It learns which of your plays actually worked.
Not “the AI runs your campaigns.” Something narrower, provable, and far more useful: every recommendation is logged with the number it was trying to move, then checked against what that number actually did. Over time you stop arguing about which plays work and start ranking them by result.
Four steps, and they have to be in this order.
“It gets smarter” is a claim, so it needs a mechanism — and the mechanism is not “we store a lot of data.” Here is the actual sequence, including where we currently are in it.
1. Write the rules by hand. A small, closed set drawn from what actually works in this trade — wrong in obvious ways rather than subtle ones, which is the point. A rule you can read is a rule you can disagree with.
2. Log every recommendation with the number it meant to move. Each one records the rule that fired, the metric it targeted and that metric's value at the time. Seven days later and thirty days later, the system goes back and checks. This is the part that is the asset, and it is running now.
3. Rank the rules by measured lift. Which of your plays actually helped, for which kind of client. Not built yet — it needs enough recommendation-and-outcome pairs to say something real, and saying it early would be worse than saying nothing.
4. Only then, propose new rules. And a rule the system writes is proposed, never activated. A human promotes it. Not built yet, deliberately last.
Why the order is the whole thing
Five agencies at ten clients is fifty accounts. Mining fifty accounts for new rules produces confident nonsense, because at that sample size a coincidence and a finding look identical. What compounds isn't your accounts — it's the record of what happened after each suggestion. So step 2 is worth building years before step 4, and anyone selling you step 4 today is selling you a coin flip with a chart on it.
“We couldn't tell” is a real answer.
Every measured recommendation lands in one of four buckets: it improved, it stayed the same, it got worse, or we could not measure it. That last one is first-class on purpose.
The temptation is to quietly file unmeasurable outcomes as neutral, because it makes the numbers look complete. It also means a rule nobody could ever score sits in your average forever, dragging it toward the middle and telling you nothing. A system that can't admit an unknown will eventually give you a confident number built out of gaps.
- A recommendation with no metric attached can't be scored, so the build fails rather than letting it ship unmeasurable.
- Rules are scoped by trade from day one — something learned on roofing does not fire on a remodeler.
- Every rule carries how many observations it rests on, so twelve and four thousand never look the same.
Cost per sold job.
Ad platforms report cost per lead, because a lead is all they can see. What decides whether a client renews is what a job cost to win.
So no threshold in Peak Logic OS is an absolute number. What a client can afford to pay for a lead is worked out from what their average job is worth, the margin it carries and how often their leads actually close — and every campaign is judged against that figure rather than a benchmark borrowed from someone else's business.
A detector that fires on “$840 spent, no leads” is asserting a judgment it hasn't earned. That's 2.4× target at one shop and 10× at another. Same number, opposite verdicts.
- Per-creative revenue and return — not clicks. “Cheapest leads” and “made the most money” are frequently different ads, and the second is the one worth repeating.
- The chain stays connected — ad, creative, click, qualified conversation, appointment, revenue. One-tap check-ins in the client's own portal are what close it.
- It refuses below a floor — the creative brief will not name a winning angle off two data points. It says the sample is too thin, which is less satisfying and more honest.
The useful consequence: it can tell you to kill the creative with the great cost-per-lead and the terrible close rate, which is the ad that quietly loses agencies clients.
One brain, working in different seats.
Not four AIs. Parts of the same intelligence, each with its own list of things it is allowed to do — and a helper's permissions are always narrower than the person who invoked it. Adding a capability means editing that list, not persuading a chatbot.
Axon
Brings you what's waiting with the reason and the evidence attached, and flags a mission to the team when something needs another pair of hands. He does not spend your budget or message your leads — that isn't a setting he has switched off, it's an action he doesn't hold.
Myelin
The relationship side: clients waiting on a reply, jobs never confirmed as closed, accounts paying with nothing running. Can send an invoice for payment and re-send a portal invite that never got accepted — a different list of permissions from Axon's, deliberately.
Neuron
The morning sweep and the daily brief. He reads every queue, and anything that has waited past the point it should have becomes a routed item with a named owner. Once a day, never blocking — a manager opening their console at 2pm doesn't sit through morning theatre again.
Synapse is named and not yet built
Sales gets its own helper, with its own permissions, and it isn't finished. The Sales console says so where he would appear rather than leaving a blank space — “not built yet” and “nothing here” being different statements is the same principle the queues run on.
Delivered, not filed.
Intelligence you have to go looking for doesn't get used. It arrives in the flow of the day.
The morning brief
What happened overnight and what matters today, written for a human. Once per day and never blocking — open your console at 2pm and you don't sit through it again.
Account health
A live read across every account — response times, pipeline movement, client engagement — so a drifting account surfaces while you can still fix it rather than after the cancellation email.
The reason, every time
No suggestion arrives bare. Each states the evidence in a sentence you'd say out loud — which is what lets you overrule it, and what makes the next one arguable instead of oracular.
Coming: voice qualification
Leads will be able to choose a call instead of texting, and the AI picks up already knowing what they told us over SMS so it never asks twice. Deliberately last in the build order — it's a skin over an intent system that has to be testable without a microphone first, and nothing that reaches a lead or a budget will ever be voice-executable.
Common questions.
Is my clients' data used to train other agencies' systems?
How long before it tells me something I didn't know?
Does it act on its own?
What if clients don't answer the check-ins?
Can it invent its own rules?
See what your own numbers would say.
A 20-minute walkthrough on your accounts, not a canned demo.
Book a demo →