minutes. The expensive failures are quiet.

A merchant sells out their monthly voucher cap on a Tuesday and their deal stops being purchasable. Nothing breaks. The merchant doesn't notice, because nothing changed on their end. The rep doesn't notice, because they own hundreds of accounts. The deal sits offline, and every day is revenue neither side ever sees. The traditional fix — more account managers — scales with headcount, which means it doesn't scale.

How it works

Detection is the easy half and has existed for twenty years. What matters is whether the loop closes without a person in it:

The agentic channels don't just work — they hold their own against the human one, because the merchant never has to leave the message. No login, no form, no call to schedule. The whole transaction is one word.

What it produced

One quarter of recovered merchant value, split by whether a human was involved:

The green sliver is the one I care about, precisely because it's the smallest number on the page. The gray bar grows with headcount. The green one doesn't — and three months ago it was zero. That slope, not the absolute value, is the whole thesis.

The 98% haircut

Any alerting system can produce an impressive number: fire an alert, observe that the deal later made money, claim the money. Most internal AI programs report their wins this way, and most of those numbers are fiction.

The honest question is counterfactual — how much would have happened anyway? So we calibrated every alert type against a control group of alerts that were generated but never actioned. Our biggest-looking result did not survive it:

Not a fun chart to present. Also the only reason the rest survived scrutiny — at the investment review nobody litigated whether the value was real, because we'd already argued against ourselves harder than any reviewer would. Every dollar above reconciles to a specific deal, on a specific date, with a specific touchpoint. Without that you don't have a result. You have a slide.

Two people

All of it is built and run by two: one engineer and me. Not as a flex — it's the product claim. If it took thirty people to build a system whose purpose is removing manual work, the argument would refute itself.

It works because one person holds the entire chain: signal definition, data modelling, message copy, the commercial case, the attribution. Most of the cost in enterprise projects isn't the work — it's the handoffs between six people who each own one slice of it. AI absorbs the analysis: the pipelines, the backfills, the warehouse queries, the reporting. What stays human is choosing which alert types to build, and deciding when to tell a stakeholder the number is worse than they hoped.

That capacity is what lets a two-person team run a catalogue this wide — and keep extending it:

The scarce resource was never engineering capacity. It's someone who understands the commercial mechanics well enough to know which fix is correct, and who can ship it without translating through three teams first.