Most advice here starts with a failure statistic that turns out to have no research behind it. Let's start somewhere more useful: what actually separates the programmes that land from the ones that don't.
Prosci’s benchmarking of more than 2,600 change practitioners found that 88% of projects with excellent change management met or exceeded their objectives, against 13% of those with poor change management. The single biggest contributor to success, in every study they have run since 1998, is active and visible sponsorship, cited over three times more often than anything else. That tells you where AI belongs and where it doesn’t. Use it hard on stakeholder maps, impact assessments, FAQ banks and survey analysis. Use it carefully, with real rewriting, on anything that goes out under a leader’s name, because AI-generated content that looks finished and isn’t makes 42% of recipients trust the sender less. And check your account type before you paste a single employee’s name into a chatbot.
If you’ve sat through a change management deck in the last decade, you’ve seen the slide. Seventy per cent of change initiatives fail. It shows up in consultancy pitches, in business school modules, and in roughly half the AI adoption articles published this year.
Mark Hughes at the University of Brighton went looking for where it came from. He reviewed the five most-cited published instances of the claim and concluded that while a popular narrative of 70% failure clearly exists, “there is no valid and reliable empirical evidence to support such a narrative.”[1] Each source either asserted the number outright or cited someone else who had.
I bring this up first for a practical reason, not a pedantic one. If you ask ChatGPT or Claude to help you build a business case for change management, it will hand you that 70% figure inside about four seconds, with total confidence and no caveat. It’s the most repeated sentence in the corpus, so it’s the one the model reaches for. Your first real lesson in using AI on this work is that the model regresses toward whatever the internet says most often, which is not the same thing as whatever is best evidenced.
The better numbers exist. They’re just less dramatic, which is presumably why nobody puts them on a slide.
Prosci has been running the same benchmarking study for over twenty-five years, and their correlation data is the most useful thing in this field. Across more than 2,600 change practitioners, 88% of those with excellent change management programmes met or exceeded their project objectives. For good programmes it was 73%, for fair 39%, and for poor programmes 13%.[2]
Self-reported outcomes from Prosci’s Best Practices in Change Management benchmarking, more than 2,600 change practitioners. Excellent programmes are roughly seven times more likely to meet objectives than poor ones.[2]
Two more findings from the same research answer the objection you’ll hear in every steering committee. Projects with excellent change management were nearly five times more likely to be on or ahead of schedule, and around 1.5 times more likely to stay on or under budget.[2] So the “we don’t have time for the people stuff” argument runs directly against the data.
And the single greatest contributor? Prosci has asked the same open question every two years since 1998, drawing on more than 4,500 change leaders across 65 countries. Active and visible executive sponsorship wins every single time, cited over three times more frequently than the next contributor on the list. Projects with extremely effective sponsorship were almost 3.5 times more likely to meet or exceed objectives than those with very ineffective sponsorship.[3]
Hold onto that, because it’s the thing that decides where AI is allowed to touch your programme.
Two other numbers from the same body of research are worth having in your pocket. 81% of participants who planned reinforcement or sustainment activities met or exceeded objectives.[4] And middle managers were identified as the most resistant group, with 65% of participants saying their organisation had not adequately prepared managers to lead change.[3] More than half believed at least a chunk of that resistance was avoidable.
Now add the context your people are actually living in. Gartner research reported in Harvard Business Review found the average employee experienced ten planned enterprise changes in 2022, up from two in 2016, while willingness to support enterprise change collapsed from 74% to 43% over the same period.[5] Gartner also found that managers creating a psychologically safe environment saw up to a 46% reduction in change fatigue, which it defines as negative employee responses such as apathy, burnout and frustration that harm organisational outcomes.[6]
Here’s the part most articles skip. Telling someone to “use AI for change management” is useless advice, and there’s data showing why. In a July 2025 Gartner survey of 2,986 employees, 65% said they were excited to use AI at work and 77% take AI training when it’s offered. But only 42% said they knew how to identify where AI could actually improve their work.[13]
Enthusiasm isn’t the bottleneck. Naming the task is. So here are six, specifically.
Paste in your org chart, the list of teams touched, and a paragraph on what’s changing. Ask for each group’s likely gain, likely loss, what they’ll need to stop doing, and the one question they’ll ask first. Then ask the question that does the real work: which of these groups can quietly kill this without ever objecting out loud?
