Raphael GiovaniniRaphael GiovaniniBack to the site

AI Operator OS

Edition 1 · October 2026 · editor: Raphael Giovanini

The working standard for teams that put AI agents to real work in sales and service. Six disciplines, the Monday practice for each, and what to measure.

§ 1

Summary

An AI agent in operations is measured by what happens after it hands the customer to a human. Everything else follows from that.

Required
The ready-lead rule written with the sales team before any prompt. A pilot run on dirty data, at the worst hours. A way to switch the agent off without taking service down.
Forbidden
A pilot under lab conditions. Measuring response time alone. An agent that makes things up when it doesn't know. A twelve-month governance project before the first agent goes live.
Measured by
Of the leads the agent handed to a human, how many the human considered good. Not by messages sent or conversations handled.

§ 2

Six disciplines

These are not frameworks. They are what I do every week in real operations, with a practice you can start on Monday.

§ 2.1

Data before model

The AI project doesn't stall on the model. It stalls on the data.

Before

A twelve-month governance project that never ends.

Now

One flow, fewer than ten fields, one owner per field.

Traditional systems live with bad data because people compensate by eye. The model fills the gap and sends the error, well written, straight to the customer. The model is a commodity; your operation's data is the only asset nobody sells off the shelf.

In practice

  • Pick a single flow: one channel, one product, one type of service.
  • List the fields the agent will read to decide. Not the ones that exist.
  • Clean only those and give each one an owner. An owner is a person, not a department.
  • When the agent gets it wrong, ask which field was wrong before touching the prompt.

§ 2.2

Written rule

Deploying an agent forces the company to write its policy for the first time.

Before

The rule lives in the broker's head.

Now

The rule is on paper before the first prompt.

How far a discount can go, when it's still worth calling the customer who went quiet, how the tone changes when an investor shows up. None of it is written down. When a rollout slips, it's almost never the model being tuned: it's the company finally deciding what its policy is.

In practice

  • List the ten questions the team answers from memory every day.
  • Write the answer to each one with the person who has authority in the room.
  • Separate what is information (the agent answers) from what is commitment (the human answers).

§ 2.3

Handoff

The sale doesn't happen in the conversation. It happens in the handoff to the human.

Before

Measuring the agent by the conversation: volume, time, questions resolved.

Now

Measuring what happened after the handoff.

Hand off too early and the broker gets junk and stops trusting it. Too late and the customer goes cold. Without context and the customer repeats everything, which irritates more than not being served at all. A fast agent that fumbles the handoff delivers less than a broker working alone.

In practice

  • Define in writing what a ready lead is, with the sales team.
  • The human gets the conversation summary together with the lead.
  • Whoever receives it rates the handoff from the first week, not at the end of the pilot.

§ 2.4

Knowing when to stop

Answering fast is the floor. Knowing when to shut up is what sells.

Before

An eager agent that keeps talking after the yes and makes things up when it doesn't know.

Now

An agent that stops at the yes, at the question above its authority and when it doesn't know.

Every model was trained to please and to produce a plausible answer, not to admit a gap. If you don't force the agent to say it doesn't know, it invents square footage, deadlines or prices with the same confidence as when it's right. In customer service that is a commercial problem, sometimes a legal one.

In practice

  • When the customer says yes, the agent only schedules the conversation with the human.
  • Deadlines, payment terms and negotiation go to whoever has authority.
  • Explicit rule: if the information isn't in the knowledge base, it says it doesn't know and hands off.

§ 2.5

Buying with an exit

Buying AI is deciding on risk, dependency and reversibility. Not technology.

Before

Comparing model, demo and price.

Now

Asking what happens when it fails, who owns the data and how to switch it off.

Almost every disappointment with an AI vendor was decided before the signature, on a question nobody asked. If the vendor answers all seven clearly, you are buying well even without the most advanced model. If it dodges three, the most advanced model won't save you.

In practice

  • Ask for a result number with source and date.
  • Ask for a real conversation between the agent and a customer, not a scripted one.
  • Ask how long it takes to take the agent offline on a peak Thursday.
  • Ask to speak with a customer who walked away.

§ 2.6

People together

The project dies from the inside before it dies in the report.

Before

The team gets the finished agent and judges the technology by its first mistake.

Now

Whoever will receive the lead takes part from the pilot.

The broker who receives the lead didn't ask for an AI agent and doesn't trust it. The hard part of AI in operations isn't technical: it shows up where the technology meets people who didn't ask for it. Fear of being replaced, changing rules and internal communication are part of the rollout, not the post-project.

In practice

  • Put the team in the pilot as evaluators, not as an audience.
  • Say clearly what changes in each person's work and what doesn't.
  • Handle the psychosocial risks of the change with people who know the subject.

§ 3

Before and now

Nine points where the work of putting AI into operations has changed.

DeliverablePilot presentationAgent in operation and a written ready-lead rule
Main metricResponse timeGood leads at handoff to the human
PilotClean data and hand-picked questionsDirty data, bad hours, questions nobody picked
DataTwelve-month governanceOne flow, ten fields, one owner per field
When the agent errsTweak the promptAsk which field was wrong
When it doesn't knowInvents with confidenceSays it doesn't know and hands off to the human
PurchaseModel and priceRisk, dependency and reversibility
SolutionOff the shelfCommon base and a rule tailored to each operation
TeamGets the finished agentRates the handoff from the first week

§ 4

Principles

The six theses of the manifesto, each with the text that supports it.

  1. Every pilot works. Operations are where it breaks.The lab proves the model. Results show up in dirty data, unwritten rules and the handoff to a human.Read
  2. Your AI project stalls on data, not on the model.Models are a commodity. Your operation's data is the one asset nobody can buy off the shelf.Read
  3. Answering fast is the floor. Knowing when to stop is what sells.The number that matters: of the leads handed to a human, how many they considered good.Read
  4. AI was trained to please, not to be right.Flattery, verbosity and false certainty cost tokens and bad decisions. It can be fixed.Read
  5. Buying AI is a decision about risk, not technology.Seven questions whoever signs the budget can ask without reading a line of code.Read
  6. AI doesn't close the sale. It sets up whoever closes.What delivered results in CRMs and WhatsApp, and what only looked good on a slide.Read

§ 5

Want the six disciplines applied to your operation?

The AI Operations Diagnostic measures the six disciplines in two weeks and delivers a 30-day plan.

Apply for the Diagnostic

Prefer to start on your own? See the free tools

§ 6

Version

Edition 1, October 2026. This is a living standard: it changes when operations teach us something new, and every change is recorded here.

Spotted something wrong or missing? Send it on WhatsApp.