The nine disciplines of an AI-native operating model
An operating model is usually drawn first and implemented afterwards. The AI-native operating model is found the other way round, by delivering one team and one use case at a time and recording what had to change for each release to land. What it contains is nine disciplines in three domains, and a firm holds the operating model in practice when every practice in all nine is true of it.
The nine are the places where a firm makes decisions about AI. Strategy is not a tenth discipline; it sits above all nine and decides which of those decisions the firm will take.
Practices
Each discipline is written as a skill: a set of instructions an AI assistant follows alongside the team. A skill states the discipline’s purpose, a few principles and a list of practices. A principle is a belief about how the work should run, and nobody is scored against it. A practice is a specific statement that is either true of a team or not, and it comes with a check against the team’s own records rather than against what anybody says in a meeting. Every result is kept, so you can see when a practice started holding and when it stopped.
We score practices, not disciplines. A score for “data” invites an average that nobody can act on; a practice that fails tells you exactly what to change.
A practice, written out in full
Before each release, the team writes down the four results it is aiming for.
- Checked by
- The team’s record shows its next release with a target for each of the four metrics, and shows that the sponsor set them.
- Who decides
- The sponsor sets the targets.
- Why
- A target set before a release can be missed, which turns the release into a test. A number chosen afterwards only describes what happened.
- The team is asked
- What is this release trying to change, in one sentence? Which of the four numbers would show it, and where does each stand today?
Organise
How the firm is arranged into teams, who holds which skills, and what the work costs and returns.
Structure 9 practices
How teams are drawn: what each owns, who decides what, and where work waits on someone else.
The unit is the delivery team. It owns one piece of the business from end to end, lists every member, AI agents included, and carries its own costs, whatever the org chart says.
- If a team must ask another team’s permission to ship, its boundary is in the wrong place.
- An AI agent that acts on the work belongs on the team list with a person supervising it; otherwise nobody answers for it.
Checked byAny piece of work that took longer than the release’s target carries a note saying what it waited for.
Talent 10 practices
Who does the work, what each role now involves, and how the firm keeps the skills it needs.
Once a machine writes the first draft, a role becomes the judgement the person applies to it. A capability counts only when people have shown it on real work, not when they have been trained in it.
- If machines take over the tasks juniors learned on, the firm has to rebuild how juniors learn, or it stops producing seniors.
- A skill nobody exercises any more is a skill the firm has lost.
Checked byThe team’s work record shows at least one finished piece of work this increment whose owners were all members of the team.
Finance 11 practices
What the work costs, what it returns, and who answers for each line.
Cost and value are attached to each release rather than to the month, so you can see what each release bought.
- A spending limit that depends on somebody noticing is not a limit.
- A saving is a cost that has left the accounts, not hours that were notionally freed up.
Checked byThe team’s policies include a spending limit that enforces itself, and the record carries a date on which it stopped spending with nobody intervening.
Build
What gets built, how it is built and checked, and the data it depends on.
Architecture 17 practices
The firm’s business capabilities, the systems behind them, the connections between those systems, and what the new work replaces.
Standards are written as tests a machine can run, shared terms mean the same thing on both sides of every connection, and the old core systems are replaced a slice at a time as each use case lands.
- A standard nobody can test is an opinion.
- Replace the old core piece by piece; never try to rewrite it in one go.
Checked byEach standard that applies to the team’s system exists as a test, and the latest run reports every part of it as passing or failing.
Engineering 15 practices
How work is scoped, built, checked and released.
Teams use the simplest tool that will do the job, write down what “working” means as tests before they build, ship in small steps that are allowed to change the plan, and keep a trace of how every output was produced.
- AI has made producing work cheap and checking it expensive, so checking is where the investment belongs.
- Use an AI agent only when plain code or a single model call will not do, and never give it more access than the person who asked.
Checked byEvery piece of work built with automation records the approach chosen and the simpler one that was tried or ruled out.
Data 11 practices
Where the data comes from, who owns it, whether it is good enough, and who may use it.
Data is made ready one use case at a time rather than across the whole firm first, and every answer the system gives carries a record of where its information came from.
- Get the data ready for the use case in front of you; a programme to prepare all the firm’s data first finishes after the use case needed it.
- Every correction a person makes should leave the system better.
Checked byThis increment, the team has tested the system against a set of questions with known answers, and traced each wrong answer either to the data it was given or to its reasoning.
Assure
What could go wrong, which agents exist and what they can do, and whether the decisions the AI makes can be traced and defended.
Assurance is a gate rather than a scale. A control either works or it does not, so a plan to improve assurance gradually has misunderstood what a control is.
Risk 11 practices
What could go wrong, who owns each risk, and whether the controls run.
Everything the firm runs is visible to the people answerable for it, and each obligation is met by a control that runs automatically rather than by a policy document.
- A review that has never stopped anything has never been tested.
- Risk adds up across the firm: fifty systems approved one at a time can share a model and a supplier that nobody assessed as a whole.
Checked byEach adopted risk policy states which check enforces it, and at least one policy records a date on which its check stopped a release.
Security 11 practices
Which AI agents exist, what each can reach, and what happens when one goes wrong.
Every agent is registered before it first runs, works under its own short-lived credentials, reaches no more than the person who asked it, and can be found and switched off.
- An agent can be hijacked by what it reads: a document can carry instructions.
- The danger is the combination: an agent that reads untrusted content, holds private data and has a way to send data out.
Checked byEvery agent the team runs appears in the register with its purpose, its permissions, an owner and a registration date earlier than the day it first ran.
Ethics 11 practices
What the AI does where no rule applies, who answers for it, and the rights of the people it affects.
The firm decides in advance how its AI behaves in cases nobody wrote a rule for, and judges its agents by everything they do together rather than one answer at a time.
- Harm from many agents builds up in steps too small for any single check to catch, so the whole population is what gets judged.
- Anyone affected has the right to know a machine was involved, to challenge the result, and to reach a person who can change it.
Checked byThe team has set a minimum score its agents must reach, somebody owns that threshold, and every agent added since it last changed has a dated result against it.
Foundations and metrics
Each discipline assumes that the established basics of running technology, which predate AI, are already in place. We write those basics down so that they can be checked, but the skills neither teach nor test them.
We teach the disciplines on the job: our workshops take a firm from choosing its first three teams to a governed plan for the next increment.
Above the nine sit four metrics, each answering one question: is the team getting better at adding value, at building, at controlling what it builds, and at the operating model itself? The sponsor sets a target for all four before every customer release. They are read as trends, and teams are never compared with one another.
- Value per cost
- What a release earned or saved, divided by what it cost to produce, measured a month after the release.
- Backlog to production
- How many days work takes from being requested to being live, leaving out time spent waiting on approval boards outside the team.
- Obligations discharged
- Of the rules that apply to the team’s work, the share enforced by a check that can stop a release.
- Practice conformance
- The share of the team’s practices that its checks show to be true.
Talk to us.
Contact us →