For several years, the main AI question was whether people would use the tools. In 2026, the question changed. AI agents can now pursue goals across several steps with little supervision.
Leaders now face a design question. What work should agents do, what should people do, and who checks the join?
What agents can now do
An agent is software that can plan, act, check its results and adjust. Earlier tools answered when asked. Agents take on whole tasks.
The improvement has been fast. Stanford's AI Index 2026 found agents' success on a benchmark of real-world tasks rose from 20% to 77.3% in a year.
Use is following. Microsoft's 2026 Work Trend Index reports active agents in Microsoft 365 grew fifteen times in a year. BCG found 61% of frontline workers think agents could do half their job.
What agents still cannot do
Agents are much weaker at working together. The 2026 CooperBench study found coding agents were 30% less successful when coordinating than when working alone.
Human-AI teams change how people work too. Ju and Aral found people in human-AI teams sent 18% fewer messages to each other. Output rose, but it became more alike.
Agents also do not push back. They will not notice that a goal is wrong or raise an ethical concern. Judgement about what is worth doing stays with people.
Build this capability in your organisation
Growth Performance designs bespoke programmes that turn insight into measurable behaviour change.
Expectations are running ahead
McKinsey's State of AI 2026 found 39% of respondents expect headcount to fall in the next year. Yet only 37% report any effect on profit from AI so far.
Expectations of agents are rising faster than the results. That gap is a leadership problem.
What leaders need to do
Decide who decides. List the key decisions in your area. Mark which stay with people, which agents can take and which need both. If you do not decide, decisions drift to agents by default.
Redesign the checking. Agents create more output to check. Count checking as real work and give it time. That is Adaptation.
Keep people talking. Watch for teams that stop talking to each other because agents sit in between. Join-up work stays human.
Hear the worries. People need to feel safe saying what agents get wrong and what worries them. That helps them take on the change, which is Adoption.
Protect people's limits. Agents never sleep. Set clear boundaries so people can keep their judgement and wellbeing over time.
How fast can you learn?
Agent capability will keep moving. The organisations that benefit will be ready for each change as it arrives. That is Acceleration.
Try this
Pick one workflow where an agent could take on part of the work. Map who checks the agent, how long checking takes and what happens when it is wrong.
Then use the Checking Capacity Calculator to see whether your team can absorb the extra checking.