AI Adoption Is Growing. Is It Actually Increasing Productivity?

Many organisations now seem past the point of asking whether AI matters.

The awkward bit is what happens after access.

The tool gets launched. People attend the demo. A few power users race ahead. Someone shares a prompt in Teams. The potential feels huge.

Then Monday arrives.

The same spreadsheet is still being updated by hand. The same document is still waiting for review. The same person is still chasing the same answer from three different systems.

That gap, between having AI and actually changing how work gets done, is where the adoption conversation gets interesting.

It is also where the productivity question becomes more useful. Are people saving a few minutes here and there, or is the organisation learning how to remove friction from the way work actually moves?

Naming that gap is not about making organisations feel behind. It is about making the picture clearer. A lot of teams are somewhere in the same messy middle: curious, experimenting, seeing pockets of value, and still working out how to make AI part of everyday work.

Adoption Is Rising. Productivity Is Less Obvious.

The numbers are useful, but they can make things sound more settled than they really are.

McKinsey's 2025 global survey says 88% of respondents report that their organisations use AI regularly in at least one business function, up from 78% the year before. At the same time, nearly two-thirds say their organisations have not yet begun scaling AI across the enterprise.

OECD data gives a more economy-wide view: across countries where data is available, 20.2% of firms reported using AI in 2025, up from 14.2% in 2024. In the UK, the government's AI Adoption Research found that around 1 in 6 businesses currently use at least one AI technology.

The Federal Reserve makes the measurement problem even clearer. In the US, one firm-level survey found about 18% of firms had adopted AI by the end of 2025, while an employment-weighted executive survey estimated that 78% of the labour force worked at firms that had adopted AI. Same economy. Very different lens.

Those figures do not have to be contradictions. They are measuring different things: access, usage, firm-level adoption, employee behaviour, and enterprise transformation. I would read them as directional rather than perfectly comparable benchmarks. A large organisation can have thousands of people using AI and still have limited workflow redesign. A small business can use one AI tool every day, get real value from it, and never call it a transformation programme.

So one better question might be: not "has this organisation adopted AI?" That question is starting to feel too blunt.

The more useful question is: where is AI changing the way work actually gets done?

Why Some Teams Move Faster Than Others

Adoption often looks fastest where the work is already digital, text-heavy, analytical, and measurable.

Technology, professional services, finance, media, marketing, legal, and customer operations all have clear starting points: document review, research, code generation, content drafting, customer support, forecasting, and knowledge search.

That does not mean those sectors are "easy". Finance and legal teams have heavy governance expectations. Healthcare has safety, privacy, and clinical accountability. Manufacturing and logistics have physical processes, complex systems, and operational risk. Public sector organisations need transparency, procurement discipline, and public trust.

The industries that look slower are not necessarily less interested. Sometimes the use cases are harder to translate into safe, repeatable workflows.

AI that summarises a meeting can be useful on day one. AI that influences a treatment pathway, a credit decision, a supply chain plan, or a public service interaction needs a different level of assurance.

That is why the next phase of adoption may be shaped less by hype and more by workflow readiness.

Big Companies and Small Teams Hit Different Problems

Large enterprises can start with real advantages: budget, data, vendor leverage, security teams, internal platforms, and enough repeatable work to justify serious investment. They are often better placed to negotiate enterprise AI tools, build internal guardrails, and run adoption programmes at scale.

But size brings complexity. Data sits in many places. Processes cross departments. Risk approval takes time. And the awkward middle between a promising prototype and a production workflow can be hard to own. The demo is often the easy part. The rollout is where ownership gets harder.

Smaller organisations can have the opposite pattern. They may move quickly because there are fewer layers. A founder, operations lead, or analyst can test a tool on Monday and change the process by Friday. The trade-off is capacity: some teams will have less access to specialist AI skills, procurement support, and time to work out where AI is genuinely worth using.

The UK research captures this well: large and mid-sized businesses are more likely to be using AI, but many adopters are investing through off-the-shelf applications or by embedding AI into tools they already use. That matters.

For many organisations, the first wave of practical AI adoption may not be custom models. It may be AI turning up inside the software people already open every morning.

