The L1 to L5 AI maturity framework for middle-market PE.
Everyone's using AI. Almost no one is banking it.
Framework v3 · updated September 2026
Most companies have given their people a chatbot: that is level one, and it makes people faster without changing what the business costs to run. The money shows up at level three, when software runs a workflow end to end and a person only reviews the exceptions; this framework is how our consultants work out which level a company is at, what the next one is worth, and what it costs to get there.
Three companies, three targets. Each moves to the level its P&L needs.Three portfolio companies, three targets. What ships in one transfers to the next.
- Month-end close, Industrials: moves from Level 1 to Level 2. 14-day close, all by hand. 10-day close, numbers trusted.
- Insurance approvals, Healthcare: moves from Level 1 to Level 3. 45 min per patient, staff on the phone. 20 min, chatbot finds each insurer's rules. 4 min by agent, nurse signs off · est. $310K/yr.
- Renewal quoting, Software: moves from Level 1 to Level 4. 2 days per quote. Same day. 1 hour, agent drafts. Minutes · est. $480K/yr.
- Month-end close, Industrials portco: moves from Level 1 to Level 2. 14-day close at the portco. 10 days; the fund sees numbers a week sooner.
- Insurance approvals, Healthcare portco: moves from Level 1 to Level 3. 45 min per patient, staff on the phone. 20 min, chatbot finds each insurer's rules. 4 min by agent, nurse signs off · est. $310K/yr to EBITDA.
- Renewal quoting, Software portco: moves from Level 1 to Level 4. 2 days per quote. Same day. 1 hour, agent drafts. Minutes · est. $480K/yr · playbook to the next portco.
Three companies, three targets. Each moves to the level its P&L needs.Three portfolio companies, three targets. What ships in one transfers to the next.
- Month-end close, Industrials: moves from Level 1 to Level 2. 14-day close. 10-day close.
- Insurance approvals, Healthcare: moves from Level 1 to Level 3. 45 min each. 20 min each. 4 min · $310K/yr.
- Renewal quoting, Software: moves from Level 1 to Level 4. 2 days. Same day. 1 hour. Minutes · $480K.
- Month-end close, Industrials portco: moves from Level 1 to Level 2. 14-day close. 10-day close.
- Insurance approvals, Healthcare portco: moves from Level 1 to Level 3. 45 min each. 20 min each. 4 min · $310K/yr.
- Renewal quoting, Software portco: moves from Level 1 to Level 4. 2 days. Same day. 1 hour. Minutes · $480K.
Azure
Five levels, in plain terms.
| Level | Name | In plain terms | Who does the work |
|---|---|---|---|
| L1 | Chat & Copilot | People ask a chatbot; it knows the internet, not your business | You |
| L2 | Connected Intelligence | The chatbot knows your documents and data; people still drive | You, with help |
| L3 | Workflow Agents | Software runs a whole step; a person checks the edge cases | Software, you review |
| L4 | Orchestrated Systems | Several agents coordinate; people handle exceptions only | Software |
| L5 | Autonomous Operations | The system notices work that needs doing and starts it | Software, governed by policy |
One workflow at a time. Turn on Claude, connect it to your data, then let an agent run the step with a person reviewing. Each move is scoped to a workflow, not the whole company, which is how the first one lands in 100 days.
L1 is two things: your people using a chatbot, and the governed data platform underneath that every later level depends on. The chatbot costs seats. The platform is where the $150K to $350K goes.
Your companies are piled up exactly where the value isn't.
Almost everyone has reached L1: chat and copilots. Almost no one has reached the tiers where AI actually moves EBITDA, and the distance between those two curves is the opportunity.
L1 is a plateau most companies mistake for the summit.
Chat and copilots are a first step, but the lift is capped and it never reaches the P&L. Letting people “play with AI” feels like progress. It’s not the same thing as banking it.
Why L1 stalls
- It’s people-dependent. The lift lives in whoever happens to use the tool well, and it leaves when they do.
- Change management eats it. Adoption stalls at the org chart, not the technology. Training decays the month after rollout.
- Cost-per-seat doesn’t move. Faster people still cost the same. The headcount, the handoffs, and the drag are all still there.
- Shadow-data risk. Company data leaks through personal accounts with no audit trail and no governance.
What breaks the ceiling
- The agent runs the step. Not a faster human: the workflow itself is executed by the system, end to end.
- Humans review the edge. People move to the high-judgment calls; the routine runs without them.
- Cost-per-task replaces cost-per-seat. Unit cost finally moves with volume. This is the EBITDA line.
- It compounds. A clean substrate and gated agents become the base for orchestration: L4, then L5.
Same workflow, different question: who does the work. That's the line between a faster headcount and a cheaper function.
Five tiers. Tier is a diagnosable fact, not a self-assessment.
