Most of the conversation around agentic AI is about what it can do. Build agents. Deploy agents. Scale agents. Automate this, orchestrate that. Fair enough – the technology is genuinely exciting, and the speed at which it’s moving is enough to give anyone a mild case of future shock.
But there’s a quieter shift happening underneath all the noise, and it’s got less to do with the technology itself and more to do with how people think about their own work. Because agentic AI doesn’t just change what gets done. It changes how you decide what gets done, and by whom – or by what. That’s a different kind of problem.
For decades, the way we’ve thought about work has been fairly straightforward. You have a role. Your role has tasks. You do those tasks (some well, some grudgingly, some while pretending to be on a call). The organisation structures itself around roles, departments, and hierarchies. Everyone knows where they sit on the org chart, even if nobody loves it.
Automation has been nibbling at the edges of this for years. RPA handles the repetitive stuff. Analytics flags the patterns. Generative AI drafts the emails you didn’t want to write. But none of this really forced anyone to rethink the model itself. You still had a to-do list. You still owned it.
Agentic AI is different. Not because it’s smarter (although it often is), but because it operates with a degree of autonomy. It can plan, act, adjust, and learn. It can take a goal – “review these 200 insurance claims for compliance issues” – and figure out the steps on its own, without you mapping every click. That changes the equation. Suddenly, the question isn’t “how do I do this task faster?” It’s “should I be doing this task at all?” And that’s where the old mental model starts to creak.
McKinsey has a useful way of framing this. They argue that traditional organisation charts – built on hierarchical delegation and reporting lines – are giving way to what they call “work charts.” These map the exchange of tasks and outcomes between humans and AI agents, rather than who reports to whom.
It’s a small linguistic shift, but a significant conceptual one. A work chart doesn’t care about your job title. It cares about what needs to happen, and who (or what) is best placed to make it happen.
ISG puts it more bluntly: “Enterprises must start reorganising work around outcomes, not job titles. Agentic AI does not slot into existing org charts easily or cleanly.” No, it does not. That’s precisely the point.
Here’s the mental shift I think matters most. When you sit down to plan your work – or your team’s work, or your department’s work – you’ll increasingly need to think not just in tasks, but in agents. Which parts of this workflow can an agent handle? Where does it need a human? Where does the handoff happen? What does the human do with the time that’s freed up?
This isn’t hypothetical. Microsoft’s 2025 Work Trend Index, based on a survey of 31,000 workers across 31 countries, found that 82% of leaders expect AI agents to be moderately or extensively integrated into their AI strategy within the next 12 to 18 months. Forty-six percent of employees already describe AI as a “thought partner.” And Microsoft coined a term for the new role that every employee will eventually play: the “agent boss.”
It’s a slightly cringeworthy label (corporate naming committees have never met a metaphor they couldn’t overwork), but the idea behind it is sound. The human’s job isn’t to do everything. It’s to know what the agents should do, what you should do, and – critically – to spot the difference.
McKinsey’s practical framework helps here. When deciding what tool fits a particular task, they suggest:
- Rule-based, repetitive, structured input? Classic automation (RPA).
- Unstructured input, but the task is extractive or generative? Generative AI or NLP.
- Classification or forecasting from historical data? Predictive analytics.
- Multi-step, goal-oriented, requires planning and adaptation? That’s where agentic AI lives.
This is useful because it forces you to actually look at what you do and classify it. Not by importance or by who’s always done it, but by what kind of intelligence it demands. A lot of people haven’t done that exercise. Most organisations certainly haven’t.
The pace here is remarkable. Gartner predicts that 15% of day-to-day work decisions will be made autonomously through agentic AI by 2028 – up from effectively zero in 2024. They also forecast that 33% of enterprise software will include agentic AI by the same year, up from less than 1% today.
MIT Sloan Management Review found that the speed of agentic AI adoption inside organisations has outpaced the adoption of both traditional AI and generative AI. BCG reports that implementations are already reducing task completion times by 25 to 40%, with agents working around the clock.
And yet – here’s the part that should worry anyone in strategy or leadership – McKinsey reports that 89% of organisations still operate with industrial-age models. Nine percent have moved to agile or platform-based structures. Only 1% function as the kind of decentralised networks that agentic work actually requires. That’s a 1% readiness rate for a transformation that 82% of leaders say is coming within 18 months. Not great.
Now, I could write a breathless piece about how we’re all going to become agent bosses and the future is beautiful. But the evidence doesn’t fully support that – at least not yet.
The reliability problem is real. The Scale AI Replicate Labor Index found that even the highest-performing AI agent workflows achieve only about a 2.5% success rate for fully autonomous, end-to-end completion of complex tasks. Debevoise & Plimpton’s analysis is characteristically measured: “Current frontier GenAI models are generally not able to reliably and consistently handle a series of complex tasks entirely on their own.” When you chain multiple agent steps together without human review, errors compound. A 95% accuracy rate across ten steps gives you a 60% chance of at least one failure. The maths is not on your side.
The cognitive atrophy problem is real, too. Gartner predicts that by 2026, the decline in critical-thinking skills caused by generative AI overuse will push 50% of organisations to require “AI-free” assessments in hiring and promotions. Gartner analyst Manjunath Bhat calls this “experience starvation” – when people delegate so much that they never build the hands-on intuition needed to do the work themselves. It’s a genuine paradox: you need to understand the work deeply enough to know when the agent is getting it wrong. But if you never do the work, you lose that understanding.
And then there’s the people problem. Research from Spend Matters suggests that employee sabotage and resistance may end up being a bigger risk than technical failures. People who feel replaced, rather than augmented, push back – sometimes openly, sometimes quietly. Meanwhile, Moxo’s research points to “automation bias”: when humans are asked to review AI output, they tend to rubber-stamp it. The safety net isn’t as safe as it looks.
None of this means the shift isn’t happening. It means the shift is hard. And “think in agents” is not a binary exercise – it’s a spectrum. Some tasks will be fully autonomous. Some will be fully human. Most will live somewhere in between, and the boundary will move constantly depending on context, stakes, and how mature the technology is in that specific area.
If this is where work is heading, then the most valuable professional skill of the next few years won’t be “using AI.” It’ll be something more specific: knowing the boundaries of what the agent can do. Knowing when to trust it, when to intervene and, when it’s getting things wrong in ways that look right.
Forbes calls this the shift from “doing” to “orchestrating.” I’d call it agent literacy – and it’s a competency that most organisations haven’t started building yet.
It also creates a paradox that’s worth being honest about. To be a good orchestrator, you need to have done the work yourself. You need the intuition that comes from experience. But if every new employee starts by delegating to agents, when do they build that intuition? Gartner’s “experience starvation” concept isn’t a theoretical risk. It’s a real design problem for how we train people.
The instinct, when talking about any big shift, is to frame it as a before and after. Before: you did the work. After: the agent does it. But that’s not what’s actually happening.
What’s happening is more like adding a layer to how you think. You still need to understand your tasks, your processes, your objectives. That doesn’t go away. But on top of that, you now need a second layer of reasoning: which of these can be decomposed, delegated, and orchestrated through agents? And how do I stay skilled enough to oversee the whole thing?
Deloitte says it well: agents are “a new form of labour.” Not a tool, not a shortcut – a new kind of worker that shares the workload. The organisations that get this right will be the ones whose people develop this additional cognitive muscle. The ones that get it wrong will either over-automate and lose control, or under-automate and lose competitiveness. Both outcomes are available. Choice is yours.
Ready to build your agent literacy? Start asking: what shouldn’t you be doing?