Skip to content
Français

Governance that helps teams ship

  • Lesson 1
  • 25 min
  • Module 1: What governance is for

Where we are

You already work with rules at work: who may sign a contract, who may open the safe, who checks an expense claim. AI arrives more quietly. A team builds a tool, or a vendor switches a feature on, and nobody asks who said yes. In this lesson, we see what goes wrong with no rules and with too many, starting with the rules of the road.

Plays from YouTube. Nothing loads from YouTube until you press play. Privacy

What you will be able to do

See what AI governance is for. With no rules, AI that nobody knows about reads customer data. With too many rules, teams hide their AI. Good governance sits between the two, so teams ship and the organisation can prove what it does.

Before you startWhat to know and to have readyAbout 25 minutes

What you need to know

Nothing yet. This is the first lesson of the course. You need to read a short policy, a table and a simple report. There is no code anywhere in the course.

What you need

  • A browser. The course’s practice files are on GitHub, in the folder C33 of the labs repository. Start with its README.
  • A document tool: a word processor, a text editor or a spreadsheet.
  • No account, no installation and no money.

This course is not legal advice. It explains laws and standards in plain words, with the official source. Ask a lawyer about your own organisation.

Think of a town with no traffic rules. Anyone drives any vehicle, at any speed, on either side of the road. Most trips still end well. But at every junction, two drivers guess what the other will do. When there is a crash, nobody can say who was wrong. There is no number plate, so nobody even knows whose car it was.

Now think of the opposite town. Every trip needs a permit from the town hall, and a permit takes three weeks. A trip to the shop needs the same permit as a truck full of fuel. What happens? People do not stop driving. They drive without a permit, on back roads, and they hope nobody looks. The town hall has a lot of rules, and it knows less about the traffic than before.

Real towns use a highway code. Every car has a number plate and a registered owner. A bicycle needs no licence, and a truck needs a special licence and regular inspections. A driver can show a licence and proof of insurance when the police ask. The rules let people drive, and they make the dangerous trips careful.

Larkfield is a home-and-garden shop that sells online to customers in Canada and in the European Union. Larkfield, its people and its systems are constructed for this course. Larkfield now uses AI in many places. Some systems its own teams built. Others arrive inside the tools it buys.

With no rules, AI arrives that nobody knows about. Larkfield’s platform team found a script in a shared folder (constructed). Each week, it sends every customer message of the week to an AI service. Then it emails a summary of complaints to a mailing list of 14 people. Larkfield has no contract with that AI service. The script uses a personal key, and the key belongs to a contractor who has left. Nobody approved it, and nobody owns it.

A second case needs no script at all. TicketHarbor (made up for this course) sells an add-on that Larkfield plugs into its own help desk. In an update, TicketHarbor added an AI reply suggester to it. It switched the feature on for every agent by default. Customer messages now go to an AI feature that nobody at Larkfield chose.

With too many rules, teams hide their AI. Imagine that Larkfield answers the script with a strict rule: every use of AI needs the board’s signature. A governance board is a small group of managers that decides about AI. This board meets once a month. A team that wants meeting notes waits weeks for a decision about a small tool. So it buys the tool on a company card and tells nobody. Larkfield already has one tool like that. Omar’s product team bought a meeting-notes tool, MinuteMoss (made up for this course), on a company card, with no review.

Both towns end in the same place: AI that nobody knows about and nobody checks.

Three panels side by side. Panel 1, too few rules: a script sends customer messages to an AI service; nobody approved it, nobody owns it; result: AI that nobody knows about or checks. Panel 2, too many rules: every AI use waits weeks for the board; teams use AI quietly, on a company card; result: AI that nobody knows about or checks. Panel 3, proportionate rules: every system is listed and has one named owner; small risk gets a quick check and big risk a careful check; result: teams ship, with a record of each decision. The first two panels are red, the third is green.
Too few rules and too many rules fail in the same way. Proportionate rules are the only panel where the organisation knows what it uses.
Text description of the diagram

The picture has three panels. In the first, too few rules: a script sends customer messages to an AI service, nobody approved it and nobody owns it. In the second, too many rules: every use of AI waits weeks for the board, so teams use AI quietly, on a company card. Both panels end with the same result, in red: AI that nobody knows about or checks. The third panel, proportionate rules, is green. Every system is listed and has one named owner. A small risk gets a quick check and a big risk a careful check. The result: teams ship, with a record of each decision.

