Ten shifts shaping artificial intelligence and the very human business of using it well.

Imagine arriving at work to discover that your new assistant has prepared a report, reorganized your calendar and drafted twelve emails before you have finished your coffee. Wonderful. Then you notice that the report uses last year’s prices, the calendar has lost your lunch break, and one email promises something your company does not sell. 

This is an imaginary morning, but it captures a real challenge: useful AI requires more than impressive capability. It requires instructions, context, boundaries and someone who can recognize when helpfulness has taken a wrong turn.

Reading Bernard Marr’s reflections on World Summit AI 2026 in Amsterdam, I was struck by how much of the discussion concerned these practical questions. His seven themes span autonomous agents, business redesign, governance, sovereignty, physical AI, infrastructure and human wisdom. The summit’s published program also puts enterprise adoption, human–AI collaboration and trustworthy systems firmly on the agenda. [1]

Those themes deserve attention. But I would take the conversation further: how must our working habits change as AI becomes more capable? Here are ten shifts worth understanding.

1. AI can take action. We need to learn to delegate.

An AI assistant can help write a reply. An AI agent, connected to the right tools and permissions, may also retrieve an order, check a policy and initiate the next step. The distinction matters because an incorrect sentence and an incorrect transaction have different consequences.

Consider a hypothetical travel assistant. Comparing three hotels is one responsibility. Booking a non-refundable room is another. Cancelling your existing reservation because it found something “better” is an excellent way to make itself unpopular. Delegating to AI therefore requires a clear brief: the goal, the permitted information, the spending limit, the actions requiring approval and the point at which it should stop and ask. 

This is becoming an engineering concern as well as a management skill. NIST’s 2026 AI Agent Standards Initiative explicitly addresses agent security, identity and interoperability. [2] For users, the habit to develop is precise delegation. “Organize my trip” is an ambition. “Compare these options and ask before purchasing” is an instruction someone can safely act on.

2. Faster tasks will expose slower organizations.

Imagine a marketing team that uses AI to produce twenty campaign drafts in the time it once took to produce two. All twenty then join the same queue for approval by the same overloaded manager. The writing became faster. The campaign did not.

This is the practical force of Marr’s point about redesigning work. Improving one step may simply move the delay somewhere else. Before adding AI to a process, follow the work from beginning to end. Where does it wait? Where is information entered twice? Which decisions need expert judgment, and which merely need a clear rule? Sometimes the useful change is surprisingly modest: a complete brief at the start, fewer handovers, or a shared source of current information. AI can help redesign the work. It should also prompt an awkward question: why are we doing this particular task at all?

3. Trust needs a working method.

“A human will check it” sounds reassuring until that human has six minutes, forty documents and no access to the original evidence. Meaningful oversight needs time, authority and something concrete to inspect. For an AI-supported customer service process, that might mean current policy documents, restricted access to customer records, a record of actions and a clear route for exceptions. A disputed refund should reach someone who can resolve it.

Governance becomes useful when it appears in the work itself: permissions, testing, review points and recovery procedures. NIST’s agent initiative reflects this focus on making systems work securely across real environments. [2] For the person using AI, the corresponding habit is verification. Ask which source supports a claim. Check that the source actually says it. Treat a confident answer as an answer to examine.

An elegant explanation of the wrong policy is still the wrong policy.

4. AI sovereignty reaches all the way to your desk.

Sovereignty sounds like something discussed by ministers in rooms with flags. Its everyday version is simpler: how much control do you have over a tool on which you depend? Where is information processed? Who can access it? Can the organization move its work elsewhere? What happens if a service becomes unavailable or its terms change? The summit’s dedicated Sovereign AI Forum explicitly connects sovereignty with control over infrastructure, data and rules. [1]

For a small business, the practical response may be a careful supplier choice and an exit plan. For a public institution, it may require much more extensive control. For individuals, the first useful step is knowing which tools are approved for which information. The fact that an assistant accepts a document does not tell you whether uploading it is appropriate. Convenience is wonderful. Understanding the dependency behind it is better preparation for a difficult Tuesday.

5. AI is learning to work beyond the screen.

Robots make excellent conference guests. They attract attention and rarely ask the speaker a question disguised as a ten-minute speech. Their harder assignment is ordinary life: an unfamiliar object, a crowded workspace, a person moving unexpectedly nearby.

Advances in AI models that combine visual information, language and action are expanding what robots can attempt. Google DeepMind’s published robotics work, for example, describes systems that connect visual understanding and instructions with physical actions and multi-step plans. These are reported research capabilities, not proof of dependable performance in every workplace. [3]

For people, the change extends beyond learning to issue commands. Workplaces will need to decide how people and machines share space, how exceptions are handled and who can stop an operation. A helpful way to evaluate physical AI is to look past the impressive demonstration. Ask how consistently it works, under what conditions, and what happens when those conditions change.

6. Every clever answer has a running cost.

AI can feel almost weightless. You type a question; an answer appears. Somewhere else, machines are doing the work. Training creates or updates a model. Inference is the computation involved in using it. An agent that repeatedly searches, checks, plans and retries may require many model calls to finish one assignment.