Departments don’t experience change. Roles do. Give it three or four real job descriptions and the new process, and ask what changes hour by hour for someone in that seat on the first Monday. You’ll get a much more concrete answer than “increased efficiency for the operations team.”
This is where tool choice matters more than prompt wording. NotebookLM, which Google is currently renaming to Gemini Notebook, is built to answer from the sources you upload rather than from general knowledge.[14] Load the actual policy, the consultation timetable and the union agreement, and answers stay anchored to documents you control. Worth being precise about the limit, though: Google’s own framing is that source grounding reduces the risk of invented answers rather than eliminating it, so spot-check the FAQ against the documents before it reaches a manager. Compared with a general chatbot answering restructure questions from nothing, it is still a materially better starting point.
Ask it to argue against your change from the position of a sceptical fifteen-year veteran of the team who has watched three similar initiatives get quietly abandoned. This is the prompt I use most in workshops, and it’s the one people find most uncomfortable, which is usually a sign it’s working.
Free-text survey responses are where genuine signal goes to die, because nobody has four hours to read 300 comments. Ask for themes with the number of comments behind each, then ask for the three most quoted verbatim comments per theme so you can check the summary against the raw material. Our guide to using AI for employee engagement surveys goes deeper on doing that without flattening what people actually said.
Useful, with one practical catch worth sorting out before you rely on it. On most platforms the AI assistant can only answer questions about a meeting afterwards if transcription was actually running during it, and that setting is often off by default or blocked by policy. Test it on a low-stakes call first. People tend to discover this the week they need it.
Notice what all six have in common. Every one produces an internal working artefact, something you and your project team read. None of them produces the thing a leader stands up and says.
Most people prompt for a comms plan like this: “Write a communication plan for our CRM migration.” What comes back is a table of channels and dates that could belong to any organisation on earth.
Prosci’s benchmarking gives you a far better brief, and it’s specific. Employees prefer to hear messages from two roles: executives and senior leaders for organisational messages, including the business reasons for the change, and immediate supervisors or people managers for messages about personal impact. Their published warning is blunt: “One of the biggest and most common mistakes you can make is to have your project team sending all the communications.”[7]
Layer on a genuinely measured finding. Tsedal Neeley and Paul Leonardi shadowed 13 managers across six companies for more than 250 hours, logging every message sent and received. One in every seven communications was completely redundant with a previous one sent through a different technology, and the managers who repeated themselves this way appeared to move their projects forward more smoothly.[8]
Leonardi is careful about how far that goes, and so should we be: they did not have data showing one group beat the other on hitting deadlines or budgets.[8] What the study does establish is the pattern, and the split by authority is the part I find most useful. Managers without formal power were redundant 21% of the time versus 12% for those with it, and they led with an instant medium such as a face-to-face conversation before following with a delayed one such as email. Managers with formal authority tended to do the reverse and spent more time on damage control.[8] Motivate first, document second.
Put those together and you get a prompt that’s actually worth writing:
Prosci also recommends that preferred senders repeat key messages five to seven times.[7] Worth saying plainly: that’s their research-informed guidance, not a measured experiment, and I’d rather flag it than let you quote it as a finding. The measured evidence for deliberate repetition is Neeley and Leonardi’s. If you want more on the mechanics of the cascade itself, we’ve covered using AI for internal communications separately.
I used to think the risk with AI in change work was factual errors. I’ve changed my mind. The bigger risk is output that looks completely finished and is hollow, because a change programme is made almost entirely of communications, and communications are exactly what this failure mode ruins.
Researchers at BetterUp Labs and Stanford’s Social Media Lab gave it a name: workslop, defined as AI-generated content that looks good but lacks the substance to move the task forward. In an ongoing survey of 1,150 US full-time employees, 40% reported receiving it in the previous month, at a cost of nearly two hours of rework per instance. Around half viewed the colleague who sent it as less creative, capable and reliable than before. 42% saw them as less trustworthy.[12]
Now hold that against the strongest finding in this whole article: active and visible sponsorship is the number one predictor of change success, cited more than three times as often as anything else.[3] Hollow AI output sent under a sponsor’s name attacks the exact variable your programme depends on most. That’s the argument, and it’s why I’d rather you used AI heavily on analysis and sparingly on anything signed.