AI governance is the set of rules, roles and records around an organisation’s AI. They decide which AI systems it uses, who answers for each one, which checks each one needs and how it proves all this. The word comes from the Latin word for steering a ship. That is a useful picture: governance steers the work. It does not stop the work.

Good governance has four goals, and each one is a part of the highway code:

Four pairs, each an arrow from a rule of the road to a goal of AI governance. Every car has a number plate leads to known systems: a list of every AI system, bought ones too. Every car has a registered owner leads to clear owners: one named person for each system. A bicycle needs no licence, a truck needs a special licence and regular inspections, leads to proportionate checks: the bigger the possible harm, the more careful the check. A driver can show a licence and proof of insurance leads to proof: a record that shows who decided what, and why.
The four goals of AI governance, each with the rule of the road that it copies.
Text description of the diagram

The picture shows four pairs. On the left is a rule of the road. On the right is the goal of AI governance that it copies. A number plate on every car becomes known systems: a list of every AI system, bought ones too. A registered owner for every car becomes clear owners: one named person for each system. A bicycle with no licence and a truck with inspections become proportionate checks: the bigger the possible harm, the more careful the check. A driver’s licence and insurance papers become proof: a record that shows who decided what, and why.

  • Known systems. You cannot govern what you cannot see. Bought features count, because the customer data goes to them too.
  • Clear owners. One named person answers for each system and can say “stop”. The next lesson is about this.
  • Proportionate checks. “Proportionate” means in proportion: a small risk gets a small check. A tool that summarises a policy for staff does not need the same review as a tool that refunds money alone.
  • Proof. A decision that nobody wrote down cannot be checked later, by a manager, an auditor or a court.

Notice what is not on the list: “more approvals”. An approval step that nobody reads adds waiting and no safety. Governance is good when the teams can ship and the organisation can show what it did.

A tiny example: three requests in one week

Section titled “A tiny example: three requests in one week”

Farida (she) joins Larkfield as its first risk and compliance lead. You follow her through this course. Nadia (she), the head of customer operations, sponsors her work. In her first week, three requests arrive (constructed). Farida reads each one with the four goals:

Request Known? Owner? Check in proportion Proof
A support agent asks: “May I use the document assistant to check our returns policy?” yes, in the registry Mei none: it is an approved use of an approved tool none needed
Omar wants MinuteMoss (made up for this course) for calls with customers no Omar says it is his a vendor check and a privacy check: it records people outside Larkfield a short decision record
Omar wants the refund workflow to refund small amounts alone yes the card says “Support team” a careful review: the system would act on customers’ money with no person a full decision, signed by a named owner

Read the table row by row. The first request needs no meeting at all: the highway code says “go”. The second needs a few days and two short checks. The third changes what the system may do alone, so it gets the most careful look. It also shows a gap: a team cannot sign anything. Later in the course, you measure exactly how much that one change moves the risk.

The same four questions gave three very different answers. That is proportion.

If you took the course Leading AI Products and Teams, you know a short version of this idea. Its lesson Governance in practice set three approval levels from a risk score: low 1 to 4, medium 5 to 9 and high 10 to 25. This course builds the full system around such levels: the list, the owners, the checks and the proof.

On her first day, Farida finds three things. The platform team’s list shows eight AI systems, and only four of them are in the platform registry. The registry is the platform team’s list of the systems that run on Larkfield’s own platform. Larkfield’s AI policy is six lines long, and it names no owner. And most people think that somebody else is in charge of AI.

That last point is the most important. A risk lead cannot govern eight systems alone, and she should not. Governance works when each owner does a small part of it as part of the normal work.

Nadia asks the team what Farida’s job should be.

  • Tomás (AI engineer): “Governance is the legal team’s job. They read the laws, and they tell us what is forbidden.”
  • Omar (product manager): “Every new rule slows us down. We ship once a week, and the customers notice.”
  • Grace (support lead): “My agents use three AI tools now. I only chose one of them.”
  • Sam (they) asks: “If a vendor switches a feature on, is that our decision or theirs?”