The electricity behind these systems is substantial. In its 2025 base-case projection, the International Energy Agency estimated that global data-centre electricity consumption would reach about 945 terawatt-hours in 2030, more than double its 2024 level. That figure covers data centres overall, not AI alone, and is a projection rather than a measured future result. [4] 

The practical lesson is proportion. A routine classification task and a difficult research problem need not receive identical computing resources. Businesses should measure the cost of an acceptable completed result, including checking and correction. A cheap answer that creates an expensive mistake is poor value.

7. Human judgment needs practice, too.

Marr’s seventh theme, the value of human wisdom, may be the most uncomfortable, because it gives us work to do. AI can suggest several ways to reduce costs, but people still have to decide what those savings would mean for customers, employees and the business they want to build. 

Yet human involvement alone guarantees very little. People overlook evidence, defend familiar habits and approve things they have barely read. A person clicking “accept” at the end of a process may contribute almost no judgment at all. The useful human contribution is active: defining success, challenging assumptions, noticing missing context and taking responsibility for a decision. One helpful habit is to form a preliminary view before consulting AI. Another is to ask it for the strongest counterargument, then examine the evidence yourself. Judgment develops through use: we should design AI-supported work so that people have opportunities to exercise it.

8. AI literacy is becoming a practical work skill.

Beyond Marr’s seven themes, I would add the ability to work with AI without either worshipping it or avoiding it. This involves more than collecting clever prompts, it means choosing a suitable task, supplying relevant context, recognizing uncertainty and knowing how to check the result. Consider two requests. “Write a proposal” leaves most of the thinking unspecified. “Use these verified notes to draft a proposal for this audience, distinguish confirmed facts from assumptions, and flag missing information” creates a much better basis for collaboration. As tools handle text, images, audio and documents together, users also need to recognize the limits of each input. A photograph may omit the crucial detail. A transcript may mishear a name. A spreadsheet may contain a beautifully organized mistake. Training should use real tasks, with permission to experiment and discuss failures. A single inspirational presentation is a pleasant afternoon; it is rarely a complete learning program.

9. Productivity needs evidence, not a feeling.

The research gives us a useful warning against universal claims. A study published in The Quarterly Journal of Economics in 2025 examined 5,172 customer-support agents and found that access to AI assistance increased issues resolved per hour by 15% on average. Gains varied across workers. [5]

In a different setting, METR’s early-2025 randomized study found that 16 experienced open-source developers took 19% longer when allowed to use AI on tasks in familiar repositories. Its February 2026 follow-up cautioned that later results were affected by selection and measurement problems, and could not reliably establish the size of the current effect. [6]

These studies concern different work, users and tools. Together, they support a practical conclusion: measure performance in the setting that matters to you. Include review time, error rates and the quality of the finished work. Producing a draft sooner is useful only if the whole task improves. A stopwatch is sometimes a more informative AI consultant than a keynote speaker.

10. We need to protect learning and relationships.

There is another question I would put beside productivity: what happens to people while the work becomes easier? If AI prepares every first draft, how will a junior colleague learn to structure an argument? If it conducts every initial customer exchange, how will a team develop an ear for what customers are really worried about?

These are design questions, not inevitable outcomes. AI can also support learning: offering hints, explaining alternatives or critiquing a person’s attempt before supplying its own. Companies should deliberately preserve opportunities for practice and direct contact. They should also make it easy for a customer or employee to reach a person when the situation calls for one. A system may complete its assigned task while leaving someone frustrated, confused or unheard; A sensible evaluation should notice both results.

The next stage of AI adoption deserves that wider measure of success: better work, capable people and relationships worth keeping. What I find most promising about this moment is the chance to become more deliberate about work we have accepted without much thought. We can ask which tasks deserve automation, which decisions require discussion and where better information would make everyone’s day easier. We can teach people to delegate clearly, test claims and use the time they recover with purpose. The imaginary assistant from our opening morning could become enormously useful. Give it current prices, protect the lunch break, limit its authority and check the promises before they leave the building.

Then finish your coffee. There is still plenty of interesting work for you to do.

Sources and further reading

Starting point. Bernard Marr, “AI Is Growing Up: What World Summit AI 2026 Revealed About The Future Of Business,” 16 October 2026. Full text supplied by the reader. The seven themes are credited to Marr; the discussion, examples and three additional themes here are original commentary.

[1] World Summit AI. 2026 program and mainstage themes: https://worldsummit.ai/programme/ and https://worldsummit.ai/shaping-the-new-ai-paradigm-track/

[2] NIST. AI Agent Standards Initiative, announced 17 February 2026: https://www.nist.gov/news-events/news/2026/02/announcing-ai-agent-standards-initiative-interoperable-and-secure

[3] Google DeepMind. Gemini Robotics 1.5, 25 September 2025. Developer-reported research: https://deepmind.google/blog/gemini-robotics-15-brings-ai-agents-into-the-physical-world/

[4] International Energy Agency. Energy and AI (2025), “Energy demand from AI”: https://www.iea.org/reports/energy-and-ai/energy-demand-from-ai

[5] Brynjolfsson, Li and Raymond. “Generative AI at Work,” The Quarterly Journal of Economics 140(2), May 2025, pp. 889–942: https://doi.org/10.1093/qje/qjae044

[6] METR. Original early-2025 experiment: https://metr.org/Early_2025_AI_Experienced_OS_Devs_Study-paper.pdf ; methodological update, 24 February 2026: https://metr.org/blog/2026-02-24-uplift-update/