AI agents promise to work without human intervention. But someone must still mind the machines. The challenge is to make supervision cheaper without making responsibility optional.
At a bank, the night shift is getting crowded. A customer applies for an account. One artificial-intelligence agent examines her documents, another consults credit records, a third checks her identity. Others deal with inconsistencies and decide which cases deserve attention. The process can involve five to seven agents, conventional software and several teams of humans. Most of the work takes place without a clerk. Then an awkward file appears. It waits for a risk specialist, who spends perhaps five minutes reviewing it.
Those five minutes are surprisingly expensive. In an illustrative banking workflow, human oversight accounts for 70–75% of the variable cost of running an agent. Tokens, the chunks of text a model reads and writes (think of them as the meter on the taxi), account for just 20–25%. The technology is cheap enough to do much of the work. Keeping an eye on it is another matter.
This exposes a contradiction at the heart of the agent boom. Businesses want machines that can act without waiting for permission. Their employees, customers and regulators want reassurance that somebody remains in control. The more independence firms grant, the more they worry about what might happen. The more closely they supervise, the less they gain from automation.
The next stage of the AI revolution will therefore be decided as much by the economics of supervision as by the intelligence of the machines. The winners will be the firms that learn which decisions still need a human, which can be left to software and who answers when things go wrong.
Trust me, I'm supervised
For all the enthusiasm about autonomous AI, agents that act alone remain rare. Some 59% of large enterprises report using AI agents, but only 9% say they have converted that investment into autonomous workflows. Companies have bought the technology without necessarily acquiring the confidence to use it. The hesitation is understandable because there is a difference between software that suggests a payment and software that makes one. The former can be ignored but the latter may need explaining to an unhappy customer, an auditor or a court.
Trust does not require pretending that machines are infallible. Experiments suggest that acknowledging an AI system's weaknesses can make it appear up to 14.6% more transparent and help people work with it up to 7.2% more effectively. A machine that knows its limits is more reassuring than one that confidently oversteps them. Yet reassurance can become an expensive habit. Require a human signature for every consequential action and the supposedly autonomous agent becomes an elaborate assistant. Give it complete freedom and the business inherits risks it may neither understand nor be able to contain.
The sensible compromise is selective autonomy. A system might gather information, prepare documents and resolve familiar discrepancies by itself, while referring unusual or irreversible decisions to a person. What matters is that a human steps in where doing so can change the outcome. Few organisations have yet settled where those boundaries lie.
The minder's bill
The temptation is large. In one illustrative model, automating bank-account onboarding reduces the cost per customer from roughly $50–150 to $10–30, even after accounting for human reviewers and supporting systems. But the same sums show why the case can disappoint. An agent's cost includes the plumbing that connects it to other systems, the tools it calls, the upkeep that keeps it reliable and the people who inspect its output. Worse, producing work faster does not necessarily mean producing better work. Across a survey of 334 product and engineering respondents, AI reduced the time spent on tasks by an average of 11.8%, but reduced rework by only 6.2%. That suggests an imbalance: creating output is becoming easier faster than ensuring its quality.
Software development provides a useful warning. In one security study, 45% of tested AI-generated code samples failed security checks. More output, faster, means more chances for mistakes to slip through. A firm can add reviewers, but if every extra unit of output needs extra inspection, the savings evaporate. Machines work around the clock; the people checking them still draw salaries. Cheaper computing will not fix this, because the model is a small part of the bill. A firm that spends heavily correcting mistakes may find that its cheapest agent is its most expensive employee. The measure that matters is therefore the cost of completing a task correctly, not the cost of generating an answer.
Garbage in, decisions out
There is another reason machines need minding: they inherit the disorder of the organisations that employ them. Consider a bank that has accumulated decades of customer records, product descriptions and internal rules across incompatible systems. One database contains an outdated address; another uses a different definition of customer risk. Staff have learned to work around these inconsistencies, often over coffee. An agent may simply treat whichever answer it finds as authoritative.
The consequences are different from those of an ordinary reporting error. A flawed dashboard misleads someone who might still question it. An autonomous agent can act on the same bad information before anyone notices. This makes the quality of the underlying data as important as the intelligence of the model. Agents need dependable information, clear rules about which records take precedence and a means of reconstructing the evidence behind their decisions. In regulated industries, someone must also be able to establish why a system reached its conclusion.
Security presents a similar difficulty. Giving an agent access to customer files, payment systems and external software may make it useful. Giving it unnecessary access makes it dangerous. A compromised agent with broad permissions can spread trouble across several connected systems. The cheaper remedy is to restrict what the machine may do in the first place. Agents should have identities, limited permissions and rules governing the information they can access, retain or share. Proposed actions can be checked automatically against those rules before execution. Properly designed, governance is what allows autonomy to expand.
From wages to wiring
The distinction is between supervising individual actions and governing the system that produces them. The first relies on people, and its cost rises with every case that needs attention. The second relies on design: a central layer of software decides which agents may do what, under what conditions and with what evidence. Routine work proceeds, sensitive decisions are escalated and each significant action leaves a record.