| Artefact | Who reads it | How much AI | The rule |
|---|---|---|---|
| Stakeholder map, impact assessment | You and the project team | Heavy | Interrogate it, don’t accept the first pass |
| FAQ bank | Managers, then employees | Heavy, source-grounded only | Use a tool built to answer from your uploaded documents, then spot-check it[14] |
| Survey and feedback analysis | Steering group | Heavy | Always pull verbatim quotes to check the summary |
| Manager talking points | Line managers | Moderate | The manager must be able to say it in their own words |
| Anything a named leader signs | The whole organisation | Structure only | Rewrite every sentence, or don’t send it |
Allocation based on Prosci’s finding that sponsorship is the top success contributor[3] and the BetterUp/Stanford finding that hollow AI output reduces trust in the sender.[12]
There’s a second problem, which is that you will be a bad judge of whether any of this is helping. METR ran a randomised trial with 16 experienced developers on 246 real tasks. They predicted AI would make them 24% faster. They were measurably 19% slower. Afterwards, having just lived through it, they still estimated they’d been 20% faster.[11]
To be fair to METR, they’ve since said that specific slowdown result is out of date and that 2026 tools probably do speed developers up.[11] Nobody has retracted the perception gap, though, and that’s the part that matters for you. Self-assessment of AI benefit is unreliable, which is an odd thing to admit if you run change programmes for a living, because measuring adoption properly is literally your professional skill. Point it at yourself.
The realistic size of the prize also deserves stating. Danish researchers surveying around 25,000 workers across 7,000 workplaces found that among people actually using the tools, average time savings came to 2.8% of work hours, against the much larger gains controlled experiments report.[9] A US survey by economists at the Federal Reserve Bank of St. Louis put self-reported savings at 5.4% of work hours in its November 2024 wave, roughly 2.2 hours in a 40-hour week, and noted openly that if people complete tasks faster without their employer knowing, that time becomes on-the-job leisure rather than productivity.[10] That page now carries a note that the 2024 results were later revised, which is worth knowing before you quote the figure at anyone.
Change work runs on the most sensitive data your organisation holds. Names against roles. Who’s at risk. Who said what in a confidential consultation. And the tool people reach for is whatever they already have open.
The single most useful guardrail I’ve found is in Microsoft’s own documentation, and almost nobody in this space quotes it. Microsoft defines “workplace harms” as generative AI making inferences, judgments or evaluations about an employee based on their workplace communication, specifically about performance, attitude, internal or emotional state, or personal characteristics. Its position: “We restrict the use of generative AI or models from being used for these purposes.”[15]
Read that alongside the thing a manager planning a restructure is most tempted to type, which is some version of “based on their messages, who on my team seems disengaged.” Microsoft’s product is documented as restricted from answering it. That’s a better answer than any policy memo.
Three specifics worth knowing, all from the vendors themselves:
On the process side, the UK’s Information Commissioner’s Office is clear that a Data Protection Impact Assessment should be carried out before using an AI tool, ideally at the procurement stage, and that you must identify a lawful basis and who is controller versus processor.[18] That guidance was issued about recruitment AI rather than change management, so apply the principle rather than claiming the scope, but the sequencing point holds: the assessment belongs before the tool, not after the incident. We’ve written more on the day-to-day version of this in using AI without leaking company data.
Abstract advice is easy to nod at and hard to use. So here’s the shape of it, using the most common change I get asked about in corporate sessions, which is a new system landing on a team that didn’t ask for it.
Sequenced from Prosci’s preferred-sender findings[7], the measured instant-then-delayed pattern[8], and the reinforcement correlation.[4]
A few things about that sequence are deliberate.
Managers get briefed in week three, before the organisation hears anything in week four. That ordering exists because 65% of practitioners said their organisations didn’t adequately prepare managers to lead change, and middle managers came out as the most resistant group.[3] A manager who finds out at the same time as their team has no useful answer to the only question that matters, which is “what does this mean for me.”
Week six isn’t a wrap-up. Reinforcement correlates with 81% of participants meeting or exceeding objectives, and it’s the step that gets cut when the go-live date slips.[4] If you’re going to protect one week from the schedule squeeze, protect that one.
And if the change you’re running is itself an AI rollout, the Danish research has a neat finding for you: employer encouragement, an enterprise tool and training together roughly doubled adoption rates, and training measurably narrowed the gender gap in take-up.[9] AI adoption is a change programme, and the same levers work on it. We’ve written that one up properly in how to train your team on AI.