What the Useful Voices Keep Saying

The voices I find most useful in this space are not simply saying "buy more AI". They are pointing to a harder but more useful idea: change the work.

Ethan Mollick, who has become one of the clearest practical voices on generative AI at work, keeps pointing to the unevenness of AI capability. In the Harvard and BCG research he worked on with Karim Lakhani and others, 758 consultants were tested on realistic consulting tasks. For tasks inside the AI frontier, performance improved substantially. But the important lesson was not "AI solves everything". It was that the frontier is jagged. AI can be excellent at some things, weak at others, and people need enough hands-on experience to know the difference.

That lines up with Andrew Ng's older but still very relevant AI Transformation Playbook. His advice is to avoid starting with a grand abstract strategy. One of his useful points is that companies often need some basic experience with AI before a thoughtful AI strategy becomes grounded. Start with projects. Build capability. Then strategy becomes more practical.

Cassie Kozyrkov, Google's first Chief Decision Scientist, frames the same problem differently. She argues that leaders often confuse tool use with automation, and that a good AI strategy starts with intent: what decision are we improving, and what does success look like? That may be a more useful starting point than "where can we sprinkle AI?"

And from the enterprise side, Microsoft and OpenAI are both saying the same thing in slightly different language. Microsoft calls it AI absorption, not just AI adoption. OpenAI's leadership guide talks about alignment, activation, sharing wins, speeding up decisions, and governing without creating roadblocks.

Different language, but a similar direction.

One way to put it: I see AI adoption less as a tool rollout and more as an organisational learning problem.

Where Adoption Actually Gets Stuck

It is tempting to think adoption is blocked by model quality.

Sometimes it is. But often the harder blockers are more ordinary. The kind of things people recognise immediately once they have tried to get anything adopted inside a real organisation.

• Unclear use cases: teams are told to "use AI" without knowing which workflow, metric, or pain point matters.
• Messy data and systems: AI struggles to answer questions or take action reliably if the underlying knowledge, permissions, and processes are fragmented.
• Security and data access: teams need clear rules for sensitive information, vendor risk, permissions, retention, and what can safely be used in external tools.
• Trust gaps: people need to understand when outputs can be used directly, when they need checking, and who is accountable.
• Skills gaps: Deloitte's 2026 enterprise AI report identifies the AI skills gap as the biggest barrier to integrating AI into existing workflows.
• Weak measurement: time saved is useful, but it may not be enough. Teams need to know whether quality, speed, risk, customer experience, or revenue actually improved.

In other words, the issue is often less "we do not have enough AI" and more "we have not connected AI to how work actually runs".

The Missing Layer: Employee Education

This is the part I think can get underestimated. AI adoption is not only a technology rollout. It is also an education challenge.

People do not just need access to tools. They need to understand what the tool is good at, where it is weak, what data it is using, how to check its output, and when human judgement should override the recommendation. Without that, AI can become either over-trusted because it sounds confident, or under-used because people do not know where it fits.

That education layer looks different for different groups:

• Executives need to understand value, risk, accountability, governance, and where AI is actually changing business outcomes.
• Managers need to know how to redesign workflows, set expectations, spot misuse, and measure whether teams are getting better results.
• Employees using AI day to day need practical training on prompts, source checking, privacy, output review, and when to ask for human escalation.
• Technical, data, and risk teams need visibility into data quality, permissions, evaluation methods, model behaviour, and monitoring over time.

The best training probably will not be a one-off webinar. It should be close to the work: example workflows, reusable playbooks, office hours, champions inside teams, lightweight checklists, and scenario-based practice. People learn faster when they can see how AI improves a task they already recognise.

In that sense, employee education is not a side activity after launch. It is part of the product. If people cannot understand, question, and safely use the system, the system has not really been adopted.

Why Enterprise Tools Help, But Do Not Fix Everything

Enterprise AI tools can matter because they make adoption safer and more scalable. They bring identity management, admin controls, auditability, data protection, connectors, permissioning, and support. They reduce the need for employees to paste sensitive information into random consumer tools. They give IT and risk teams a way to say yes with conditions instead of saying no by default.

But enterprise tools are not the whole answer. A tool can open the door. The organisation still has to walk the process through it.

