Adoption is the ROI.

The tool was never the hard part.

You have seen this rollout. A tool gets chosen after a strong demo. IT connects it, someone sends the launch message, there is a training call and the recording gets shared. Three months later, four people use it, two of them badly, and everyone else works exactly the way they did before. The subscription renews anyway.

The scale of this pattern is documented, noisily and quietly. The noisy version: MIT's Project NANDA reviewed more than 300 enterprise AI initiatives in 2025 and reported that, despite an estimated $30–40 billion in enterprise spending on generative AI, 95% of organisations were getting zero return. That study was preliminary and its method drew fair criticism, so hold it loosely. The quiet version is harder to argue with: McKinsey surveyed 3,613 employees and 238 executives and found that while almost every company now invests in AI, only 1% of leaders call their company mature, meaning the tools are actually embedded in workflows and driving outcomes.

Here is what makes those numbers interesting rather than just grim: the technology is not the bottleneck, and neither is willingness. In the same MIT research, workers at more than 90% of the companies surveyed reported regularly using personal AI tools for work, often while the official rollout stalled. McKinsey found employees were three times more likely to be running gen AI on at least 30% of their daily work than their leaders assumed. People are not refusing AI. They are refusing your rollout.

Harvard's AI Essentials for Leaders programme calls the underlying problem the capability-adoption gap: model capability improves at an astonishing pace, while an organisation's ability to absorb it moves at the speed of habits, trust and training. That gap, not the model, is where the budget dies. Closing it is a change-management job, and it starts with a blunt accounting question: where does the value of an AI change actually come from?

Adoption is the ROI.

The invoice buys capability. Behaviour produces the return. Everything between those two sentences is what an AI rollout actually is.

BCG's guidance for AI transformation puts proportions on it: 10% of the effort is algorithms, 20% is technology and data, and 70%, the lion's share, is people and process. The tool is the cheap, fast, exchangeable part. The slow, expensive, differentiating part is whether your team changes how the work gets done.

So define adoption honestly. It is not licences activated, logins per week, or a completed workshop. Adoption is a changed workflow: the account brief that took twenty-five minutes takes eight, everyone produces it the new way, and nobody had to be reminded on a random Tuesday. The deployments that succeed tend to look exactly like that definition. MIT's research found the rare winners bought purpose-built tools for one specific workflow and embedded them deeply, rather than running broad "AI enablement" programmes.

That also settles the strategy question. Your competitor can buy the same model tomorrow. The vendor will happily sell it to them. What they cannot buy is a team that actually works differently. The 70% is the moat.

Two beliefs decide everything.

When someone on your team is not using the tool, they are usually not being difficult. They are answering no, often rationally, to one of two questions.

The first is about buy-in: do I believe this is worth it? Worth it to me, in my job, this quarter, not to the company in a slide. The second is about skill: do I believe I can do this? Without producing embarrassing output, without looking slow in front of colleagues, without breaking something a customer sees.

This is the Hearts and Minds framework from Harvard Business School Online's AI Essentials for Leaders. Of everything in that programme, it is the piece we use most in practice. Plot buy-in against skill and your team lands in four mindsets: Inspired (high on both), Frustrated (sold on the why, stuck on the how), Indifferent (capable but not convinced), and Overwhelmed (low on both).

The map is only useful at the level of the person. A team average hides that your best salesperson is Indifferent and your newest hire is Overwhelmed, and the same launch email does nothing for either of them. Diagnosis first, then the move.

Four mindsets, four moves.

Each mindset has a signal you can observe and one move that works. Running the wrong move is how rollouts burn goodwill: training the Indifferent insults them, and messaging campaigns bounce off the Frustrated, who are already sold.

