Most corporate experience of artificial intelligence so far has involved asking it something and reading the answer. Useful, occasionally impressive, but the human still does the work. Agentic AI changes the arrangement by giving software the ability to plan a sequence of steps, use ordinary digital tools to carry them out, and come back with a finished piece of work.

The difference may sound small, but it is actually quite significant. A model that drafts an email saves a few minutes, an agent that reads the account history, drafts the email, checks it against the pricing system, sends it and logs the outcome removes an entire task from someone’s day. Adoption remains early, with most organisations still running small prototypes, but the direction of travel is clear enough to plan around.

Which processes are suitable, how output quality gets checked when nobody watched the work being done, and where the productivity actually shows up on a profit and loss statement.

From answers to outcomes

Three skills separate agents from the tools that came before them, they handle the awkward exceptions that rule-based automation never coped with, because language models respond sensibly to situations they have not encountered before. They use the same interfaces a person would, including web browsers and internal systems, which removes the integration work that historically made automation projects expensive and slow.

The third skill matters most for anyone responsible for output. Agents produce a written work plan before they act, in plain language, which can be read and corrected. A manager can see the intended approach, adjust it, and let the agent proceed. Automation that could previously only be audited after the fact can now be steered while it happens.

More sophisticated setups run several agents together, with one acting as a manager that builds the plan and delegates pieces to specialists. QuantumBlack Labs used exactly this structure to automate credit memo drafting at a bank, with subagents handling data analysis, verification and output. Analyst productivity rose by as much as 60%, which is the sort of figure that changes a hiring plan.

The first real wins

Software development has moved fastest, partly because code can be tested automatically, which gives an agent immediate feedback on whether its work was any good. That property is worth noting because it explains the pattern. Work with a clear, checkable definition of success is work an agent can own.

Customer operations follow the same logic. Salesforce’s platform puts agents to work resolving support tickets, scheduling meetings, sending follow-ups and qualifying leads, all tasks where the outcome is unambiguous. Darktrace applies the approach to security, with agents monitoring network traffic continuously and deciding how to respond to anomalies, freeing human analysts for the judgement calls that actually need them.

The common thread across these deployments is not job replacement but redistribution. Routine surveillance, first-pass drafting and administrative sequencing move to software, while people concentrate on exceptions, relationships and decisions with consequences. Firms reporting the strongest results have generally been explicit about that redistribution instead of leaving teams to work it out themselves.

Choosing the work to hand over

A useful filter is to look for tasks that are repetitive, well documented, and produce an output someone already checks. Credit memos qualify. Support ticket triage qualifies. Board strategy does not, and neither does anything where the right answer depends on context nobody has written down.

A second filter concerns volume. Agents pay for themselves through repetition, so a process performed forty times a week is a better candidate than one performed twice a quarter, even if the quarterly task takes longer. Organisations that start with their most painful problem often pick something rare and complicated, then conclude the technology does not work.

The third consideration is data. Agents perform best where there is a decent record of how the work has been done before, because that record becomes the reference for doing it correctly. Departments with tidy documentation tend to see faster results, which is an unglamorous argument for fixing process documentation before buying anything.

Managing what nobody watched

Handing work to an agent does not remove the need for oversight, but instead relocates it. Somebody has to define what a good outcome looks like, review a sample of completed work, and decide when the agent may act alone and when it must ask. That is a management responsibility, not a technical one, and it is the part most commonly left unassigned.

Accountability deserves particular attention as agents begin executing transactions and operating across systems. Questions of liability when an autonomous system makes a costly error have not been settled by regulators or by the courts, so the sensible position is to define internally where authority stops. Spending limits, approval thresholds and escalation rules are ordinary governance tools that apply perfectly well here.

‘Agentic AI moves AI from a passive tool to an active collaborator with enterprise workflows. As these systems gain autonomy and decision-making capabilities, it is also critical to invest more in figuring out how to work with AI when it’s seen as a colleague versus a tool. At the same time, we will need strong governance, transparency, and ethical guardrails to ensure that these agents operate with accountability and build lasting trust.’

Delphine Nain Zurkiya, Senior Partner, McKinsey & Company

Starting narrow and building outward

The most productive approach observed so far involves picking one process, instrumenting it properly so the before and after can be measured, and running it long enough to produce evidence. Measurement matters more than ambition at this stage, because the case for expanding will need to be made to people who were not in the room for the pilot.

The number of roles advertised for building these systems grew by 985% between 2023 and 2024, and the skills in shortest supply are the ones agentic work depends on most. Companies that wait until they have a proven business case to begin hiring will be competing for the same small pool as everyone else who waited.

‘AI agents won’t just automate tasks, they will reshape how work gets done. Organisations that learn to build teams that bring people and agent coworkers together will unlock new levels of speed, scale, and innovation.’

Lareina Yee, Senior Partner and Director, McKinsey Global Institute at McKinsey & Company

Adoption sits early enough that a company beginning now is not behind, and the tooling has matured enough that a first project no longer requires a research team. What it does require is a clear-eyed choice about which work to give away first, and the discipline to measure whether giving it away made anything better.