Just when many of us were getting comfortable handing small tasks to AI, it found a creative new way to get things wrong (full marks for confidence, though!).
An AI agent was recently asked to work from one specific document. It went looking, found a much older one on a similar topic in a completely different folder, and relayed that content back as if it were exactly what had been requested. There was no hesitation and no caveat (and it would have probably been missed if it hadn’t been for the reference to Windows 95’).
What stood out was how convincing it was. Nothing in the response suggested a problem, because the problem was the source.
It was a bit like working with a bright new team member who knows a lot, moves quickly and occasionally heads off in the wrong direction. Plenty of us now hand work to AI chats, workflows and agents much as we would to an intern, with the same need to explain the task properly and check what comes back. Some people even give their AI a name (“Mr Smee” has a certain appeal, although moments like this make it worth being clear about who is steering the ship).
Of course, the comparison only goes so far. An intern who grabbed the wrong file would probably feel a bit sheepish and remember it next time. AI carries no accountability for the work, and a task that went well last week tells you very little about how the next one will go. Anyone who has parented an 11-year-old will recognise the principle.
What kind of working relationship do we want to build with AI?
AI needs more than a task description
Software has long helped people organise information and complete tasks. AI can now contribute to the work that happens between instructions, such as interpreting a request, proposing an approach and taking steps across applications. Microsoft’s 2026 Work Trend Index supports this direction. It reported a fifteenfold increase in active agents in Microsoft 365 over a year, with its most advanced users already handing agents multi-step workflows.
That changes what good delegation looks like. Imagine asking AI to prepare a stakeholder briefing. It may produce something clear and well organised, yet miss that a previous disagreement that makes one recommendation particularly sensitive. The information might be available somewhere in the organisation. It may also sit in someone’s memory of a conversation that never made it into the meeting notes.
People still need to decide what context matters, what AI can take on and where to pause for a human decision. As systems gain permission to act, those boundaries become more consequential (we have plenty of examples of what happens when those boundaries are not in place now).
The wrong-document moment is a small version of this. One simple habit helps. When you point AI at a source, ask it to confirm which file it used, where that file lives and when it was last updated, before it tells you what the file says. It takes ten seconds. More importantly, a question worth exploring before handing over a task is:
What does AI need to understand, and what still needs a person’s approval?
When AI does more, who’s thinking it through?
AI can assemble information, compare options and draft a recommendation quickly. That creates more opportunity to examine the work, provided people make the time and have the expertise to do so.
A polished AI-written briefing is not always what it’s cracked up to be. Clear language and a sensible structure can give an answer more authority than its reasoning deserves. This can impact the quality of decision making happening in an organisation.
There’s also “AI Slop” to consider (poor quality AI writing - we’ve all seen it!). In Microsoft’s 2026 Work Trend Index, half of AI users named quality control of AI output as a skill that matters more as AI takes on work. The person reviewing the work still needs to notice the missing perspective, question the assumption and recognise when a recommendation would create a problem elsewhere. When AI is used for writing without critical thinking or human discernment we accidentally (or maybe intentionally at times?) shift the onus for understanding from the writer to the reader. If we are also using the AI for the reading and understanding part too, we risk complete disengagement from start to finish. All of this can do damage in an organisation.
There is a capability question underneath this too. People often develop judgement through doing the work. A newer employee learns by gathering information, making an initial assessment, discussing it with someone more experienced and discovering what they missed. If AI takes on more of those early steps, how will that employee develop the expertise needed to evaluate its recommendations later?
A Human-Centric AI approach pays attention to both the immediate result and the learning happening through the task. AI might help someone compare their own assessment with another perspective, explore why a recommendation is weak or practise a difficult decision with feedback. How it’s used shapes what the person gets to practise.
AI is reshaping the team discussion
By the time someone reviews an AI recommendation, it may already have shaped the discussion. What AI includes in a summary, or puts first in a draft, can influence what people pay attention to and which questions they ask. Teams don’t normally rely on an intern to fully drive a discussion, and yet here we are letting AI do it. Would we be so inclined if we thought of AI more as an intern?
Imagine a team coming to a meeting having read the same AI-generated summary. Everyone has saved some preparation time, but an unresolved concern is missing. Unless someone remembers it and raises it again, the team may move on as though it has been dealt with. Leaders may need to consider what might be missing from the material everyone is working from. In that meeting, does anyone ask whether the summary reflects what was discussed? And would someone feel comfortable reopening a concern the rest of the team seems ready to move past?
The time saved could help here. A team could use it to work through something they are unsure about or hear from a colleague who deals with the issue every day. But if saving preparation time simply means bringing the next meeting forward, that opportunity may disappear. How we use that time deserves as much thought as how we save it.
AI can help a team arrive prepared, but people still need room to add what they know, disagree with the way an issue has been framed and bring an unresolved concern back into view. Leaders influence whether that happens through what they ask and how they respond. The time saved becomes more valuable when it gives that team discussion room to happen.
Existing leadership habits travel with the technology
Introducing AI into a team also introduces it into the habits already there. A manager who welcomes questions may treat AI-generated work as something to examine together. A manager who rewards fast answers may unintentionally encourage people to accept the first plausible response. Employees will notice which behaviour gets recognised.
Gallup’s State of the Global Workplace 2026 found that employees whose manager actively supports AI use are far more likely to use it often (79% compared with 46%). In the US, fewer than a third strongly agree their manager does.
Leadership support does not look like the “everyone must use AI” mandates we see in many organisations. Instead, the responsibility shows up in ordinary moments. Moments like when a leader admits AI got something wrong or someone is given time to verify a result. It can even include an employee raising a concern about using AI for a sensitive task and the leader supporting them.
Those interactions help people understand what good AI use means in practice. They also reveal whether the organisation’s expectations support the judgement it says it values.
Try it with one piece of work
Human-Centric AI is the deliberate use of AI to strengthen what makes us human.
A practical place to explore that intention is a task where AI is already doing more of the work. You can use the following prompt to examine the relationship developing around it.
An AI prompt for setting the work agreement
The prompt below treats AI a little bit like onboarding a new team member, and finishes with a short working agreement you can test on your next piece of work:
“I’d like to set up a better way of working with AI on [describe the task and what AI does in it].
Treat this like onboarding a capable new team member. Ask me one question at a time to understand:
what the task involves and what a good result looks like
which sources AI should use, and how I’ll know it used them
where AI has got this kind of task wrong before, or could
which parts I want to keep doing myself, because they build my judgement or rely on context only I have
what I’ll do with any time saved
Push back if I’m trusting AI with something I haven’t checked, or checking things that don’t need it. Finish with a short working agreement for this task: what AI does, what it shows me, what I check and what stays with me. Keep it to five lines so I can actually use it.”
A final thought
The intern metaphor captures something recognisable about working with AI. There’s the promise and the speed, and every so often the need to ask where on earth that answer came from (or, in this case, which folder).
The responsibility stays with the people designing, directing and using it. How the relationship develops will depend on how clearly we explain the work, how carefully we check it and whether people keep learning while AI does more of it.
The AI intern may have arrived, but there is plenty of human work ahead.
Written by Jo Grubb
Human Shift Advisory - 2026
A note on AI use in this article: For this article, AI helped with image creation, research discovery, pressure-testing ideas and editing. The thinking, synthesis, interpretation and conclusions remained human. Industry experience and expertise shaped the ideas and determined what to use, challenge, change or leave behind. This reflects the philosophy behind Human-Centric AI, which is the deliberate use of AI to strengthen what makes us human.