One form of control is paid for largely in wages; the other is paid for in systems, maintenance and rules that can be reused across transactions. The wiring is not cheap either: spending on tokens, computing, infrastructure and AI operations can reach 20% of existing labour costs. But unlike wages, that bill does not rise with every extra case. Agents can be granted more freedom as their record improves. They can handle familiar exceptions automatically, while uncertain cases, sensitive information and high-impact decisions receive closer attention.
Even the checking can increasingly be automated. Software agents can test the work of other agents, compare results against requirements and flag possible defects. One retailer that embedded such checks into its software-delivery process reported a 50% reduction in the effort needed to resolve defects.
There remains an obvious question: who checks the checker? Banks already possess a useful answer in their established model of risk management. Business units own their risks; independent control functions challenge the business; auditors examine whether the arrangements work. The principle does not require three people to approve every transaction. It requires responsibilities to be distinct, controls to be effective and failures to be traceable. Autonomous systems need something similar. An agent may execute a decision, another may verify it, and software may enforce the rules. But a named person must remain responsible for the system's performance. Delegating the work should not mean delegating the blame.
Management by exception
For much of corporate history, supervision has meant directing employees, reviewing their work and approving decisions. AI agents challenge that arrangement because they can perform tasks, coordinate activities and respond to events without waiting for the next meeting. Organisations that retain every old approval chain risk paying for autonomy while continuing to operate as before.
Tools alone do not deliver. Teams are shrinking: at the best-performing software organisations, 79% of respondents say squads got smaller after adopting AI, with the median falling from about ten people to seven. In the same survey, firms that redesigned roles around AI were likelier to become top performers than those that simply bolted the tools on: 40% against 27%. Yet the future shape of management remains contested. One possibility is that faster coordination will allow firms to operate with fewer managerial layers. Another is that more employees will become supervisors of systems as machines take over execution. Both may prove true: flatter hierarchies, with more people minding machines.
The managerial skill in greatest demand may be designing the conditions under which work is done: establishing boundaries, measuring results, deciding when intervention is warranted and recognising failures before they multiply. That calls for a different way of judging performance. Counting how many people a manager supervises, or how much output a team produces, tells little about the reliability of a mixed workforce of people and machines. More revealing measures would include the proportion of cases resolved correctly, the cost of exceptions, the frequency of intervention and the consequences of mistakes.
A manager paid on those numbers has every reason to make supervision sharper.
The buck stops somewhere
Optimists may argue that this problem will fade as AI becomes more reliable. Better models should make fewer errors; cheaper computing should make automated checks less costly. Workflows that are uneconomic today may become attractive within months. Perhaps, but falling error rates do not settle the question of responsibility.
Imagine an agent that misrepresents a product during a sale, publishes a defamatory accusation or makes an unauthorised purchase. The machine acted, but who should answer for the harm: its user, its developer, the company deploying it or some combination of them? Existing notions of intention, foreseeability and control become harder to apply when software can independently choose how to pursue an instruction.
Courts will take years to settle this. Managers cannot wait: no firm should discover its accountability arrangements only after something has gone wrong. The real prize of agentic AI is not an organisation without people. It is one in which human attention is reserved for decisions where judgement matters, rather than consumed by checking work that reliable systems could handle themselves.
The machines will need less watching. Someone will still have to answer for what they do.
References
McKinsey (2026), “Where AI agents pay off: A practical guide to the economics of agentic workflows” — https://www.mckinsey.com/capabilities/quantumblack/our-insights/where-ai-agents-pay-off-a-practical-guide-to-the-economics-of-agentic-workflows#/
Harvard Business Review (2026), “To Adopt AI at Scale, Employees Need to Trust Agents” — https://hbr.org/2026/09/to-adopt-ai-at-scale-employees-need-to-trust-agents
MIT Technology Review (2026), “How to navigate the age of agentic AI“— https://mitsloan.mit.edu/ideas-made-to-matter/how-to-navigate-age-agentic-ai?utm_source=mitsloantwitter&utm_medium=social&utm_campaign=smragentic
McKinsey (2026), “Beyond the copilot: Scaling the agentic product development life cycle” — https://www.mckinsey.com/industries/technology-media-and-telecommunications/our-insights/beyond-the-copilot-scaling-the-agentic-product-development-life-cycle#/
MIT Technology Review (2026), “The era of agentic chaos and how data will save us — https://www.forbes.com/sites/larryenglish/2023/05/25/the-impact-of-ai-on-company-culture-and-how-to-prepare-now/
MIT Technology Review (2026), “Data readiness for agentic AI in financial service” — https://www.technologyreview.com/2026/05/14/1137034/data-readiness-for-agentic-ai-in-financial-services/
MIT Technology Review (2026), “Building agent-first governance and security” — https://www.technologyreview.com/2026/04/21/1136158/building-agent-first-governance-and-security/
McKinsey (2026), “How agentic AI challenges four rules your organization runs on” — https://www.mckinsey.com/capabilities/people-and-organizational-performance/our-insights/the-organization-blog/rethinking-early-career-talent-in-the-agentic-organization
The Economist (2026), “Agent of whom?” — https://www.economist.com/business/2026/09/23/agent-of-whom
