AI has taken the shallow work. What remains for leaders is deep thinking, synthesis and judgement. Three disciplines for protecting deep work, in order.
A board pack ran to forty-seven slides. Someone clicked through it. Several people nodded. Nobody in that room sat and thought.
The decision that mattered was not made in the meeting. It was made later that evening, at home, with the screen closed and thirty minutes of quiet.
I spent years inside senior teams and learned one thing about this. The shallow work of a company is always loud and always urgent. Meetings, decks, inbox, fast decisions. The decisions that actually shaped the business were rarely made in the middle of that noise. They were made at the edge of it. A quiet hour. Half an hour with the screen off. A pen and a sheet of paper.
Most of that shallow work is now moving to AI. Drafting the deck, writing the email, summarising the report, gathering the data. One by one these are leaving our hands.
What remains is the part nobody can delegate.
AI is absorbing the repeatable layer first: collecting, summarising, drafting and formatting. These tasks share one property. They have a defined output and a checkable result, which is precisely what a model handles well.
The mistake is assuming this hands leaders more thinking time by default. It does not.
Workday's Beyond Productivity study, fielded by Hanover Research in November 2025 across 3,200 respondents in North America, APAC and EMEA, found that close to forty per cent of the time AI saves goes straight back into rework: correcting errors, rewriting output and verifying results. Only fourteen per cent of employees report consistently net-positive outcomes.
So the hour AI gives back is not automatically an hour of thinking. Without a verification habit it becomes an hour of checking.
That distinction is the whole leadership question.
Deep work is not a new capability. Cal Newport named it in 2016 and the underlying idea is older than that. What changed is its scarcity value.
When drafting a strategy document took three days, the ability to draft was itself worth something. Now it takes twenty minutes. The scarce input is no longer production. It is judgement: deciding which question is worth asking, which assumption is load-bearing, and which output is wrong in a way that matters.
MIT Media Lab's 2025 GenAI Divide study found that ninety-five per cent of corporate AI pilots produced no measurable P&L impact. The reason it identified was not model quality. It was the learning gap and the absence of proper enterprise integration. Put plainly: the organisations had output and lacked judgement about that output.
Deloitte's State of AI in the Enterprise 2026 puts a number on the same gap from another angle. Only twenty per cent of organisations say their talent is highly prepared for broad AI adoption, and talent scores lowest of the five readiness dimensions measured.
The tools arrived. The thinking did not scale with them.
Here is the uncomfortable part. In most organisations the biggest consumer of everyone else's attention is the leadership layer itself.
The urgent request at six in the evening, the meeting that could have been a paragraph, the document circulated for comment with no decision attached. Each of these is a withdrawal from someone's concentration, and concentration is the input that AI cannot supply.
A leader who protects their own deep work while fragmenting everyone else's has not solved the problem. They have relocated it.
This matters more now than it did three years ago, because the work that remains for humans is precisely the work that requires uninterrupted attention. If the organisation's operating rhythm makes that impossible, the AI investment underdelivers regardless of the tooling.
Protect the block before you protect the calendar. Most leaders try to find thinking time in the gaps between meetings. Gaps do not produce deep work; they produce recovery. The block has to be scheduled first and defended like an external commitment. Ninety minutes, twice a week, is a realistic floor for most senior roles.
Separate production from judgement. When AI drafts something, the review should not happen immediately afterwards in the same sitting. The mind that just produced is poorly placed to evaluate. A gap of even a few hours materially improves the quality of the critique, and the critique is where the value now sits.
Make the verification standard explicit. Teams need to know what "checked" means. Which claims require a source. Which numbers require a second pair of eyes. Which outputs can ship without review. Without that standard, the forty per cent rework figure becomes your organisation's default.
The talent implication is direct. If judgement is the scarce input, then judgement has to be developed deliberately rather than assumed.
That changes what learning and development budgets should buy. Tool training has a short half-life; the interfaces are getting easier every quarter and much of that training expires. Critical evaluation, structured questioning and clear written reasoning hold their value across tool generations.
It also changes hiring signals. Fluency with a particular AI product tells you less than the ability to explain why a plausible-looking output is wrong. The second is harder to teach and far harder to fake.
For UK organisations preparing 2027 workforce plans, this is the practical question worth putting on the table: are we buying capability that expires with the next model release, or capability that compounds?
Leadership used to include a substantial production component. Writing the strategy, building the case, preparing the board material. That component is shrinking and will keep shrinking.
What replaces it is not less work. It is different work: framing the question, holding the standard, deciding what is true, and protecting the conditions under which other people can think.
That is a harder job than it sounds, and it is the one that is left.
What to do this week:
1. Put two ninety-minute thinking blocks in next week's calendar and treat them as external commitments.
2. Write down what "checked" means for one type of AI-assisted output your team produces regularly.
3. Review one line of your L&D budget against a single test: does this capability survive the next model release?
→ Enquire about an AI and leadership speaker for your event
---
Jessica Chen · Speaker Reel (YouTube)
The term was popularised by Cal Newport in 2016, but the underlying practice is much older. What has changed is its economic value. When production was slow, the ability to produce was scarce. Now that AI produces quickly, judgement about that production is the scarce input, and judgement requires uninterrupted attention.
Most often the verification layer has not been defined. Workday's 2026 research found close to forty per cent of AI time savings going back into rework, with only fourteen per cent of employees reporting consistent net gains. Agree explicitly what must be checked, by whom and against what source. Inconsistency usually traces back to an unwritten standard rather than to the tool.
Both, weighted differently over time. Tool training is necessary and has a short half-life because interfaces keep simplifying. Critical evaluation, structured questioning and clear written reasoning hold value across tool generations. A practical split is to treat tool training as recurring operational cost and reasoning capability as the longer-term investment.
Start with the leadership layer rather than a company-wide policy. Leaders generate a disproportionate share of the interruptions. Two protected blocks a week at senior level, plus a rule that non-urgent requests wait for the next working day, changes the rhythm without adding process.
Roles with a large production component and a small judgement component change fastest: analysis, drafting, reporting and first-pass research. Roles where the output is a decision rather than a document change more slowly, but the inputs to those decisions change immediately. Both need attention in workforce planning.
After Dinner Speakers
Contents
