Agentic AI

A chatbot answers; an AI agent finishes the job. Given a goal, it plans and acts using tools the organisation provides, such as customer records or payment systems. After each step, it decides what comes next. Traditional automation follows fixed rules and stalls at an unforeseen exception. An AI agent adjusts as it goes. At an insurer, AI agents process claims overnight: collecting emails, photographs and policy details, checking claims against the policy and paying out on straightforward ones. The claims handler starts the day with a handful of files ready for a decision. AI agents have already assessed and settled the remaining claims.

A coding agent builds software on its own. It reads the existing code, makes a plan, writes the change, runs the tests and keeps revising until they pass. The developer describes what is needed, reviews the result and decides what goes into use. In programming this is already routine, partly because tests quickly show where something is wrong. A developer hands an agent an ageing component to rebuild at the end of the day and finds a tested proposal the next morning, with an explanation for every change. What is happening in software development now is an early sign for other knowledge work, such as legal work and financial analysis.

An agent can operate a computer much as a person does: it looks at the screen, recognises buttons and fields, moves the pointer and types. So it can also work in systems with no connection for other software, such as an old accounts package, a supplier's web portal or a government form. On a widely used test of everyday computer tasks, the best agent completed one task in eight in 2024. On the revised version, the best agents had overtaken the human score by the end of 2025. In pilots, they transfer orders from emails into an old order system and ask a colleague for help with a field they do not understand. That brings within reach work that was too fragmented to automate.

A multi-agent system brings several AI agents together to work on a single process, each with its own task. One coordinates and allocates work, specialists carry it out, and another checks the result before anything leaves the organisation. At a wholesaler, this team handles an urgent order: checking stock, arranging delivery, raising the invoice and informing the customer. Increasingly, AI agents also communicate outside their organisation. In trials, a buyer's procurement AI agent requests a price from a supplier's sales AI agent. Together they produce a proposal that only needs the buyer's approval. Transactions between companies begin to move at a different pace.

A customer agent finds and compares offers for a consumer or buyer, and can order within agreed limits. Every offer gains a new judge: an agent that weighs up hundreds of options in seconds. A consumer asks an assistant to book a trip; a buyer has an AI agent request quotations. AI agents can read ordinary web pages, but they compare options more reliably using up-to-date information on prices, delivery times, terms and stock levels. Retailers and travel companies are therefore making their offers readable for AI, and payment companies are building payments that an agent makes on the customer's behalf. Sellers can put their own AI agent on the other side, answering questions from customers' agents and handling the order.

In an AI-first organisation, the agent is the first to get to work. Agents take on repetitive and predictive work: processing requests, drafting quotes, forecasting demand and stock, scheduling people and resources, from start to finish and through the night. People focus on what they do better: judging what is out of the ordinary, choosing between competing interests and working with customers and colleagues. At a wholesaler, an agent links the quote, the stock check and the delivery schedule into one process, with fixed points where an employee decides; the account manager calls the customer who is hesitating over a large order. Organisations are now mapping their processes: which tasks, from processing requests to drawing up schedules, can an agent take on, and where does an employee remain indispensable? How many agents they deploy depends on the tasks those agents can handle and the oversight they require.

Many processes assume people pass work along, inbox to inbox, department to department. Adding an AI agent speeds up one step but leaves the handovers and waiting intact. The gain starts with the process, not with the agent. Designing this way means beginning with the result the customer needs, then reallocating the work: what an AI agent does, where human judgement is needed and how handovers work. For a permit application, an AI agent gathers documents, checks they are complete and requests missing information. The case officer receives a complete file and spends time weighing the application. Chasing paperwork gives way to assessing a case that is ready for review.

In a human-AI team, employees and AI agents work towards the same result. People bring judgement, make choices and handle difficult conversations. AI agents take on repetitive tasks and detailed analysis. At a transport company, an AI agent schedules hundreds of vehicles overnight. Planners begin with the puzzles it could not solve and the customers who need a personal conversation. The familiar questions of teamwork become explicit: who allocates work, how it passes between people and AI agents, and who intervenes when an AI agent's work falls short. The team grows without adding a desk.

As AI agents take on operational tasks, employees spend more time setting briefs, checking results and improving performance. Someone who once did the work now reviews what an AI agent has prepared. New roles appear. AI agent managers adjust instructions and boundaries and track performance. Process owners decide which work to assign to them. Specialists find weaknesses before an AI agent goes live. A claims handler in such a role needs professional knowledge to spot what it has missed. A handbook cannot teach these skills. That is why more and more organisations train their own people, in a future academy of their own where they practise with agents in their everyday work.

Handing work to a colleague takes adjustment. Handing it to an AI agent running in the background is a bigger step. The challenge is largely behavioural: being willing to let go, and willing to intervene. Each AI agent has defined permissions and signals for human takeover, and records every step. A team can then see which data was used to handle each customer enquiry and what action followed. How much gets checked depends on what is at stake. Trust builds gradually. An AI agent that delivers good work for months gets more freedom, much as a new colleague earns it.

An AI agent register records an organisation's digital colleagues, just as personnel records do for people. It goes beyond a software list: AI agents receive instructions, act independently and collaborate. Each entry shows the owner, version, systems, data and permissions. At an insurer, it reveals which AI agents handle claims and who monitors their work. AI agents built by employees are recorded too. In organisations that may soon have more AI agents than people, the register adds a digital layer to the organisation chart, showing who and what keeps the business running.

Leaders working with AI agents focus on results rather than hours or headcount. They decide which tasks to delegate, how much freedom to allow and where people decide. Increasingly, leaders oversee processes instead of working within them: setting direction, watching where it matters and intervening when circumstances change. Work taken on by AI agents affects jobs, roles and habits, shaping where the organisation could be in five years. An executive team has AI agents model ten growth scenarios, including staffing, capacity and cash flow, and then tracks them week by week. Boardroom discussion shifts away from reviewing last month towards choosing between possible futures.