Why AI adoption stalls, and what each step needs

Maybe the board asked you to "do something with AI", and it landed on your desk because you run IT, or operations, or simply because you seemed interested. Maybe you bought the licences and ran a lunch-and-learn, and six months later a few people love it while most use it like a search box. Maybe a pilot impressed everyone in the demo and then went quiet. Or maybe you haven't bought anything yet, people are quietly using their own AI apps, and the renewal is coming. Either way, every AI question in the building ends up with you.

If that sounds familiar, you're not behind and you're not on your own. Across UK organisations, AI use is wide but shallow, and it stalls in predictable places. The cause is rarely the technology. In Stanford's 2026 study of 51 successful AI deployments, 77% of the hardest challenges were change management, data quality and process redesign (Stanford Digital Economy Lab).

Below: each step of the climb, what holds organisations on it, and how to tell when you're ready for the next. By the end you'll be able to place your organisation area by area, and you'll know what the next step needs. To keep it concrete, we'll follow one job all the way up: winning a bid or tender, the kind of work a charity, a contractor and a consultancy all do.

The film tells the story in about seven and a half minutes. It's the version to share with your leadership team; this article is the reference.

Is it just us? Wide, but shallow

No. The UK picture, as of autumn 2026:

• Use is spreading. About four in ten UK businesses with 10 or more staff now use at least one AI technology (38%, Office for National Statistics, September 2026). Across all businesses, including the smallest, it's 30%.
• It isn't deep. When the ONS asked in June 2026, when about 35% were using AI, only one in ten of the businesses using it said they used it extensively, and 11% of businesses with 10 or more staff said more than half their workforce had received AI-related training (ONS).
• Even where the firm uses it, most staff don't. In the government's 2026 AI adoption research, 47% of firms with 50 to 249 staff used AI, including pilots, but on average only about one in five of their staff did. In firms with 250 or more staff, 52% used it and about one in six staff did. (Firms' own estimates; the survey excludes public administration, education, and health and social work.)
• Skills are the most-cited barrier, even for firms already using it. 67% of AI-using firms cite limited skills, expertise or knowledge (UK government, 2026).
• Most people haven't been trained. Only 13% of UK employees say they've had any training in using AI in their current job, and two in three of those who use it at least monthly say they've had none (CIPD Good Work Index 2026).
• Charities show the same shape. 79% of charities are at least exploring AI, but 38% use it actively or strategically, and about half have no AI policy and aren't developing one (Charity Digital Skills Report 2026, 807 self-selected respondents).

So the organisation has AI, and most of its people don't use it much yet. That's the gap the person handed "AI" is asked to close. For the full set of UK figures, see our guide How many UK businesses use AI, and how deeply?

What are the stages of AI adoption?

We describe seven steps, in a model we call the AI Stairway™ to Success. It's our own model, based on our own experience. Most published maturity models, from Microsoft's agentic AI adoption maturity model to MIT's and BCG's, share its broad shape: from scattered individual use, to shared practice, to work that's been redesigned. There's no single agreed standard; we compare several in AI maturity models: what they are, and how to use one honestly.

• Step 0. No formal AI adoption. People may experiment with personal tools, but the business has no agreed approach.
• Step 1. Approved AI access. People have suitable business accounts, guidance and tools they're allowed to use.
• Step 2. Informal professional use. People use AI for real tasks, learn its limits and find where it helps. Know-how is useful, but personal.
• Step 3. Shared best practice. Effective instructions, examples and skills are written once and available to the team.
• Step 4. Reusable agents. A defined task gains its own process, tools, checks and owner.
• Step 5. Embedded workflows. The process runs as part of normal work, with exceptions and value measured.
• Step 6. AI-first strategy. Shared systems support several workflows and give people controlled ways to use them.

The AI Stairway™ to Success model © 2026 Scorchsoft Ltd. All rights reserved. The AI Stairway™ is a trade mark of Scorchsoft Ltd. Please don't reproduce or adapt the model without permission.

It helps to think of it as a kitchen. At the bottom, people cook at home with their own kit. Then the organisation buys everyone an oven. People learn to cook, and the good recipes get written down. Then come stations, each set up for one dish; then a full service every night; and at the top, a restaurant designed around its guests. The lower steps are about one person getting good. The middle steps are about the team sharing it. The top steps are about changing the work itself.

