Framework

An AI maturity model where the top level is one you should not want

Most maturity models are a ladder with the vendor at the top. This one has five levels defined by who initiates the work, and level four — autonomous finance — is marked as something we do not build and think nobody should. The useful question is which level fits you, not how high you can climb.

Where are you now?

Three questions about size, stack, and process. We will tell you which level you are at and what the next step costs.

1 / 3
Defined by who initiatesLevel 4 marked as undesirableReadiness checklist included
Software initiates, person governs

A bill arrives; an agent codes it and routes it.

The work starts without a person. Requires an authority model, an audit trail, and a policy engine — which is why it cannot be retrofitted onto a screen-driven system.

The distinction that matters is not how clever the model is. It is who is responsible for noticing the work exists.

Readiness

Six things that must be true before level three.

None of these are about buying software. They are the conditions under which handing work to an agent produces a good outcome rather than an expensive disappointment.

A chart that means something

Fewer than about 250 accounts with real dimensions. An agent learning from a chart where the same cost sits in four places learns the inconsistency.

A close that completes

If the close does not currently finish, automating parts of it accelerates a process that does not work. Fix the sequence first; it is cheaper.

History to learn from

Twelve months of consistently coded transactions is the input. Without it the first quarter is spent building what should already exist.

An approval matrix that is real

Written down, current, and actually followed. An agent routing to your documented matrix is useless if everyone routes around it.

An owner with authority

One person who can decide that a policy is acceptable. Automation stalls at exactly the same place implementations do — an unmade decision.

Systems that can be read

An API or a usable export. A ledger nobody can read is a constraint no amount of capability overcomes.

Moving up a level is mostly not a software purchase

Every company we work with can technically operate at level three on day one — the software supports it immediately. What determines whether it goes well is the six conditions above, and five of them are about your own processes rather than about us.

The one that stalls most projects is the last kind: an owner with authority. Agentic automation requires somebody to decide that a policy is acceptable, and that decision has consequences. Where it gets escalated to a committee, the agents sit at level one indefinitely and the investment produces speed of preparation rather than autonomy.

Almost every company stalls at the same place: not the technology, but the person who has to sign off on a policy having consequences.

A realistic first year

  • Months 1–2. Level 2 across the board — connect, report, and ask questions. No workflow changes and no risk, and it is where the corpus starts building.
  • Months 2–4. Level 3 at authority Level 1 on accounts payable. Everything drafted and reviewed. This is where the labelled data comes from.
  • Months 4–6. Authority Level 2 for recurring vendors under a threshold, plus dunning inside a policy. The first genuine autonomy.
  • Months 6–12. Reconciliation and close, then the analytical agents once three or four clean periods exist to reason over.

Notice that most of the first year is at level three with low authority. That is not caution for its own sake — it is where the corpus is built, and skipping it means automating on a model that has not learned your conventions.

Why level four is not the goal

Autonomous finance — software that closes the period, releases the payment, and signs the assertion — is technically approachable and we think it is the wrong destination.

Closing a period asserts that the numbers are right, to an auditor, a board, or a lender. That assertion has value because a person is accountable for it. Automating it does not remove the work so much as remove the meaning, and the first time it matters is the first time somebody asks who is answerable.

So this model tops out at level three by design. A vendor promising level four is either not thinking about accountability or is describing a product for a market that does not exist.

Our interest, disclosed

We sell software that operates at level three, so a model placing level three as the sensible destination is convenient for us. The check worth applying is whether the argument against level four holds independently — we think it does, and it costs us the more ambitious claim.

Questions

Common follow-ups.

Can we skip straight to level three?
Technically on day one. Whether it goes well depends on the six readiness conditions, five of which are about your processes. The assessment will tell you which are missing.
What if our chart of accounts is a mess?
Fix it first, and it is usually the highest-value part of an engagement anyway. An agent learning from an inconsistent chart learns the inconsistency and reproduces it faster.
How long to reach level three with real autonomy?
Four to six months for a first workflow at Level 2 authority. Most of the intervening time is corpus building, which cannot be compressed by spending more.
Is level two worth reaching if we stop there?
Yes, and some companies should. Reporting, integration, and a copilot on top of an existing ledger deliver real value with no workflow change and no risk.
Why do you exclude autonomous finance?
Because closing a period and releasing funds are assertions that need somebody accountable. Automating them removes the meaning rather than the work.

Find out where you actually are.

Three questions about size, stack, and process, and a straight answer about what the next level would take.