Five levels of enterprise AI maturity, and the eight problems at the last mile

"Enterprises have adopted AI" and "enterprises have not matured it" sound contradictory. They are measurements taken at different heights on one ladder, which is why both hold with two orders of magnitude between them.

McKinsey's 2025 survey, as the whitepaper reports it, found 88% of organisations using AI regularly in at least one function and 39% reporting enterprise-level EBIT impact. A January 2025 McKinsey survey, retold by a practitioner, puts at 1% the executives calling their own firm's AI use mature. BCG's 2025 study: 5% of companies reach value at scale, 60% see almost no substantive value.

The numbers are taken at different heights

Four things separate them. The unit: 88% counts organisations using AI in at least one function, a floor rather than an average, since one team qualifies a whole firm; 39% counts firms where the effect reached earnings; 5% counts scale; 1% counts executives willing to say "mature". The judge: broad panels, self-interested self-assessment, researchers' classification. The bar: using a tool has none, value at scale has a high one. The moment differs.

So the figures coexist: measures near a ladder's floor approach 100%, measures near its top sit in single digits. Gartner's 2025 finding that 63% of organisations lack AI data-management practices, or are unsure, measures a rung the same way, as does its 2024 estimate that 41–42% of prototypes reach production.

Five rungs, defined by what can be pointed at

A maturity level is useful only if two people looking at the same company land on the same rung. Adjectives do not achieve that; evidence does.

The whitepaper prints its maturity path as a figure whose level names do not survive into the text. The five below are this page's construction, built from four production questions, the eight last-mile problems, and the three gates of process, system and responsibility.

Rung	What you can point at	The question that settles it
1. Individual use	Work pasted into a chat window, results retyped by hand; no shared data, no session record	Turn the tool off for a week. If nobody files a ticket, rung one.
2. Assist at one step	AI at one step of one workflow; a named person approves and enters the result; edit rate unmeasured	Name the human between model and record. If you can, and no edit rate exists, rung two.
3. Write-back under permission	The model reads customer data and writes behind an approval step, with logs, rollback and a named approver	Reverse the last automated write. If nobody can say who approved it, this is rung two wearing rung three's name.
4. Operated	An eval set, versioning, monitoring, an on-call owner, and a finance-accepted outcome number over four weeks	Change the model version. Does anything rerun before users see it?
5. Organisational	Scenarios two and three ship without an outside team; permission designs and eval sets are reused; review rules adapted	Who would ship scenario four, and what would they build it from?

Microsoft's 2024 Work Trend Index gives rung one its signature: 75% of knowledge workers use AI, and 78% bring their own tools — shadow AI is what the bottom looks like.

Eight problems, eight different jobs

The last mile is usually described as one queue. It is eight, and they do not bind equally at every rung:

1. Strategy — the sponsor can name the goal but not the first scenario.
2. Data — three answers to one question, in the CRM, a spreadsheet and a group chat.
3. Process — the old process, unchanged, with a chat box beside it.
4. Systems — the system of record has no interface, or an undocumented one.
5. Permissions — nobody can say who may write, and the answer changes with the meeting.
6. Adoption — licence counts are healthy and weekly active users are not.
7. ROI — the business case is hours saved, and finance cannot find the line.
8. Operations — after go-live there is no owner, no eval set, no version.

Each rung moves which of the eight binds. At rung two permissions are theoretical; at rung three they stop the project; at rung four operations becomes the job. A firm that "did the data project" has not climbed a rung; it cleared one of eight items and returned.

Why a rung cannot be skipped

Skipping looks attractive: the lower rungs read as paperwork. A platform purchase can take a firm from rung one to something resembling rung four in a quarter. What moves is not the firm but the problem.

Put monitoring on a process whose metric was never defined, and the alerts are real while the thing monitored is not; the first incident review becomes an argument about what counted. The definition work was not skipped, only deferred into a system that must change while it runs.

The second cost is trust. Writing back to records before the habit of approval exists spends the first bad output, an irreversible one, teaching the organisation that the system is dangerous. Recovery then happens at rung two, minus a quarter.

The order is about evidence, not the calendar: regulation may push permissions ahead of adoption. What cannot be skipped is the artefact — an approved edit rate, a defined metric, a reversible write.

Placing a firm in ten minutes

Three questions, answered from memory, place a company faster than any strategy deck:

1. What does the tool touch — a person's clipboard, or a system's record?
2. Who approved the last automated write, and can it be reversed today?
3. Which number did finance accept, over what window?

A clipboard answer means rung one, whatever the investment says. A write with no name means rung two. No accepted number means rung four was never reached. Scope the next project to the rung above the answer, not the rung that photographs well — and if the metric is in doubt, the five-column metric table is the first artefact to build, because nothing above rung two is observable without it.

The test runs both ways: let a scenario run four weeks unchecked and the firm is at rung three with a pause, not rung four, which the four-week rule settles. Read as a hiring brief, the eight problems point to one person who can hold them all at once — which is why the constraint in China is not the number of engineers.