A company is in L2 if its data platform, function-level AI deployment, and governance posture match the L2 definition. We complete the classification in a two-hour working session: no survey, no scoring rubric.
Chat & Copilot
AI makes individuals faster (drafting, summarizing, coding), but people still do all the work and nothing is integrated. A real first step, and the place ~80% of the mid-market stalls.
BOD’s role: We make sure L1 is governed (a prompt library, usage reporting, no data leaking through personal accounts) so the next move is clean.
Explore L1Connected Intelligence
AI grounded in your own data through retrieval, but a human still drives every step. Answers become cited and trustworthy instead of generic.
BOD’s role: We build the substrate (the clean data and retrieval layer an agent will need) and stand up the first grounded copilots that prove value.
Explore L2Workflow Agents
An AI agent runs a whole workflow step; a human reviews. Cost-per-task replaces cost-per-seat, the point where a function gets faster and cheaper at once.
BOD’s role: L3 is where cost per seat becomes cost per task. We get a company here one workflow at a time: an agent runs the step, a person reviews the exceptions.
Explore L3Orchestrated Systems
Agents run end-to-end and humans only edge-gate the high-risk decisions. Fleets of agents work in parallel against backlogs you couldn't hire against.
BOD’s role: We take a company here one workflow at a time, adding orchestration, durable execution and per-agent cost attribution once the workflow layer is proven.
Explore L4Autonomous Operations
Agents orchestrate other agents at organization scale. Largely aspirational in 2026, real today only in narrow domains like coding and customer experience.
BOD’s role: we build and govern (cost, evals and audit) the few functions where narrow L5 is genuinely shippable.
Explore L5Ten words, one sentence each.
Every first use of these terms across the framework links here and shows the definition on hover.
- Copilot
- A chat assistant a person uses to work faster. It doesn't do the work on its own.
- Agent
- Software that completes a multi-step task on its own: reads the input, decides the steps, uses your systems, produces the output.
- Workflow
- A repeatable business process with a start and an end, like closing the month or processing an invoice.
- Cost per seat vs cost per task
- Copilots are priced per person using them. Agents are priced per unit of work done. The second is what moves EBITDA.
- Retrieval (RAG)
- Wiring an AI model to your own documents and data so its answers come from your business, not the internet.
- Evals
- Automated tests that check an AI system gives the right answer, run every time something changes.
- Guardrails
- Rules that stop an agent from doing something it shouldn't, like sending an email to the wrong list.
- Human in the loop
- A person approves the agent's work before it takes effect. At L3 it's the default; at L4 it's the exception.
- Data foundation
- The governed, connected data platform underneath everything. If it's wrong, every agent above it is confidently wrong.
- Orchestration
- Coordinating several agents and systems so they work together on one process.
Estimate it yourself. Then let us qualify it.
Four inputs, a rough range. The biggest thing that decides the number is how much of the work is repeatable and already runs in software: that is what an agent can take over.
Pick the closest fit. The estimate updates as you go.
An estimate, not a quote. Assess replaces every assumption here with your own numbers and tells you which workflows to do first.
A typical mid-market profile.
Some workflows are rule-based and ready for an agent; others need a person in the loop. The first Build goes after the ready ones and the readout tells you how far to keep going.
Assumes $85K loaded cost per seat, 1,800 hours a year, and that 20% of back-office hours in scope can move to an agent at L3. Year one realizes 50 to 100% of run rate. Our fees: Build $75K to $350K per stage.
Savings are hours recovered at loaded cost. They become EBITDA only if the hours are redeployed or not backfilled. That decision is yours.
Qualify it in a two-hour diagnosticHow this relates to our prices. Assess is $25K to $50K and Build is $75K to $350K per engagement. The estimator's investment range adds the data foundation you may need underneath, plus platform and run cost, because that is what a company actually spends to move a level.
Calibrated to $25M to $500M+ revenue companies. Every input has a default, so the first number you see is the typical mid-market case. The two-hour diagnostic replaces the defaults with your data.
Vendor picks, the six pillars and the D1 to D10 capability spine live on the level pages behind a “For your CTO” toggle.
For your CTO: the technical annexTwo hours. One defensible readout.
Not a deck of maturity adjectives, but a diagnostic an IC can act on: the tier you're in, the tier to target, the six-dimension read, and the 90-day move.
Project Atlas, Industrials
- 1Stand up the L2 retrieval substrate on the existing warehouse, clean the FP&A source of truth.
- 2Ship one gated L3 workflow agent: the monthly variance close, human-reviewed at the edge.
- 3Instrument cost-per-task and evals before expanding to the next function.
The questions sponsors actually ask.
Start with an assessment.
Two hours, read-only. You leave with the tier you're in, the workflows ranked by ROI, and the 100-day pilot plan. Credentials deleted at engagement close.