Two cautions. Higher isn't better for everyone: a 60-person membership body may be well served with most of its work at step 3 and a couple of workflows at step 5. And the step numbers are labels, not a score out of six. We have no data on how many organisations sit on each step.

Why does AI adoption stall? It's rarely the tool

Look across the steps and the same move repeats: taking something that works for one person and making it work for everyone, without them. Each step makes the work depend less on any one person. That's why the blocker is usually people and habits, not the technology. Three patterns keep turning up.

The oven: access without capability

An oven doesn't make you a chef. Buying licences gives people access; it doesn't teach them to cook. You learn by cooking, with someone showing you. With only 13% of UK employees trained to use AI at work, many people now have an oven nobody showed them how to use.

The head chef: everything goes through one person

In many organisations, one or two people get good at AI first. Then everyone sends them their work and their questions, and they become the head chef: the only one who can cook. If it only works when she's in, the organisation isn't on that step yet. The same happens to groups: an AI working group that holds all the know-how becomes the queue. A good one does the opposite. It spreads the cooking and makes itself less needed. Microsoft's own maturity model names this "champion dependency": "Adoption stalls when champions leave or burn out" (Microsoft Learn). The pattern is well described in case material and in our own work, but we haven't seen it measured.

The absent owner: approval without backing

"Executive sponsorship is about actions, not approval," as the Stanford study puts it. In a 2026 survey of more than 1,000 UK managers, only about one in eight (13%) strongly agreed that their senior leaders actively use AI tools themselves (CMI). People copy what leaders do, not what they sign.

These three are one problem. The head chef is what happens when know-how stays in one person's head. The way out is kaizen: many small improvements by the people doing the work, written down and tried again. That takes time, and it takes permission to say what didn't work, and only leaders can give either. Approving AI doesn't give them; using it in front of people does.

Is it really mostly people?

Mostly, but not only. Stanford studied successful deployments only, and it counts data quality among those hard problems, which is partly technical. In BCG's 2026 survey of workers, a clear AI strategy went with 25 percentage points more measurable impact, against about 5 for better tools alone (BCG, consultancy research). Microsoft's 2026 Work Trend Index, which compared organisational with individual factors, put about two-thirds of the impact AI users report down to the organisation (Microsoft, vendor research). Tools still matter: from step 4 up, data, integration and the model you choose make a real difference. They're just the part you can buy.

Behaviours before processes

You can write every AI policy and KPI you like; without the behaviours underneath, they won't hold. That's the argument of Andrew Ward's book The Control Standard, and its simplest test fits AI work well. Give every piece of AI work four control points: a named owner (not "the team"), a dated next step, a decision or escalation date, and a fallback. Anything that can't clear all four is being discussed, not steered.

What does each step need?

Here's the climb, step by step, with the bid followed all the way up. Each section ends with tests you can check this week. Count a step only if you can show it working consistently, not in your best example. If you'd rather answer questions than read six sections, the AI Readiness Score looks at the same five areas in about five to seven minutes.

Step 0 to 1: from no formal adoption to approved access

What it sounds like: "We haven't got a policy yet, so we've told people not to use it." Meanwhile, the bid manager is pasting last year's tender into an AI app on her own phone. Almost a third of UK workers who use AI at work say they do it without their employer knowing (Deloitte UK, 2026, consultancy research). So your people can be further up the stairs than your organisation.

What holds it here: nobody owns the question, or the answer is a ban. A ban doesn't stop AI. It stops you seeing it. The fix is an approved route, a short set of rules people can find, and someone who owns it.

You're ready for the next step when:

• Five people in five different teams, asked "which AI tool may you use for work, and where are the rules?", give the same answer within a minute.
• You know who has an approved business account, and the guidance says plainly what must never go into an AI tool (client details, personal data, anything confidential).
• The position has a named owner and a review date, and it's "yes, here's how", not a ban with no alternative.

Your next move: ask five people which AI tool they're allowed to use, and write down the answers. Two modules of The AI Stairway™ cover what's safe to share: the introductory Know Your AI Blind Spots and Module 1, From Unsure to Useful.

Step 1 to 2: from licences to everyday use

What it sounds like: "We bought 300 licences and about 50 people use them." "We did a lunch-and-learn." "Is it saving us money yet?", asked at month three.