Buying Copilot, Gemini, ChatGPT Enterprise, Salesforce Agentforce, ServiceNow AI agents, or any other platform rarely creates adoption on its own. It mainly creates access, which is the starting point.

The value is more likely to show up when those tools are tied to real work: a finance analyst reconciling commentary faster, a lawyer finding precedent more efficiently, a contact centre agent resolving cases with better context, a sustainability team drafting client reports from structured evidence, or an engineer using an agent to move from ticket to pull request.

OpenAI's enterprise report is interesting here because it points to intensity, not just access: workers reporting higher time savings tend to use more tools, more models, and AI across a wider range of tasks. The pattern is useful to think about. Light usage tends to give light value. Deeper value is more likely when people use AI across the messy middle of actual work.

That is one useful way to think about enterprise AI: not as a shiny new layer on top of work, but as something that can become part of how work flows.

So What Actually Moves This Forward?

The next step is probably not more excitement. There is plenty of that. The next step is choosing better starting points, with enough governance that people can move safely.

That means picking use cases where the value is visible, the risk is understood, and the team can learn quickly. A useful way to structure it is across five areas.

Start with the workflow:

• Look for work that is expensive, repetitive, knowledge-heavy, or slow because people are waiting for information.
• Choose a small safe pilot that can teach the organisation something useful in weeks, not months.
• Be clear on whether AI is helping with a better first draft, a better search experience, a better decision aid, or a faster handover.

Make the safe path obvious:

• Separate use cases that are safe for experimentation from the ones that need formal controls before they scale.
• Define what data the AI needs, who owns it, what permissions follow it, and what should never be pasted into external tools.
• Test outputs before they are trusted in front of customers, regulators, patients, or employees.

Educate people in context:

• Teach employees how to use AI against real workflows, not abstract prompt examples detached from their job.
• Show what good output, weak output, and risky output look like so people build judgement, not just enthusiasm.
• Create simple role-based guidance for executives, managers, frontline users, technical teams, and risk owners.

Give adoption an owner:

• Name who has the mandate to make adoption happen across teams, not just inside one enthusiastic pocket.
• Communicate the purpose clearly, so people understand what AI is for, what it is not for, and how it helps them.
• Ask managers to model the behaviours they want to make normal.

Measure the productivity gain:

• Decide where AI should show up as a practical improvement: faster turnaround, less rework, better decisions, or more time for higher-value work.
• Measure whether the workflow actually changed, not just whether people opened the tool.
• Share what worked, where it failed, and what the next team can reuse.

The organisations that move fastest may not be the ones with the longest AI strategy deck.

They may be the ones that build a repeatable loop: choose a starting workflow, define the guardrails, give people the tool, explain the "why", educate them in context, measure the result, improve the system, and share what worked.

That leadership mandate feels important. Without it, AI adoption can drift into optional enthusiasm. With it, the message is clearer: this is not just another tool people have to remember to use. It is a way to remove friction from work, improve quality, and give people back time.

That loop still needs governance, but not governance that lives only as a document. Useful governance answers practical questions: what can people use, with what data, for what decisions, with what level of human review? Frameworks like the NIST AI Risk Management Framework are helpful because they push organisations to think about trustworthiness, evaluation, and risk management as part of the lifecycle, not as a document at the end.

The Productivity Question Is Really a Workflow Question

I do not think AI adoption is a single moment. It is more like a maturity curve.

A useful maturity curve might look like this: access, where people can use the tools; confidence, where people know how to use them safely and well; integration, where AI is embedded into the systems and processes where work happens; and redesign, where teams stop asking only how AI can speed up the old process and start asking whether the process should exist in the same form at all.

Many organisations seem to be somewhere between access and integration. That does not have to be failure. It may simply be the normal middle of a technology shift. But it does suggest the hard work is still ahead.

The question for leaders is less whether AI will matter, and more where it will matter first.

The bigger question may be whether their organisation can turn scattered usage into shared capability, and shared capability into measurable productivity gain.

If your organisation is somewhere in the messy middle of AI adoption, it is not alone. That is probably normal. The useful question is whether the next experiment teaches the organisation how to work differently.