Lesson 1 of 6
What Changes When AI Runs in a Customer's Environment
See why a demo that works is not yet a deployment, and learn the four things to agree with the customer before any AI feature goes live.
5 minForward Deployed Engineering
After this lesson you will be able to explain what separates an AI demo from an AI deployment, and list what you must agree with a customer before an AI feature goes live in their environment.
A forward deployed engineer (FDE) is often the person who takes an AI feature from "it worked in the meeting" to "it runs every day at the customer". This course follows one such job from the first agreement to the handover.
The case
One made-up customer runs through every lesson. Tessel Energy supplies gas and electricity to homes. About two thousand written messages arrive each day by web form and email. A triage desk of six people reads every message and sends it to one of five queues: emergency, billing, meter, moving home, or complaint.
At busy times a message can wait three hours before anybody reads it. That includes messages about a smell of gas. Tessel has asked your company to deploy an AI model that reads each message as it arrives and chooses the queue.
In a demo, the model sorted twenty messages correctly and everybody was impressed. Your job starts after that.
Why a demo is not a deployment
A demo answers one question: can the model do this task at all? A deployment has to answer harder ones.
- What happens on the day the model is wrong? Over two thousand messages a day, even a very good model will be wrong many times a week.
- Which mistakes matter? Sending a billing question to the meter team costs a day. Sending a gas leak to the billing team could cost a life.
- Where do the messages go when the model reads them? They hold names, addresses and account details, and they belong to Tessel's customers.
- Who will notice if it gets worse in six months, and who can fix it once you have moved on?
None of these is about how clever the model is. They are about the system around it, and that system is what you build.
Agree the task exactly
Write down what goes in and what comes out. Here a written message goes in, and exactly one of five queue names comes out, with a number saying how sure the model is.
Be strict about the list. If the model may answer with anything else, such as "unclear" or a queue that does not exist, decide now what the system does with that. Vague outputs cause trouble later.
Agree what each mistake costs
Ask the customer to describe each kind of mistake and what follows from it. At Tessel the head of customer service and the leader of the emergency team gave three answers.
- An emergency sent to another queue may not be read for a day. This must never happen.
- An ordinary message sent to the emergency queue wastes a few minutes of the emergency team's time. A small number is acceptable.
- Any other message in the wrong queue is passed on by hand and answered about a day late.
The first answer becomes a hard rule: a result that breaks it is a failure, whatever the other numbers say. The others are covered by a pass mark, a share of cases that must be right. Tessel agreed on 85 in every 100, which is about what the triage desk manages on a busy day.
Agree both before anyone sees results. If you choose the pass mark after the first test, you will choose the number you happened to reach.
Agree where the data may go
A model has to run somewhere, and the customer's messages travel to it. Ask the customer's security lead which models are approved, where they run, and whether anything the model reads is stored or used by the provider. Many customers will only allow a model that runs inside their own cloud account. Some will not allow certain data to be sent at all.
Get the answer in writing before you build. A deployment that works perfectly and uses an unapproved service will be switched off.
Agree who checks, and what happens when the model is unsure
No model is right every time, so decide what surrounds it. Three questions cover most cases. What does the system do when the model is not confident? Is there any rule that runs without the model, for the mistake that must never happen? And can the customer turn the automatic part off in one step?
At Tessel, an unsure message goes to the triage desk exactly as it does today. You will build all three answers in lesson 5.
The deployment sheet
Put the agreement on one page.
# Deployment sheet: Tessel Energy message sorting
## Task
Read each written message. Choose one queue: emergency, billing,
meter, moving, complaint. Give a confidence from 0 to 1.
## What mistakes cost
Emergency sent elsewhere: may wait a day. Must never happen.
Other message sent to emergency: minutes lost. A few are accepted.
Any other wrong queue: answered about a day late.
## Rules for "good enough"
1. No emergency case in the test set goes to another queue.
2. At least 85% of the test set is sorted correctly.
Both agreed before testing, with the head of customer service.
## Model and data
A model approved by Tessel's security lead, running in Tessel's own
cloud account. The provider stores nothing. Confirmed in writing.
## When the model is unsure
The message goes to the triage desk, as today.
## People
Head of customer service: decides whether it is good enough.
Triage desk lead: supplies and labels the test cases.
Emergency team lead: agrees the emergency rules.
Security lead: approves the model and the data flow.Set up for this course
You need Python 3.9 or later and nothing else. The course calls no AI service, so you need no account and no key. A small stand-in plays the part of the model, and everything you build around it would stay the same with a real one.
Create a folder called deploy-ai and keep every file from this course in it. Where the course writes python, use python3 if that is the command on your machine.
Key takeaways
- A demo shows that the model can do the task; a deployment has to deal with the times it cannot.
- Mistakes do not cost the same, so find the one that must never happen and write it down as a rule.
- Agree what "good enough" means, in numbers, before anyone sees a result.
- Know which model may be used, where it runs and where the customer's data goes, and get it approved.
- Decide who checks the output and what happens when the model is unsure.
Try it
Write a deployment sheet for a different case: a veterinary practice wants an AI model to read messages from pet owners and decide whether each one needs an appointment today, an appointment this week, or advice by message. Use the same headings as the Tessel sheet. A good result names the one mistake that must never happen, gives a pass mark with a number, says what happens when the model is unsure, and lists at least one question about where the owners' messages may be sent.
Official documentation
Written with AI and checked automatically. Nobody has run these steps by hand, and tools change, so a command may behave differently in the version you install. Spotted a mistake? Tell us.