What holds it here: access without practice. People learn by using it on their own work, with someone beside them. Middle managers matter most here and are usually forgotten: they weren't trained either, and if they haven't given their team the time, people hear "not on my time". Only 12% of UK managers say they're very confident managing AI-enabled teams (CMI, 2026). The US evidence points the same way. In Gallup's 2026 survey of employees in organisations that have adopted AI, 78% of those who strongly agreed their manager supported their team's AI use used it frequently, against 44% of those who didn't strongly agree (Gallup).

For the bid: the bid team each try it on one section of a live tender, with their manager doing the same, and compare notes on Friday.

You're ready for the next step when:

• In most teams, most people used the approved tool on real work last week (check the activity report, or ask three people per team to show you the last thing they did with it).
• Each team's manager can name two tasks their team now does with AI as a matter of course, and practice time is in the diary, not "when you get a minute".
• People can tell you one thing it gets wrong in their own work. Knowing its limits is the sign of real use.

Your next move: book one hour of practice for every team this month, and have each manager go first.

Step 2 to 3: from informal use to shared best practice

What it sounds like: "Siobhan's brilliant with it, just send it to her." (Siobhan, like everyone in this article, is invented.) Or a 40-page prompt pack, written before anyone asked a single team what they do, that has been opened twice.

What holds it here: the know-how is personal, and the incentives push it to stay that way. If being good at AI only earns you more work, you'll keep it to yourself. Many people also hide their use: in Deloitte's 2026 UK survey, almost a third of workers who use AI at work said their employer didn't know, and experiments suggest people who use AI are judged lazier by colleagues who don't (see "What can only leaders do?"). On our reading of the figures, this is the step few organisations have taken: from "Siobhan knows how" to "the team knows how". It doesn't need a big budget. It needs time, an owner for each shared recipe, and permission to share what didn't work. Then the team improves it a little every time: that's kaizen (more below).

For the bid: Siobhan's method gets written down. The best past bids, what good looks like, and the checks she does in her head become the house recipe. Now anyone can cook it, and the bids still go out when she's on holiday.

You're ready for the next step when:

• For at least one recurring task in each team, there's a written current best way (instructions, a good example, what good looks like, the checks) that someone other than its author used this month with a comparable result.
• Each one has an owner and a review date, and it has been improved at least once since it was written.
• The last time your best user was away, that work still went out to the usual standard.

Your next move: ask your three best users to write down one worked example each, and give them the hour to do it. The AI Stairway™ covers this in Module 6, One Chat, Whole Team, which also builds the team's one-page rule for safe use.

Step 3 to 4: from shared practice to reusable agents

What it sounds like: "We built an agent, but it doesn't really work." "It's waiting for the AI group to approve it." "Nobody has time to test it."

What it is: an agent is a station. One job gets its own recipe, its own kit, its own checks, and a chef who answers for it. For the bid, that's an assistant that drafts from your real documents (past bids, policies, prices), tested on old bids before it touches a new one, with a named owner.

What holds it here:

• The wrong ingredients. In our experience, failed agents are rarely bad AI. They're unowned work. A common pattern is an agent built on a list of FAQs instead of the source documents, which is a database with a chatbot in front of it.
• Testing by exhaustion. If testing means volunteers reading every output, they tire and testing stalls. A small fixed set of test cases with expected answers, re-run after every change, is lighter and more reliable.
• Governance out of proportion. One set of rules for everything either blocks useful low-risk work or waves through risky work. Size the rules to the dish: allergen controls for the wedding cake, not a risk assessment for toast.
• Buying a ready meal. Buying an "AI agent" before you've defined the dish rarely ends well. Give any supplier the outcome, the standard, the constraints and the check-ins, or you'll pay for the supplier's days spent waiting on you.

Gartner forecast in June 2025 that over 40% of agentic AI projects will be cancelled by the end of 2027, citing costs, unclear value and weak risk controls (Gartner, analyst forecast).

You're ready for the next step when:

• Your first agent has the four control points on one page: a named owner who isn't the AI lead, a dated next step, a decision date and a fallback.
• It works from named source documents, not a summary or an FAQ list, and a fixed set of test cases was re-run after its last change.
• You can say which risk tier it sits in and who approved it, and approval took days, not a queue of meetings.

Your next move: pick the one shared recipe your team uses most, gather its source documents into one place, and name who will own its agent. This is the jump where outside help is most often worth it, and it's work we do through AI opportunity planning; The AI Stairway™ takes learners to this point in Modules 8 and 9.

Step 4 to 5: from reusable agents to embedded workflows

What it sounds like: "It saves people time, so it must be paying off." "The pilot went well. Then everyone went back to the old way."