Four failure patterns, in the order I see them.
The plan looks finished on day one. AI is very good at producing a document with the right headings. A change plan with all the right headings and no real stakeholder knowledge in it is worse than no plan, because it stops anyone asking the questions that would have surfaced the problem. If your first draft came back complete and you didn’t argue with it once, you haven’t got a plan, you’ve got a template.
Nobody tells people what to do with the time. Gartner found that only 7% of organisations provide guidelines to employees on how to use time saved by AI, from a July 2025 survey of 114 HR leaders.[13] If your change programme promises time savings and never says what the time is for, people will assume the answer is redundancy, and they’ll be resistant for a perfectly rational reason.
The tool becomes the programme. Gartner reported in October 2025 that 88% of HR leaders said their organisations had not realised significant business value from AI tools.[13] Some of that is genuinely early days. Some of it is teams who bought a platform instead of changing a behaviour, which is the oldest failure in this discipline wearing new clothes.
Sponsorship gets delegated to a mailbox. This one predates AI by about forty years, but AI makes it much easier to do at scale. A sponsor who has never spoken the words out loud, whose entire visible presence in the change is a set of drafted emails, is not active and visible sponsorship. The research is unusually consistent on this being the thing that matters most.[3]
If you take one thing from this: decide, before you open a single chat window, which artefacts in your programme have a named human’s credibility attached to them. Those get your time. Everything else gets the machine’s.
It can draft the structure, and it should not write the final text of anything a named leader sends. Prosci’s benchmarking identifies active and visible sponsorship as the single greatest contributor to change success, cited over three times more often than any other factor. Research from BetterUp Labs and Stanford found that 42% of people who received hollow AI-generated work saw the sender as less trustworthy. Use AI on the analysis, the FAQ bank and the manager talking points, then write the sponsor’s words yourself.
There is no reliable evidence for it. Mark Hughes at the University of Brighton reviewed the five most-cited published sources of the 70% claim and concluded there is no valid and reliable empirical evidence supporting it. Use Prosci’s correlation data instead, which measures something real: 88% of projects with excellent change management met or exceeded objectives, against 13% with poor change management, across more than 2,600 practitioners.
It depends on the artefact rather than the brand. For a FAQ bank or anything that must not invent policy, use a source-grounded tool such as NotebookLM, now being renamed Gemini Notebook, which is built to answer from documents you upload rather than from general knowledge. Google’s own framing is that this reduces rather than removes the risk of a wrong answer, so spot-check it. For stakeholder mapping and pre-mortems, any capable general chatbot works. For meeting capture on governance calls, check whether transcription needs to be running for the assistant to be promptable afterwards, because on most platforms it does.
Not into a personal consumer account. Business and enterprise tiers generally do not train on your inputs by default, while consumer tiers behave differently and your employees are usually on those. Anthropic extended retention to five years for consumer users who allowed training data use, and Google tells consumer Gemini users plainly not to enter confidential information they would not want a reviewer to see. The UK ICO’s position is that a Data Protection Impact Assessment belongs at the procurement stage, before the tool is used.
Less than the marketing suggests, and you will be a poor judge of it. Danish research covering around 25,000 workers found average savings of 2.8% of work hours, and a Federal Reserve Bank of St. Louis survey put self-reported savings at 5.4%, roughly 2.2 hours a week. A METR randomised trial found experienced professionals were measurably slower with AI while believing they had been 20% faster. Measure adoption and outcomes the way you would for any other change, rather than trusting the feeling of speed.
This guide draws on Prosci’s Best Practices in Change Management benchmarking, peer-reviewed work on the origins of the 70% failure claim, Gartner survey data on change fatigue and AI value in HR, measured studies of AI effects on knowledge work from the Federal Reserve Bank of St. Louis, METR and researchers at the University of Chicago and University of Copenhagen, and vendor documentation from Microsoft, Google and Anthropic. Two things worth flagging: the Prosci correlation figures are that company’s own self-reported practitioner research, which sits in interesting tension with the peer-reviewed scepticism about the 70% claim, and several widely-circulated change-fatigue statistics were deliberately excluded because they turned out to be third-party claims repeated on consultancy pages rather than original research.
Weekly tips. Real examples. Practical help for busy professionals.
We care about your data, check out our privacy policy.