All insights

A shift in the unit of work

Jevons uses the Agentic Economy as a strategic framework for a business environment in which AI agents can take bounded steps toward a goal: interpret a request, obtain information, use software tools, assess a result and decide what to do next. It is a useful strategic lens, not a claim that every company or process is already autonomous. The relevant question is which decisions, transactions and customer experiences can be redesigned when software can participate in a sequence of work.

An agent is part of a wider operating system of people, data, rules and software. Its authority comes from the permissions and controls an organization grants it. It can prepare an action without having authority to execute it; an accountable person remains responsible for the outcome.

How an agent differs from traditional automation

Traditional automation typically follows an explicitly specified route: when a purchase order meets a rule, send it for approval. An agent can choose among permitted steps in response to an objective and changing information—for example, compare a request with policy, look up missing details and prepare an exception for review. Anthropic distinguishes predefined workflows from agents that direct their own processes and tool use; real systems can combine both.

A copilot usually assists a person within a task, while an agent can select several permitted steps toward an objective. The distinction is a matter of design and degree, not a guarantee of intelligence or reliability. A fixed rule may be cheaper, easier to test and more appropriate for stable, high-volume decisions. Use an agent when the variability and value of the task justify the extra complexity, and keep deterministic checks around consequential actions.

Where leaders can start

Consider a distributor handling a delayed order. An agent with read access could gather inventory, shipment status and contract terms, identify plausible options and draft a customer response. A service employee checks the evidence and decides whether to change the order or offer compensation. The example is illustrative: its value depends on data quality, authority and the actual process.

Other bounded uses include preparing a supplier-risk brief from approved sources, assembling evidence for a financial close, or routing a service case with its relevant history. Begin with one end-to-end outcome and its exceptions. A demo that completes the easy cases is not evidence that the full operation will work.

The economics: capacity is not the same as profit

Agentic systems may change the cost, speed and availability of parts of a service. They can also create new costs: model and tool usage, integration, monitoring, human review, security, corrections and vendor management. Calculate cost per accepted outcome rather than cost per generated response, and include failures, rework and the expense of maintaining the workflow. Research by Brynjolfsson, Rock and Syverson explains why general-purpose technologies require complementary investments that can be difficult to measure; it is not an estimate of agent ROI or a prediction for an individual firm.

Released capacity has economic value only if the organization can use it: serve demand, improve quality, avoid spending or support an offer customers value. Lower cost can make a previously uneconomic segment feasible, but pricing, competition and demand determine whether that possibility becomes a viable business model. Test those assumptions rather than treating a faster task as realized savings.

What changes in the operating model

Agentic work changes handoffs and decision rights. Someone must own the process, its data, the rules governing access and the outcome. Define which work the agent may perform, which tools it may use, when it must stop and who receives an exception. Staff also need time and training to review evidence, challenge outputs and correct the system.

Map the full path from customer request to accepted result, including queues and human approval. If the bottleneck is an unavailable source or a policy decision, a better agent alone will not remove it. The OECD’s work on firm adoption emphasizes that capabilities and organizational conditions matter alongside the technology.

Boundaries, security and human control

Agents can produce plausible but incorrect outputs, use an unsuitable tool or encounter data that is incomplete or out of date. NIST describes agent hijacking: malicious instructions embedded in material the agent reads can redirect its behavior. Connecting an agent to email, records or payment systems therefore changes the consequences of an error; testing should include adversarial and unusual cases, not just average ones.

Apply least-privilege access, separate reading from writing, restrict destinations and spending, log actions and provide a way to pause or reverse them. Enforce authorization in the connected system, outside the model: an agent that summarizes email does not need permission to send or delete it. Separate proposal, authorization and execution, and bind approval to the exact action reviewed. Require a human approval at consequential or irreversible steps, such as changing bank details or committing funds; the reviewer must have context and real authority to reject. NIST’s generative AI profile is a voluntary risk-management resource, not a certification or a guarantee of safety.

A practical adoption path

First, choose a process with a named owner, measurable pain and enough examples to establish a baseline. Second, compare an agent with simpler alternatives, including eliminating steps or using ordinary rules. Third, evaluate historical and representative cases in a contained environment; include exceptions, adversarial inputs and a clear failure threshold. Next, run the agent in read-only or shadow mode, comparing proposed actions with actual decisions without changing live records.

Then introduce a supervised pilot with narrowly scoped access, explicit approval points and a documented rollback; only then consider bounded authorized execution. Finally, compare outcomes with the baseline and decide whether to improve, stop or expand. Increase permissions only after evidence supports the change and the business owner accepts the residual risk.

Measure outcomes, not agent activity

Track the share of cases completed correctly and accepted without rework; end-to-end turnaround; human review and correction time; cost per accepted case; exceptions and their severity; and customer or employee experience. Record data incidents, unauthorized action attempts and the time needed to detect and recover from errors. A high task count or low token price is not a business result.

Connect those operating measures to the original value hypothesis: revenue actually realized, expenditure avoided, capacity redeployed or risk reduced, with a credible counterfactual. Review quality and distribution of outcomes across case types. The governing question is whether the whole system delivers a better result at an acceptable risk and cost—not whether an agent can complete a demonstration.

Frequently asked questions

Does every company need an AI agent?

No. A simpler rule, a redesigned process or better data may be more reliable and economical. Compare alternatives against the same business outcome and risks.

Can an agent make decisions without a person?

An organization can grant limited authority for selected actions, but it should set permissions and escalation rules according to the consequences. People remain accountable, and consequential or irreversible actions should require appropriate review.

Sources and further reading

Let’s discuss your next decision.

Tell us what is at stake, what remains uncertain and where a clearer strategic direction would help.

Start a conversation