Something strange is happening inside organizations. Artificial intelligence is boosting productivity for individual employees, yet many organizations are struggling to translate those gains into measurable results.
McKinsey’s 2026 global survey of 1,719 respondents across 97 countries found that 80 percent say AI has improved their personal productivity. Only 37 percent could point to any impact on enterprise EBIT—a number essentially unchanged from the year before. Just 6 percent qualified as high performers.
Millions of hours are being saved, but too few of them ever reach the profit-and-loss statement. The time saved has nowhere to go, the knowledge that would make the output trustworthy was never written down and the people holding that knowledge have no reason to hand it over.
Every AI transformation is a knowledge-transfer program wearing a technology budget, and the chief talent officer owns the conditions that determine whether it works.
The talent leader’s insight
Your AI model has read the internet. It has not read your organization. The most valuable data in your enterprise has never been digitized. It’s there for your best underwriter to apply in the final 10 percent of a decision, and in what a seasoned account director knows about a client’s board politics.
A competitor can license your model by Friday. They cannot license your enterprise knowledge.
Five checkpoints deserve a hard review at every stage of the program. Skip any one of them and adoption plateaus at the enthusiasts.
1. Value: Count customers served, not pilots launched
Most boards can tell you how many AI use cases are live; far fewer can tell you which customer now waits less or which margin has moved.
In McKinsey’s 2026 survey, the share of respondents reporting any EBIT impact stayed flat, while one in five said AI-related operating costs were already constraining their use of the technology. Activity is not value. The leadership team’s job is to tie every AI initiative to a named business or customer outcome before a line of code is written, and to stop the ones that cannot pass that test.
What to do
- Rewrite the AI charter so each initiative names one financial metric and one customer metric it must move.
- Appoint a business owner, not a technology owner, for every use case above an agreed budget threshold.
- Replace the use-case count in board packs with value delivered against plan, reported every quarter.
How you’ll know it’s working
- Every live initiative reports a baseline and a target on a financial or customer measure.
- At least one initiative a quarter is stopped because it failed the value test.
- Resolution times, retention or net promoter scores improve in the customer journeys where AI is deployed.
2. Institutional knowledge: Capture the last 10 percent before it retires
The judgment that separates your top claims handler from an average one is rarely written down, and generative AI can only spread what has first been captured.
When Stanford and MIT economists studied more than 5,000 customer-support agents, in research published in the Quarterly Journal of Economics in 2025, AI assistance raised issues resolved per hour by 15 percent on average, and agents with two months’ tenure performed as well as unassisted colleagues with more than six.
The researchers found evidence that the tool was passing the habits of the most able workers to everyone else. Notably, the gains went mostly to newer and less-skilled agents; the most experienced, whose habits were being spread, gained least. That is why the conditions in checkpoint five must be in place before you ask experts to teach. Treat knowledge capture as a talent program with named experts, protected time and its own budget.
What to do
- Identify the few dozen people whose departure would hurt most, and give each of them paid time to teach, record and review.
- Pair each expert with a knowledge engineer who turns decision rules, exceptions and hard-won stories into material the model and new hires can use.
- Make a structured knowledge handover a standard step in every retirement, resignation and internal move, starting with anyone within two years of retirement.
How you’ll know it’s working
- Time to competence for new hires falls in roles where expert knowledge has been captured.
- Named experts review AI outputs in their domain at least once a month, and their corrections are logged.
- Every exit from a critical role includes a completed and signed-off knowledge handover.
3. Intent: Let the machine draft and people decide
AI is excellent at some tasks and quietly wrong at others that look just as easy, and nobody publishes a map of where the line sits.
In a Harvard Business School and BCG field experiment with 758 consultants, those using ChatGPT-4 finished more tasks, faster and to a higher standard inside the model’s capabilities, but on a task set deliberately outside them they were 19 percentage points less likely to reach the right answer than colleagues working without AI. Your underwriters and claims handlers cross that line every day, often without knowing it. CTOs must make decision rights explicit in every redesigned workflow: the system proposes, a named person decides and that person is accountable.
What to do
- Map each redesigned workflow step by step, marking where AI drafts, recommends or acts and who signs off.
- Train people to recognize when a task falls outside what the tool does well, using real cases from your own business.
- Give frontline staff clear authority to override the system and capture their reasons so the next version improves.
How you’ll know it’s working
- Override rates are tracked and reviewed monthly and sit neither at zero nor climbing unexplained.
- Error and complaint rates in AI-supported decisions match or beat the pre-AI baseline.
- Employees can name who owns the final call at each step of their workflow.
4. Roles and workflows: Build careers around the tools, not beside them
When a role changes faster than the people in it, they either grow with it or quietly work around it. The World Economic Forum’s “Future of Jobs Report 2025″ estimates that if the global workforce were 100 people, 59 would need reskilling or upskilling by 2030 and 11 of them are unlikely to receive it. Training alone will not close that gap if the career ladder still rewards the old way of working. CTOs must redesign roles, workflows, pathways and pay so that people who learn to work with AI can see exactly where it takes them next, and what it pays.
What to do
- Rewrite the job architecture for roles most exposed to AI, including new hybrid roles such as knowledge steward or AI operations lead.
- Fund hands-on practice with coaching, delivered inside real workflows rather than in a classroom, for everyone in an affected role.
- Publish internal mobility routes that show which current roles lead to which new ones, and which skills each move requires.
How you’ll know it’s working
- A rising share of new AI-related roles is filled from inside the organization.
- Regular AI use among frontline staff closes the gap with managers quarter by quarter.
- Pulse surveys show employees in affected roles can describe their next career step.
5. Culture: Make sharing knowledge the smartest career move
An expert who believes that teaching the model will cost them their job will teach it very little. Trust decides this, and it starts at the top of the organization. BCG’s “AI at Work 2025” survey of more than 10,600 employees found that the share who feel positive about generative AI rises from 15 percent to 55 percent when leadership support is strong, yet only about a quarter of frontline employees say they receive it. CTOs must make it safe to experiment, safe to admit mistakes and plainly rewarding to share what you know.
What to do
- Announce publicly how productivity gains will be used, including what happens to roles, before asking anyone to hand over their expertise.
- Add knowledge contribution and responsible AI use to performance reviews and promotion criteria.
- Ask senior leaders to show their own AI use, including where it let them down, in town halls and team meetings.
How you’ll know it’s working
- Trust and psychological safety scores rise in teams going through AI redesign, measured before, during and after the change.
- Contributions to shared knowledge bases come from every grade, not just the enthusiasts.
- Use of unapproved AI tools declines as sanctioned tools and guidance improve.
The lasting advantage
The organizations that look enviable in five years will not be the ones that moved first. They will be the ones where a new joiner reaches full competence in months rather than years, where the best underwriter’s instinct informs every decision the firm makes, and where good people stayed because the work became more interesting, not less. Their machines will have learned from their smartest people.
That outcome sits squarely within the gift of the talent function: build the capability, hone the judgment, engage the experts, retain the people who hold what matters most.
AI has handed CTOs the largest lever they have ever had over enterprise value. The organizations that recognize it will compound the advantage every year they hold it.