What holds it here: AI bolted onto the old process. When factories swapped steam engines for electric motors but kept the old layout, the gains took decades; they came when the floor was redesigned around the new power (Paul David, American Economic Review, 1990, summarised by the BBC). Putting a new oven in the old corner doesn't change the kitchen. The other blocker is saved time with nowhere to go: unless someone decides what it's for, it evaporates. In BCG's 2026 survey, 42% of regular frontline AI users said it saved them at least a full day a week, yet 66% said they got limited or no guidance on what to do with that time (BCG, consultancy research).

For the bid: every bid goes through the station, the tricky ones go to a person, and you measure turnaround, rework and wins against how it ran before. In our experience, the first weeks cost more time than they save.

You're ready for the next step when:

• You have a measure of the task from before the agent (turnaround, rework or cost per finished piece) and the same measure since, taken the same way.
• It runs where the work lives, with no copying and pasting into the system of record, and exceptions go to a named person, with the exception rate visible. (Where a rule must hold every time, it belongs in code; our article AI workflows: when to use ChatGPT, builders or code explains why.)
• Someone has decided in writing what the saved time is for, and the process was redrawn around the agent, not just one step swapped for it.

Your next move: measure the task as it is now, before you build. Module 10 of The AI Stairway™, From Glue to Workflow, builds a 90-day test with a baseline and a stop rule. Our Managed AI & Automation Improvement Programme runs the same shape with you, and "no action" is one of its four possible outcomes.

Step 5 to 6: from embedded workflows to AI-first

What it sounds like: "Every team has its own AI tools now, and none of them talk to each other." "Our AI strategy is really a list of licences."

What holds it here: AI sits outside the real strategy, and incentives still reward the old way of working. Step 6 is about shared foundations (data, access, testing, logging), one pantry feeding every station, that several workflows reuse; decision rights that reflect the new work; and strategy written in outcomes rather than tool names.

For the bid: the same document store and checks now serve bids, proposals and contract renewals, and the bid plan starts by asking what AI could draft and who signs it off.

You're ready for the next step when:

• At least two workflows in different areas run on the same shared foundations rather than each team building its own.
• AI-run work is reported on quality, cost, and staff and customer experience together, and a customer who wants a person can reach one (try it yourself this week).

Your next move: list every AI tool and data connection by team, and circle the duplicates.

What does "AI-first" really mean?

AI-first is the first option you consider, not the first person you remove. When you plan a piece of work, you ask what AI could do and who carries the outcome. The work runs on shared systems that many workflows reuse. People own and check the AI's work, stay accountable, and customers can still reach a person. Quality, cost and how staff and customers experience it are measured together.

It isn't a memo. The 2025 headlines showed what goes wrong when AI-first is heard as cost-first. Klarna's chief executive said that "cost unfortunately seems to have been a too predominant evaluation factor… what you end up having is lower quality", and the company dialled back, hiring people so customers can always reach a human (CX Today, reporting Bloomberg, May 2025). Duolingo's chief executive later said of the company's AI-first memo, "I did not give enough context", and that it had never laid off any full-time employees (Fortune, August 2025). Both are large firms outside the UK, but the lesson travels: measure quality alongside cost, and say plainly what changes for people.

Few organisations are here. BCG's latest study put 7.5% of large companies in its top, "future-built" group (BCG, September 2026, consultancy research; the mapping to step 6 is ours). Climb as far as the work pays for.

What can only leaders do?

If you lead the organisation, three things can't be delegated:

1. Use it in front of people, and say what went wrong. Leaders' own use changes how they judge others. In four experiments with more than 4,000 participants, people who used AI were rated lazier and less competent, and participants playing managers who rarely used AI themselves favoured the job candidate who didn't use it (Reif, Larrick and Soll, PNAS, 2025). If you want people to share their use, show yours.
2. Give people time to learn, starting with their managers, and time to write the recipe down. Practice in the diary, and recognition for the people who share what works.
3. Name an owner for every piece of AI work, and make sure it isn't you. Then size the rules to the risk.

Leadership's job changes as you climb. On the lower steps it's permitting and modelling. In the middle it's rewarding sharing and naming owners. At the top it's funding the redesign, riding out the dip before the gains, and changing what the organisation rewards.

