From agentic maturity to deterministic, self-sustaining systems
Part 1: The AI adoption maturity scale, L1 to L4
Every customer we talk to says they’re using AI.
For one, that means a few people asking Copilot to summarize their email. For another, it means agents updating records in the ERP overnight with nobody watching.
Both are true. They just aren’t the same thing, and they don’t need the same planning, budget or controls.
This series gives you a simple way to tell them apart, and a path for what comes after.
- The AI adoption maturity scale (this article): four levels, defined by what the AI can do for you, and where the tools on the market fit.
- L5, the self-improving mesh: turning what your agents learn into plain code, so the system gets cheaper and more predictable the longer it runs.
- Golden master testing for AI-assisted software: keeping AI-written code from turning into a maintenance problem.
- Eval-driven development: testing software whose behavior is itself AI.
Parts 3 and 4 stand on their own, at any level.
The scale in one picture
The four levels of AI capability, from answering questions to coordinating whole processes
Each level is defined by what the AI can do without a person doing it for them.
- L1, Answer. It answers what you ask. You do everything with the answer.
- L2, Know. It answers from your own organization’s content.
- L3, Act. It does tasks in your systems. It creates, updates, files and sends.
- L4, Coordinate. Several agents hand work to each other and run whole processes, often without anyone starting them.
Each step up adds capability. Each step also asks more of you before you turn it on.
The four levels up close
|
L1 Answer It answers what you ask. You do the rest. |
|
|
CAN DO Draft, summarize, explain, brainstorm and rewrite, using what you paste in and what’s on the public web. |
CAN’T DO See your files, email or systems. It can’t take any action. |
|
NEED IN PLACE An acceptable-use policy and a short list of approved tools. • The real risk is staff pasting customer data into a consumer chatbot your contracts don’t cover. |
HOW YOU PAY Per seat, or free. Easy to budget. |
|
IN PRACTICE A sales rep pastes in an RFP and asks for a first-pass outline of the response. It saves an hour. |
|
|
L2 Know It answers from your own content. |
|
|
CAN DO Answer from your SharePoint, OneDrive, Teams, email and business data, limited to what each person can already open. |
CAN’T DO Change anything. |
|
NEED IN PLACE Clean permissions and current content. • Copilot finds whatever a user can technically open, including the salary spreadsheet shared with Everyone in 2019. • Stale and duplicate documents give stale and contradictory answers. |
HOW YOU PAY Mostly per seat, with some usage charges once you add agents. |
|
IN PRACTICE A project manager asks what was agreed with the customer about change requests, and gets the answer with links to the contract and the meeting recap. |
|
|
FROM OUR WORK Most of the readiness work we do for customers is L2 work: oversharing, sites nobody uses, sharing links set to Anyone. It isn’t glamorous. It’s what decides whether Copilot is trusted after the first month. |
|
|
L3 Act It does tasks in your systems. |
|
|
CAN DO Carry out multi-step tasks: create a ticket, update a CRM record, file a document, start an approval. A person sets the goal and the agent works out the steps. |
CAN’T DO Work with other agents on its own. It handles one task at a time with the tools it’s been given. |
|
NEED IN PLACE Controls on the agent itself: • An identity, so you know whose permissions it’s using • Only the access it needs • An approval step for anything that can’t be undone • A log of every action, and a way to reverse one |
HOW YOU PAY By usage: per message, per action or in credits. The bill moves with how often the agent runs and how many steps each task takes. |
|
IN PRACTICE An agent reads incoming invoices, checks each one against the purchase order, and sends only the mismatches to a person. |
|
|
L4 Coordinate Agents hand work to each other and run whole processes. |
|
|
CAN DO Split a process across specialized agents that pass work along, often on a schedule or an event. One sorts the request, one gathers the facts, one drafts, one checks. |
CAN’T DO Promise the same path twice. The same request can take a different route on a different day. |
|
NEED IN PLACE Everything from L3, plus: • Monitoring across the whole chain • A named owner for every agent • Checkpoints where a person makes the call |
HOW YOU PAY The costs add up. Every handoff is another model call, and a retry can trigger several more. |
|
IN PRACTICE A request comes in by email. One agent classifies it, another pulls the account history, a third drafts the reply, and a person approves it before it goes out. |
|
|
FROM OUR WORK We run our own internal work this way. Each of our agents has its own area of work and at least two human owners. Every outbound message and every sign-in still goes past a person first. Agents do the work, people make the calls. |
|
What you need in place, level by level
What you need in place at each level. The requirements stack up as you climb.