Before you read on, answer these:

  1. Is governance the legal team’s job?
  2. Does every rule slow Omar down?
  3. Answer Sam’s question.

The course’s answer. Legal advice is one input, not the whole job. The lawyer answers “what does the law ask?”. Only Larkfield can answer “which systems do we use, who owns each one and what did we check?”. Those answers come from the teams. So governance belongs to the business, with help from legal.

Omar is right about bad rules. A rule that treats the document assistant like the refund workflow slows him down for nothing. A proportionate rule lets him ship the small changes the same day, because their path is clear.

Sam’s question has a short answer: the vendor switched the feature on, but the decision to use it is Larkfield’s. Customer messages are Larkfield’s customers’ data. A bought feature needs an owner at Larkfield as much as a built one.

What would change the answer: in a very small company, one person may do all of these jobs. The four goals stay the same. The list is shorter, and the owner’s name is often the same.

Start governance when the first AI system touches customers, staff or money, not after the tenth. It is cheaper to name an owner on day one than to find one after an incident.

Governance costs time: a list to keep current, owners who read evidence, and records to write. Keep that cost low. Give low-risk uses a fast path with no meeting, and save the careful review for the few systems that can do real harm. A rule that costs more than the risk it reduces is a bad rule, and people will go around it.

Common mistake

A policy that bans AI tools

Symptom: the policy says “No AI tools without the board’s approval”. No request reaches the board for months. Meanwhile, the company card statement shows new AI subscriptions every month.

Cause: the ban made the approved path harder than the hidden one. People still need the tools, so they stop asking.

Fix: make the approved path the easy one. Publish a short list of approved tools and the data that may go into each, with a fast path for low-risk requests. Then ask the finance team for the card statement, and add every AI tool on it to the list.

Case studies in this course

At the end of the course, you use what you learn on a realistic project, from the data to the recommendation. Complete at least one for the certificate; the others are optional.

  • A tribunal case about a chatbot's answerRead a public tribunal decision about a company held responsible for what its chatbot told a customer, and write the governance review: the policy, owner, review and record that would have prevented it.
  • Learn from a public algorithm failureRead the official reports on a public-sector automated-decision failure, and write the impact assessment that should have been done before launch.
  • Review a published AI registerUse published algorithmic impact assessments under an open licence to build an inventory and risk tiers, and find what the published records leave out.

Check your understanding

Name the four goals of AI governance, and give one Larkfield example of a goal that is not met today.

Show the answer

Known systems, clear owners, proportionate checks and proof. Examples: the weekly digest script was unknown until the platform team found it, so the systems were not known. The script also has no owner, and neither has the reply suggester that the vendor switched on. Only four of the eight systems are in the registry. The policy names no owner, so nobody keeps it current.

Check your understanding

A manager says: “After the script, we must be safe. From now on, every AI use needs the board’s signature.” Predict what happens, and propose a better rule.

Show the answer

People still need the tools, and the board meets rarely, so the requests wait for weeks. Teams stop asking and buy tools quietly, as Omar’s team did with the meeting-notes tool. The organisation then knows less, not more. A better rule is proportionate. Approved tools and low-risk uses get a fast path with no meeting. New tools that touch personal data get a short check. Only the few high-risk changes go to the board.

Check your understanding

Tomás says: “We only use AI from big vendors, so the vendor is responsible.” Use the reply suggester to answer him.

Show the answer

The vendor built the feature and switched it on, but Larkfield decides to use it with its customers’ messages. A customer who gets a wrong reply complains to Larkfield, not to the vendor. So the feature needs an owner at Larkfield, a check of what the vendor does with the data, and a record of the decision. The vendor’s size changes nothing in this.

Recap

  • AI governance is the set of rules, roles and records around your AI: which systems you use, who answers for each, which checks each needs, and the proof. It steers the work; it does not stop it.
  • Too few rules and too many rules end in the same place: AI that nobody knows about and nobody checks.
  • The four goals: known systems, clear owners, proportionate checks, proof. More approvals is not one of them.
  • A bought AI feature needs an owner at your organisation, as a built one does.

You can now say what governance is for. But “clear owners” is easy to say and hard to do: a refund touches support, finance, privacy and the customer. Next, you give each system one accountable owner and fill a responsibility table.

Next: Who is accountable