How do you keep climbing? Kaizen, not a programme

This is the fix for the head chef. In our experience, the organisations that climb don't usually run one big programme. They make many small improvements, made by the people doing the work, written down and tried again. That's kaizen: it's how you climb a step, and how you stay on it. Toyota's version starts with doing the job by hand: "Rather than starting with a machine from the beginning, you must first try doing it thoroughly by hand" (Toyota). Know the work before you automate it.

Ethan Mollick's "Leadership, Lab and Crowd" describes the shape that works: leaders set the direction and make it safe, a small team builds tools and benchmarks them, and everyone improves their own work, with good ideas distributed "widely, very quickly" (One Useful Thing, 2025, opinion).

Steps also decay. Champions leave, a bad incident knocks confidence, and the tools change every few months. A step you don't maintain is a step you'll slide down, which is why the improvement loop never really finishes.

This week: find the best cook in one team, and ask what it would take to write the recipe down. This month: run a ten-minute show-and-tell where people share one win, one mistake and one recipe, the routine Module 7 of The AI Stairway™, Forming an AI Culture, is built around.

Where is your organisation, area by area?

Your organisation isn't on one step. Its parts are. Microsoft's guidance on its own maturity model makes the same point: "Don't use the maturity model to create a single overall score" (Microsoft Learn). The five areas below are the five foundations our AI Readiness Score uses.

For each area, tick the highest level where you can show the evidence today, consistently. If you're not sure, that's a check to make, not a fail. Your areas will land at different levels, and that's normal: the shape tells you more than any average. The levels are a rule of thumb, not a measurement. Early covers steps 0 and 1, practising covers steps 2 and 3, and building covers steps 4 and 5. Step 6 is when building is true in most areas. To work through it with your leadership team, download the one-page self-check (PDF).

Leadership and ownership

• Early (steps 0 to 1): you can name the one person who owns AI adoption, and find the written position on which tools are allowed, dated within the last year.
• Practising (steps 2 to 3): in the last month a senior leader showed their team how they use AI, including something that didn't work.
• Building (steps 4 to 5): every live AI workflow has a named owner who isn't the AI lead, with a dated next step, a decision date and a fallback.

Process and measurement

• Early (steps 0 to 1): you can list, from asking people, the three to five recurring tasks where AI is already used.
• Practising (steps 2 to 3): at least one recurring task has a written current best way, and you recorded how long it took (or how often it needed rework) before.
• Building (steps 4 to 5): at least one workflow reports turnaround, rework and exceptions against a pre-build baseline every month, and someone has decided what the saved time is for.

Data and systems

• Early (steps 0 to 1): someone has checked who can see what in your main shared files before the approved tool is connected to them.
• Practising (steps 2 to 3): the documents your best AI-assisted task relies on are in one place, current, and owned.
• Building (steps 4 to 5): an AI workflow reads from the system of record through an approved connection, and a second workflow uses the same source.

Security and oversight

• Early (steps 0 to 1): staff can find, in under a minute, the guidance on what must never go into an AI tool, and you've asked anonymously how many use personal tools.
• Practising (steps 2 to 3): there's a risk-tiered approval route: low-risk uses need no committee, and you know how long the last approval took.
• Building (steps 4 to 5): each agent has a fixed set of test cases re-run after every change, a record of what it did, and a fallback someone has actually used.

People and delivery capacity

• Early (steps 0 to 1): you know what share of staff used the approved tool last week, and who has had any training.
• Practising (steps 2 to 3): every team has practice time in the diary this month, and your best users can say what sharing has earned them, other than more work.
• Building (steps 4 to 5): someone other than the original champion has built or changed an AI workflow, and the work kept going through the champion's last holiday.

Mark your people as well as your organisation. If most people are practising but the organisation is early, that's hidden use. If the organisation is building and most people are early, that's tools without habits. Look again in six months, and if nothing moved, ask what got in the way rather than who.

You can buy the oven. You have to grow the cooks.

To do this properly, take the AI Readiness Score: twelve questions on the same five areas, plus four about your organisation, in about five to seven minutes, free, with your report emailed at the end. It gives an overall figure too, but read the five areas first: the shape matters more than the total. Ask two colleagues to take it as well; where your answers differ is where to start.

To grow the cooks, The AI Stairway™ is our course for staff and teams: ten modules, each under an hour, from first safe habits to a workflow plan with a 90-day test. Start with its introductory module, Know Your AI Blind Spots, which is free and takes about 22 minutes.

We make both, so weigh that accordingly. "Not yet" is an honest answer too.