The first wave of workplace AI answered questions. The second wave, now arriving, does jobs: agents that read the inbox, draft the report, reconcile the spreadsheet, chase the invoice and book the meeting, checking in with a human only when judgement is required. Employers who spent two years experimenting with chatbots are beginning to reorganise real workflows around software that can act, not merely advise.
The shift is visible in procurement as much as in technology. Software is increasingly sold by outcome — tickets resolved, documents processed, leads qualified — rather than by seat. Our technology coverage tracks the deployments, and a pattern is emerging: agents succeed first where tasks are frequent, rules are clear and errors are cheap to catch, from IT helpdesks to accounts payable.
For staff, the honest question is not whether agents arrive but which parts of their job the agents absorb — and what becomes more valuable in the humans who remain.
Where Agents Are Working Today
Customer support is the beachhead: agents resolve routine queries entirely, escalating the emotional or exceptional ones to people. In finance departments, agents match invoices to orders and flag anomalies. In software teams, coding agents draft and test routine changes under review. The World Economic Forum’s Future of Jobs research describes employers expecting exactly this division — machines absorbing information-processing tasks while human roles tilt toward oversight, relationships and accountability.
The failures are instructive too. Agents deployed on ambiguous, high-stakes work without supervision have produced confident nonsense at scale. Successful programmes share a design discipline: narrow mandates, clear escalation paths, and audit logs reviewed by people whose job explicitly includes doubting the machine.
The New Skill: Supervising Software
Managing agents is becoming a recognisable professional skill, part editor, part auditor, part process designer. Someone must write the instructions the agent follows, set the limits it cannot cross, sample its output and retrain the workflow when the world changes. Our technology section interviews practitioners who describe the role as closer to managing a very fast, very literal junior team than to using a tool — tireless, knowledgeable and entirely without common sense.
That framing helps organisations plan. Headcount models built on ‘AI replaces X per cent of staff’ consistently miss; the firms gaining most treat agents as capacity that must be directed, and invest in the directing. The bottleneck moves from doing the work to specifying and verifying it, which is why experienced staff — who know what good looks like — often become more valuable, not less.
What Employees Should Do Now
Practical advice from deployment veterans is consistent. Learn the tools in your own field before they are assigned to you, because fluency is visible and valued. Volunteer for the oversight work — reviewing agent output builds exactly the judgement that stays scarce. And keep a record of outcomes: in a workplace measured by results, the person who can show what their agent-assisted team delivered negotiates from strength.
Coverage by the BBC’s technology desk strikes the balanced note the moment deserves: agents are genuinely capable, genuinely fallible, and arriving regardless. The workplaces that handle them best will not be those that adopt fastest or resist longest, but those that stay clearest about the division of labour — machines for the tireless and the routine, people for the accountable and the humane. That clarity, more than any single product, is the real workplace revolution now under way. The organisations learning fastest share one habit: they pilot agents on a single workflow, measure honestly, publish the results internally and only then scale. That discipline sounds unglamorous beside the launch headlines, but it is what separates durable transformation from expensive theatre, and it is teachable to any management team willing to slow down for a month in order to speed up for a decade.