1. 01 — Inspired → give them room to lead. The signal: your visible early adopters. Already getting value, sharing prompts nobody asked them to share. The move: give them a stage, not a task force: a ten-minute slot in the team meeting, an open invitation to show their workflow. Peer proof travels further than any mandate. One caution: do not quietly convert them into unpaid support. Give them permission and visibility; route the tickets elsewhere.
2. 02 — Frustrated → train them, hands-on. The signal: they want in but cannot yet. They tried twice, got a mediocre output, and went back to the old way without telling anyone. The move: role-specific, hands-on training on their own live work. It is the single most fixable gap on the map, and the one generic training misses entirely. The demand is real: in McKinsey's research, nearly half of employees said they want more formal training than they are getting. Skill demand outstrips teaching supply in most rollouts.
3. 03 — Indifferent → build the trust. The signal: the experienced, skeptical-but-able colleague. Could use the tool tomorrow; sees no reason to. Often your best performer, whose current way of working is exactly what made them good. The move: build trust, not skill. Clear messaging on the why, real examples from peers rather than vendor case studies, a straight line to goals they already own, and transparency about what the tool does with their data and who is accountable when it is wrong. Trust is the real currency here: DORA's 2025 survey of nearly 5,000 technology professionals found 90% now use AI at work, yet three in ten say they trust AI-generated code only a little or not at all. Usage without trust is fragile. The first bad output confirms the skeptic.
4. 04 — Overwhelmed → start small. The signal: silence. This group does not complain; it avoids. The move: begin with the lowest-stakes task that takes a job they dislike off their plate. Not the core workflow, not a dashboard: one repetitive chore, done for them once, then with them. Patience compounds and pressure does not. Two rules: never anchor the rollout timeline to this group, and never lead the rollout with them.

What actually moves people.

Whatever your team's mix of mindsets, four moves do most of the work. None of them costs much. All of them are habits, which is why they get skipped.

Use it visibly yourself. McKinsey's conclusion was blunt: the biggest barrier to scaling AI is not employees, who are ready. It is leaders, who are not steering fast enough. A leader who delegates AI while never touching it signals that this is optional theatre. Open the Monday pipeline review with the tool's output on screen and the signal reverses.

Seed one credible early adopter per team. Credible is the operative word: not the most enthusiastic junior, but the person whose judgment the team already trusts. One trusted voice per team, and momentum compounds on its own.

Train on the real workflow. Short sessions, this week's accounts, the actual job. Abstract "introduction to AI" sessions produce certificates, not behaviour.

Let people help shape it. People adopt what they helped build. The mechanism is a feedback loop that visibly changes the workflow within days. A suggestion that disappears into a backlog teaches people to stop suggesting.

AI is an amplifier: it magnifies how a team already works. Cohesive teams get faster. Fragmented teams get louder.

Scale only when it's earned.

After a good pilot, the instinct is to roll the tool out everywhere at once. The discipline is to hold.

The benchmark we borrow from Harvard's programme: keep working the current team, refining the workflow and building the habits, until adoption holds at 75–85% and the return is measurable. Then, and only then, standardise: document the new workflow as the default, wire it into onboarding, and make the old way the exception that needs a reason.

Both failure modes are expensive. Scale too early and you multiply a workflow nobody follows. Worse, the second team inherits the first team's doubts along with the tool. Wait too long and the pilot decays quietly as the champions' attention moves on. The 75–85% band, held stable over weeks rather than spiking after a launch push, is the tell that the behaviour is real.

The reason to be patient is strategic, not procedural. The lasting edge was never the model. Anyone with a credit card has the model. The edge is the workflows, habits and data discipline your team builds around it. That is the part a competitor cannot invoice.

1. BCG · The Leader's Guide to Transforming with AI (2024, updated 2025). The 10-20-70 rule.
2. MIT Project NANDA · The GenAI Divide: State of AI in Business 2025 (2025). 95% zero return; 300+ initiatives; shadow AI usage. Released as preliminary, non-peer-reviewed findings; cited here with that caveat.
3. McKinsey · Superagency in the Workplace (January 2025; 3,613 employees, 238 C-level executives). 1% maturity; employees vs. leadership readiness; training demand.
4. Google Cloud DORA · State of AI-assisted Software Development 2025 (September 2025; ~5,000 technology professionals). 90% adoption; trust distribution; "mirror and multiplier."
5. Harvard Business School Online · AI Essentials for Leaders. Hearts and Minds framework; capability-adoption gap; 75–85% adoption benchmark.