A top‑twenty firm got almost its entire staff using generative AI within months, and the spend that mattered was not on licences.
Most firms measuring AI adoption are measuring the wrong thing. Seats deployed is a purchasing metric. It tells you what finance approved, not what changed.
The number worth chasing is the share of people still using the tools some months after the announcement, once the novelty has worn off and the work has piled back up. At Armanino, a top‑twenty US accounting firm of roughly three thousand people, that figure was 98 percent.
What follows is how they got there, described in enough detail to copy. Almost none of it is about software.
Decide the Governance Before the Tooling
The firm built a two‑tier structure, and the sequencing matters more than the shape.
Digital group
A representative from every business unit, plus legal, IT and risk in the room from the start. Guardrails move down.
Its own digital group
Works out what AI means for its own jurisdiction, practice line and risk profile, and prioritises accordingly. Ideas move up.
Figure 1. The two‑tier governance structure described in the masterclass.
At the top sat a firmwide digital group with a representative from every business unit, plus legal, IT and risk in the room from the start. Beneath it, each business unit ran its own digital group, with the latitude to work out what AI meant for its own jurisdiction, practice line and risk profile, and to prioritise accordingly. Ideas moved up; guardrails moved down.
The reason for the lower tier is the part firms usually miss. A tax practice in one jurisdiction and an advisory practice in another face genuinely different constraints on what can touch client data, and a single central committee issuing uniform rules will either be too permissive for one or too restrictive for the other. Devolving the interpretation let each unit move at its own pace without anyone going off on their own.
Note also the deliberate word choice. They called them digital groups, not AI groups, on the reasoning that AI is becoming a component of every piece of software the firm already owns rather than a separate category to be governed separately. That framing ages better than the alternative.
The operating philosophy came from software development rather than accounting: Modern Agile’s four principles, of which two did most of the work here. Make people awesome, meaning every deployment is judged on whether it improved someone’s day rather than on whether it used AI. And experiment and learn rapidly, meaning you do not buy a licence, declare the problem solved and lock the door.
Train Before You Scale, and Train Properly
Here is where the money and the attention actually went.
- 1Put it in every employee’s education plan
Not an optional lunch‑and‑learn for the curious. Generative AI literacy became part of what every person was expected to learn, on the reasoning that this is now foundational rather than specialist.
- 2Teach the why, not just the how
People treat these tools as an omniscient oracle until somebody explains that it is a very good statistical model for language. Understanding the mechanism is what stops both over‑trust and dismissal.
- 3Micro‑learning, not a course
Two‑minute videos giving one piece of advice a person can use the same day, on the format people already consume everything else in.
- 4Small cohorts, hands on keyboards
No mass licence drop. Small interactive groups where every participant is working in the tool during the session rather than watching a demonstration of it.
- 5Everyone builds an agent
Each participant leaves having built their own custom agent. The point is not the agent. It is that they have crossed from consumer to builder, and know they can do it again.
Figure 2. The training design as described. The expensive ingredient is not the content, it is the decision to run small cohorts rather than a single broadcast, which multiplies facilitator time by the number of groups.
The stated ambition was to turn one idea about generative AI into three thousand. The image used for it is worth keeping: a trolley takes you somewhere useful along a fixed track, but a car lets each person drive where they actually need to go. Firms that pre‑select the approved use cases get the trolley, and get exactly the value they imagined in advance and no more.
The result was a firm where custom bots proliferate from the bottom up. One example: a knowledge bot on 1099 reporting, built by a staff member, which answers common questions from published guidance rather than internal material. Nobody at the centre would have specified it.
Then Stop Making People Swivel Chair
Broad enablement is stage one and it has a ceiling. Everything above depends on a person deciding to open a separate tool, paste something in, and bring the answer back. That works, and it is the reason the usage number matters, but it taxes every single use with a context switch.
Stage two is putting the model inside the work rather than beside it, and the example given is instructive precisely because it is unglamorous.
In the firm’s own audit platform, an integration reads what is already on the auditor’s screen, the test being performed and the evidence attached to it, and drafts the note that goes to the client. No chat window, no copying, nothing for the user to remember to do. The same integration scans documents submitted for review and tells the auditor what arrived, so nobody opens a PDF only to discover it is the wrong asset and has to send it back.
The reported effect on the work is worth more attention than the time saved. Junior auditors could produce a professionally worded client note at a standard that previously took seniority to reach. That is not headcount reduction; it is compression of the distance between a first‑year and a competent output, which in a profession with a two‑decade talent shortage is the more valuable outcome.
The same pattern shows up in engagement letters, where inbound letters are summarised, classified and pushed into the billing system, removing the manual step entirely.
Measure on Two Axes, Because One Will Mislead You
The macroeconomic evidence supports the second branch more than most executives expect. Research by Erik Brynjolfsson and Avinash Collis measured what people would need to be paid to give up generative AI tools for a month and put the resulting consumer surplus at roughly $97 billion in 2024, against about $7 billion of actual revenue booked by the major providers in the US. The value is real and it is landing with users, which is precisely why it does not show up cleanly in the aggregate figures a CFO watches.
That is worth stating plainly to a partner group before a rollout rather than after. If your justification depends on seeing the productivity in the firm’s top line within two quarters, you have set a test the technology will fail on your timeline even where it is working.
The Uncomfortable Part
One piece of this does not travel as easily as the rest. The presenter’s own standard is that everyone should use AI twenty times a day, and that a draft arriving without AI involvement gets sent back.
That is a strong position and it will not suit every firm. But it names the real variable. Adoption at this level is not achieved by making tools available; it is achieved by changing what is expected, and then equipping people to meet the expectation. Firms that raise the expectation without the training get resentment. Firms that provide the training without raising the expectation get a well‑taught population that carries on working the old way.
To follow the full rollout, from the digital group structure through the training design and the return‑on‑investment framework, watch A Strategic Blueprint for AI Adoption in Accounting on Miles Masterclass.