AI adoption fails when it becomes a launch event
Why professional firms need repeated review, testing, and practice change more than one-off AI rollouts.
TLDR
- A launch creates attention, but attention is not adoption.
- Professional AI use improves when teams bring real work back into repeated review: what helped, what failed, what needs checking, and what should change.
- The better metaphor is rehearsal, not rollout.
AI adoption fails when it becomes a launch event.
The pattern is familiar. A firm buys licences. There is a webinar. A policy goes around. A few enthusiastic people try the tool. A few cautious people avoid it. A few risky people use it quietly. Usage rises for a while, then becomes uneven.
The firm has launched AI.
The work has not changed much.
That is because professional AI adoption is not mainly a launch problem. It is a practice problem.
People have to learn where AI helps, where it fails, what it needs to know, who should check it, and which parts of the work should not move faster until the firm understands the risks.
That does not happen in one event.
It happens through repeated review.
Launches Create Attention
A launch is not useless.
It tells people something matters. It creates a shared moment. It gives permission to experiment. It may reduce the awkwardness of being the first person to try a new way of working.
But attention fades.
After the launch, people return to live matters, client calls, tax deadlines, audit files, project reviews, design crits, procurement issues, and the everyday pressure of work that already has a shape.
If AI does not fit into that shape, it becomes an extra thing.
Extra things die quietly.
Work Has To Rehearse
The better metaphor is not rollout.
It is rehearsal.
A professional firm gets better by repeatedly bringing real work into review. A legal team reviews matter strategy. An accountancy team reviews workpapers and assumptions. A studio reviews options, critique, client feedback, and production details.
AI should enter those existing moments.
Not as a grand announcement, but as a question:
- Did the draft help?
- Was the source map complete?
- What did the system miss?
- Where did it sound confident but weak?
- Which review step caught the issue?
- What should we change next time?
That is how adoption becomes real. The team does not merely use AI. It learns how AI behaves inside its work.
Why Access Is Not Enough
Research on generative AI at work shows that access can improve productivity in some settings 1. That matters. It means the tools can be useful.
But usefulness does not automatically spread evenly.
Adoption depends on fit. Daniel Russo's study of generative AI adoption in software engineering found that compatibility with existing workflows was a major factor 2. That should not surprise anyone who has tried to change professional habits. People do not adopt tools in the abstract. They adopt them when the tool makes sense inside the work they already have to do.
Broader research on workplace AI adoption also points to uneven uptake. Skills, training, digital maturity, employee voice, and the nature of the work all matter 3. Another study of a multinational workforce found that role fit, trust calibration, search skills, guidance, and knowledge infrastructure shaped adoption 4.
The plain version: people need more than access. They need a place in the work where the tool belongs.
The Hidden Failure Mode
The obvious failure is that people do not use AI.
The quieter failure is that people use it in ways the firm cannot see.
Shadow use can be worse than non-use. Someone drafts client work without checking sources. Someone pastes sensitive material into the wrong tool. Someone relies on an answer because it sounds complete. Someone creates a private workflow that saves time but leaves no review trail.
The firm then has two problems:
- AI is not improving the shared way of working.
- AI is still creating risk.
A launch does not solve that. A policy alone does not solve it either. People need regular, practical places to bring examples back for review.
What Repeated Review Looks Like
Repeated review is not a big programme.
It is a recurring habit around real work.
For a law firm, that might mean matter teams bring one AI-assisted research packet or draft to a weekly review and compare it against the responsible lawyer's judgment.
For an accountancy practice, it might mean reviewing where AI helped prepare workpapers, client queries, reconciliations, or management letter drafts, and where it introduced uncertainty.
For a design studio, it might mean critiquing AI-assisted research, concept options, copy, mood references, or production checks in the same way the studio already critiques design work.
The rhythm should ask:
- What did AI prepare?
- What did a person change?
- What source was missing?
- What should be prohibited?
- What can be templated?
- What needs a better example?
- What is reliable enough to repeat?
That is adoption. Not usage as a vanity metric, but learning as a habit.
Technology Becomes Real In Practice
Wanda Orlikowski's practice lens for technology in organisations is useful here. Her argument, in plain terms, is that technology does not have one fixed effect when it arrives. People make it meaningful through repeated use in practice 6.
That is exactly the issue with AI.
The same model can become:
- a toy;
- a private shortcut;
- a source of risk;
- a drafting aid;
- a review assistant;
- a better search surface;
- a way to prepare decisions;
- a way to preserve learning.
The difference is not only the model. It is the practice around it.
What To Measure
If adoption is treated as a launch, the firm measures logins.
If adoption is treated as practice change, the firm measures better things:
- Are first drafts easier to review?
- Are sources clearer?
- Are missing facts caught earlier?
- Are handovers cleaner?
- Are repeated corrections turning into better prompts, examples, or checks?
- Are people more explicit about what AI may and may not do?
- Are risky uses being brought into the open?
- Are clients, matters, projects, and files better understood?
Those measures are harder than usage counts. They are also closer to the value.
The Stanford AI Index is useful for the broad picture: AI capability and adoption keep advancing, but governance, evaluation, and responsible use remain uneven across organisations 5. That is the adoption problem in one line. The tools move fast. Work changes slowly.
The Plain Rule
Do not ask, "Have we launched AI?"
Ask:
"Where does AI now show up in our weekly work?"
If the answer is nowhere specific, adoption has not started.
If the answer is "in private, differently for each person", adoption is unmanaged.
If the answer is "in this review, this draft, this source check, this handover, this approval path", the firm is learning.
Launch days create attention.
Operating rhythms create adoption.
Sources
- Brynjolfsson, Li, and Raymond, "Generative AI at Work"
- Russo, "Navigating the Complexity of Generative AI Adoption in Software Engineering"
- Henseke, "Generative AI at Work"
- Ali et al., "AI Adoption Across a Multinational Workforce"
- Stanford HAI, Artificial Intelligence Index Report 2025
- Orlikowski, "Using Technology and Constituting Structures"
/ Start
Start with one operating area. Expand from there.
Begin with a focused review rhythm, workflow, or team where better operating context would immediately change the quality of preparation and judgment.