Nearly every company has tried AI by now. Far fewer can show what it changed. This guide is about closing that gap. It uses one example throughout: a regional insurance broker with 400 staff, going from a list of ideas to a claims process that runs faster with people still in charge.

The gap

In McKinsey's 2025 State of AI survey, 88% of respondents said their organisation used AI regularly in at least one business function, but only 39% reported any effect on EBIT at the enterprise level. These are self-reports, but the pattern is consistent: use is wide, and measurable value is narrow.

A 2025 MIT report went further. From a review of more than 300 public AI initiatives, 52 interviews and 153 survey responses, it concluded that "just 5% of integrated AI pilots are extracting millions in value, while the vast majority remain stuck with no measurable P&L impact". The authors describe the findings as preliminary and directional, and critics point to the small sample and the short six-month window. The direction matches what we see, though: pilots stall when they aren't tied to real work, measured and built into how people do the job.

Start with the work

Don't start with the technology. Start with where the time goes. Run short workshops with each team and collect:

• tasks that repeat, with the same steps each time;
• documents that get read and re-keyed into another system;
• questions that get answered again and again;
• drafts that follow a pattern: letters, summaries, replies.

Score each idea on two axes. Value is the hours saved multiplied by how often the task happens. Feasibility covers whether the data is ready, how costly a mistake would be and how much effort it takes. Pilot the ideas that score high on both, plan for the valuable but hard ones, and park the rest.

Redesign the workflow

Adding a chatbot to an unchanged process rarely moves a number. The gains come from redesigning the workflow around what AI does well, with a person checking what matters.

In the broker's claims intake, a handler used to read each email and its attachments, re-key the details, check the policy, route the claim and write an acknowledgement: about 25 minutes. Redesigned, AI reads and extracts the details, suggests the policy check and drafts the acknowledgement. The handler checks, confirms and routes it in about 6 minutes, and still makes every decision.

Run an honest pilot

A pilot is a working prototype, used by real people on real work, measured against a baseline taken before it starts. Track three things every week: how many people use it, how much time it saves on the task, and how often its output is accepted without changes.

Measure the work, not the impression of it. In METR's 2025 study of experienced open-source developers, those allowed to use AI tools took 19% longer to complete their tasks. Beforehand they had expected a 24% speed-up, and afterwards they still believed they had sped up by 20%. METR has since said those early-2025 results no longer reflect current tools, but the lesson about self-reports stands.

The evidence for gains is strongest where the work is well defined. In a study of 5,172 customer-support agents, published in the Quarterly Journal of Economics in 2025, access to an AI assistant raised the number of issues resolved per hour by about 15% on average, with the biggest gains among newer and less experienced agents.

Gate quality before scaling

Before a pilot scales, build a test set from real cases with known good answers, and score every option against it: accuracy, critical errors, cost and speed. Set a gate in advance, for example "at least 93% accurate and no urgent claim routed wrongly", and let only options that clear it go forward.

The cheapest model is often good enough, and sometimes it isn't. The test set tells you which, and it tells you again every time a model or prompt changes.

Governance that helps

People use AI either openly or quietly. Clear rules make it open. A workable set is short:

• a usage policy that fits on a page;
• a data classification, from public to restricted, with the tools allowed for each class;
• a list of approved tools, and a quick way to request a new one;
• a register of AI uses, with an owner for each.

In the EU, the AI Act's AI-literacy duty has applied since 2 February 2025. As amended in 2026, it asks organisations that provide or deploy AI to support the AI literacy of their staff, softened from the original wording, which asked them to ensure it. For a recognised structure, ISO/IEC 42001 sets out a certifiable AI management system, and the NIST AI Risk Management Framework, with its profile for generative AI, maps the risks to manage.

Train for adoption

Rollouts succeed team by team. Train leaders first, so they use it and ask for it; then a network of champions, roughly one for every 20 to 25 people, who help colleagues day to day; then each team, with examples from its own work. Track weekly active use by team, and look closely where it stalls: usually the workflow hasn't changed, or people don't trust the output yet.

A short checklist

• Ideas collected from the work, and scored for value and feasibility.
• Pilots only in the high-value, high-feasibility corner.
• The workflow redesigned, with a person checking what matters.
• A baseline measured before every pilot.
• Weekly numbers for use, time saved and acceptance.
• A test set and a quality gate agreed before scaling.
• A one-page policy, a data classification and approved tools.
• Leaders, champions, then teams, with active use tracked by team.

Common questions

How many pilots should we run?

Two or three at a time. Enough to learn what works in your organisation, few enough to measure each one properly and support the people using them.

Should we build or buy AI tools?

Buy for general work: an enterprise assistant covers drafting, summarising and research. Build where the workflow is specific to you, such as claims intake, and connect it to your systems.

How do we know a pilot worked?

Compare it against the baseline you measured before it started, on the same task. The pilot worked if the time saved and the quality hold up after the first weeks, when the novelty has worn off.

---

Sources, as of October 2026: McKinsey: The state of AI (opens in a new tab), MIT NANDA: The GenAI Divide, State of AI in Business 2025 (opens in a new tab), METR: early-2025 AI and experienced open-source developer productivity (opens in a new tab), METR: uplift update (opens in a new tab), Brynjolfsson, Li and Raymond: Generative AI at Work (opens in a new tab), EU AI Act, Regulation (EU) 2024/1689 (opens in a new tab), Gibson Dunn: the AI Act omnibus agreement (opens in a new tab), ISO/IEC 42001 (opens in a new tab), NIST AI Risk Management Framework (opens in a new tab).