What does AI adoption actually mean inside a company?
Adoption means AI has become part of how specific tasks are done, by the people who do them, with a check that keeps quality up. It does not mean licences bought, logins counted or a launch email sent. A team where three people use AI daily on approved tasks and review the output carefully has adopted it further than a team where everyone logged in once.
It helps to separate three questions. Access: does each person have an approved tool and know the rules? Use: do they reach for it on the tasks it suits? Quality: does the work that comes out survive review? Most rollouts solve the first question, measure the second roughly and ignore the third.
Why do so few people use AI after the rollout?
Low use after rollout is common and has ordinary causes: people cannot see which of their tasks AI helps with, the rules feel unclear so they avoid the tool to stay safe, an early poor output convinced them it was not worth the effort, or they simply have not had time to practise. In the UK government's AI Adoption Research, businesses already using AI reported that on average 30% of their staff were using it, and 77% said fewer than half of their staff were. The same survey found the most common reasons for not adopting AI at all were a lack of identified need and limited AI skills and expertise. These are self-reported figures from 3,500 UK businesses interviewed in early 2025.
Notice that none of these causes is fixed by buying a better tool. A lack of identified need is a task-selection problem. Unclear rules are a policy problem. A bad first experience is a training and expectations problem. Each has a specific fix.
| What you see | Likely cause | What to try |
|---|---|---|
| People logged in once and stopped | No clear task to use it on | Give each role two or three named tasks with a sample prompt |
| Use only by a few enthusiasts | Others are unsure what is allowed | A one-page rule sheet with examples from their own work |
| 'It just makes things up' | Early output was used unchecked or without context | Teach context-giving and a short review checklist |
| Heavy use, poor quality | Over-reliance; checking has lapsed | Spot-check samples and ask people to explain their edits |
| Managers do not use it | No visible example from leadership | Managers share one task they use AI for and how they check it |
What does a practical AI adoption strategy look like?
A practical strategy for a small or mid-size company is a sequence, not a document: pick tasks, set rules, train on those tasks, review real work and expand what holds up. The items below are an editorial plan for a team of 10 to 200 people, not a published standard.
Choose role tasks first
Ask each team lead for two or three repeated tasks where AI could produce a first draft and someone can check it. These are the adoption targets, not 'use AI more'.
Publish one page of rules
Approved tool and plan, what may and may not be entered, who reviews what, who to ask. Short enough to remember.
Train on the tasks you chose
Short practice sessions on those exact tasks beat a general course. Include a draft with planted errors to practise the check.
Make managers visible users
A manager who shows how they used AI on a real task, and what they corrected, does more than any announcement.
Review work, not logins
After four weeks, look at a few outputs per task with the reviewer. Ask what people stopped using and why.
Expand what held up
Add tasks or people only where the routine works. Retire tasks where the reviewer rewrites everything.
How do you handle the people side of AI change management?
Talk about jobs plainly and early, and involve the people who do the work in choosing the tasks. Staff worry is real: in an Acas survey of 1,023 employees in Great Britain in spring 2025, 26% named job losses as their biggest concern about AI at work, and 17% named AI making errors. Acas advises employers to consult staff and their representatives before introducing AI and to explain how it can improve roles while human involvement will still be needed.
Say what is and is not changing. If no roles are being cut as part of the rollout, say so. If some tasks will change, say which and how. Then let people raise practical concerns, such as client confidentiality or accuracy, and turn the good ones into rules. People adopt tools they helped shape far more readily than tools announced to them.
Want to see how this connects to a role?
Explore the planned team trial ↗What does fixing a stalled rollout look like?
Pellow Analytics, a fictional 60-person software company, bought business-plan seats for an AI assistant for everyone. Six weeks later, the admin console shows that most people logged in during the launch week and only a handful use it weekly, mostly in engineering and marketing. The COO's first instinct is a reminder email. Instead, she asks three team leads why.
The answers are specific. Customer success did not know whether they could paste customer emails (the rule sheet said 'no confidential data' without examples). Finance tried it once on a variance explanation, got an invented cause and gave up. Sales thought it was 'for marketing'. The fixes are equally specific. Customer success gets a rule: paste the email text with the customer's name and company replaced by placeholders. Finance gets a task with a prompt that says 'describe the change in the figures; do not state a cause unless it appears in the notes I provide' and a check: recalculate every number. Sales gets two named tasks: call-note clean-up and follow-up drafts from those notes.
Each team lead runs one 30-minute practice session on their team's tasks. Four weeks later, the COO looks at five outputs per team with the lead, not at the login chart. Two tasks are working, one needs a better prompt, and one, drafting renewal pricing emails, is dropped because the reviewer had to rewrite every draft. That is adoption: fewer, better routines that people actually use.
What are the stages of AI adoption?
There is no single agreed stage model; consultancies and vendors publish different ones. For a small or mid-size company, a useful working description is four stages: trying (individuals experiment, often on personal accounts), approved (one tool, a business plan and written rules), routine (named tasks with prompts and checks run weekly by the people who own them) and improving (the company reviews outputs, retires weak workflows and adds new ones deliberately). This is an editorial description, not a published framework.
The important transition is from approved to routine. Many companies stop at approved: they have the licences and the policy, and they wonder why little has changed. Getting to routine takes task selection, practice and review, which is the work this page describes.
Which signals should a manager watch?
Watch work quality and repeat use on named tasks, not total logins. Useful signals are: how many approved tasks are being run as intended each week, how much the reviewer corrects, whether errors are caught before they leave the team, and whether people can explain what they changed in a draft. Logins tell you about access; they say little about whether the work is better.
Be careful with self-reported gains. People who like a tool tend to overestimate the time it saves, and people who dislike it tend to underestimate. A few before-and-after samples of the same task, reviewed by the same person, are better evidence. Our answer on measuring whether AI training worked goes further.
When is pushing adoption the wrong goal?
Pushing adoption is the wrong goal when the tasks are not suited to AI, when nobody can review the output, or when the team's real problem is workload or process. A target such as 'everyone uses AI daily' rewards use for its own sake and can push people into tasks where AI adds risk without value. It is fine for some roles to use AI rarely.
It is also the wrong goal if the rules are not settled. Encouraging use before people know what they may share invites exactly the data mistakes that make leadership pull the tool.
Where does Tosheo.ai fit in AI adoption?
Tosheo.ai is being built for the practice part of adoption. At launch, a manager is planned to pick roles and the AI tools the team uses; each person takes a 5-minute diagnostic and then works through 10 to 15 minute lessons that end in a hands-on task with a real AI tool, graded against a visible rubric. The planned manager dashboard shows skill progress and completion, never raw chat logs, and the platform is planned to send nudges to inactive learners and a weekly digest to the manager.
It is pre-build, so there are no adoption results to show. It is not the right answer if your rules are unsettled or your team has no tasks suited to AI; fix those first, using the steps above. Planned prices are on the pricing page.
Sources and scope
DSIT, AI Adoption Research report (fieldwork February to May 2025), published January 2026 ↗
GOV.UK, AI Adoption Research publication page, revised February 2026 ↗
Acas, 1 in 4 workers worry that AI will lead to job losses, 28 April 2025 ↗
Acas, One third of employers think AI will increase productivity, 15 May 2025 ↗