The requirements stack. You don’t get to skip L2’s clean-up because you’ve jumped to agents. Every L3 and L4 agent reads before it acts, so if your content and permissions are a mess, your agents act on bad information.
Where the tools on the market fit
Where well-known AI tools fit on the maturity scale
The first thing to notice is that almost every tool spans more than one level. You don’t buy a level. The level comes from what you connect the tool to and what you let it do.
A Microsoft 365 Copilot license is a good example:
- Used as a chat window on the web, it’s L1.
- Grounded in your Microsoft 365 content, it’s L2.
- With Copilot Cowork, or an agent built in Copilot Studio that can update a list or send an email, it’s L3.
- Agents in Copilot Studio that call other agents take you to L4.
Same license, four very different risk profiles.
|
Tool |
Typical range |
What moves it up |
|
Microsoft 365 Copilot Chat |
L1 to L2 |
Adding agents grounded in SharePoint content |
|
Microsoft 365 Copilot |
L2 to L3 |
Copilot Cowork, and agents with actions |
|
SharePoint agents |
L2 |
Read-only by design |
|
Copilot Studio |
L2 to L4 |
Actions, then connected and child agents |
|
Power Automate with AI steps |
L3 |
AI steps inside a fixed flow |
|
Azure AI Foundry Agent Service |
L3 to L4 |
Multi-agent orchestration |
|
ChatGPT Enterprise |
L1 to L3 |
Company connectors, then agent mode |
|
Claude for business |
L1 to L3 |
Connectors to company systems, then task-running tools |
|
Gemini Enterprise |
L1 to L3 |
Workspace data, then agents |
|
Salesforce Agentforce |
L3 to L4 |
Agents working across CRM processes |
|
ServiceNow AI Agents |
L3 to L4 |
Agents orchestrated across IT and HR workflows |
These placements reflect how each tool is typically set up as of November 2026. Vendors ship new features every month, so check the current version before you plan around it.
AI coding tools are on a separate axis. They change how fast you build software. This scale is about what AI does inside the business once it’s running. Parts 3 and 4 cover the coding side.
What we see at customers
Levels coexist
Most organizations run L1 and L2 for everyone, L3 in a few departments and L4 in pilots. The useful question is what level each solution is at, and whether you’re running it with the controls that level needs.
The riskiest setup is an L3 tool with L1 controls
A department builds an agent in Copilot Studio. It can write to a SharePoint list and send email, it runs on the maker’s own account, there’s no approval step, and nobody reads the logs. It works fine until the day it doesn’t.
The budget model flips at L3
Per-seat pricing turns into usage pricing. Finance should hear that before the first agent goes live.
Assess each solution, not the company
For any AI solution, ask four questions:
- Who starts it: a person, an event or another agent?
- What can it reach: the chat only, your content, or systems it can change?
- Who makes the decisions: a person at every step, or the AI within limits?
- How many AI agents are involved?
The answers tell you the level, and the level tells you what has to be in place.
Why L4 isn’t the finish line
The natural reading of the scale is that L4 is the goal. We don’t think it is.
L4 is where capability peaks. It’s also where cost, unpredictability and upkeep peak. The same request can take a different path on different days, which is hard to test and hard to explain to an auditor.
The teams that get the most from AI use agents to work out how a job should be done, then turn the stable parts into plain code that runs the same way every time, for a fraction of the cost. That’s L5, and it’s the subject of Part 2.
Next in the series
Part 2, L5: the self-improving mesh. How to move AI work into ordinary code, cut usage costs, and send each remaining task to the cheapest model that can do it.
Where do your AI solutions sit on the scale? We’re happy to help you map them.
