A Workforce of Agents — Agents at Work

The bet is not that AI will become a new kind of employee. The bet is that more work will be broken into goals, delegated to software that can plan and act across systems, and supervised through controls rather than direct clicks. The stakes are material: cost curves, speed of service, error rates, customer trust, and the shape of operations. The near-term question is not whether agents are impressive. It is where delegated digital work creates advantage, where it creates unmanaged risk, and how fast the operating model must change before competitors learn the same lessons.

The real decision is delegation, not adoption

The bet is that software can move from recommending the next step to carrying out a bounded piece of work. The stakes are highest where work is high-volume, rule-shaped but judgment-heavy, and spread across many systems: claims handling, fraud review, patient scheduling, supplier exception management, collections, customer onboarding, and compliance evidence gathering.

An agent is software that can interpret a goal, decide a sequence of steps, use tools, check progress, and escalate when it reaches a boundary. That makes it different from a chatbot and different from a workflow script. A chatbot answers. A script follows fixed instructions. An agent attempts to complete a task within an autonomy envelope, which is the agreed boundary for what it may do without human approval.

The strategic issue is that delegation changes accountability. If a fraud analyst accepts a model recommendation, the control point is visible. If an agent investigates a case, queries systems, drafts a decision, and triggers a customer communication, the control point moves into the design of the work itself. The decision to weigh is which work can safely become delegated work, and which must remain advisory until evidence improves.

Delegation moves software from answering questions to owning a bounded piece of work.

The first S-curve is in coordination cost

The bet is that agents will first matter where coordination is the constraint, not where labor cost is simply visible. Many operations are slowed less by the difficulty of any single task than by the handoffs between teams, systems, approvals, and evidence. Agents compress those handoffs when they can gather context, prepare options, and keep the work moving.

A hospital readmission model may identify risk, but the value is often lost in the steps that follow: checking discharge notes, arranging follow-up, confirming insurance, coordinating transport, and alerting the right care team. A factory downtime signal has similar friction: diagnostics, parts availability, maintenance scheduling, safety checks, and vendor contact. Agents are economically interesting when they reduce this connective tissue, not merely when they draft a better summary.

This means the early ROI will look uneven. Some agents will remove minutes from expensive decisions. Others will remove days from stalled processes. The second may matter more. Speed changes conversion, loss rates, working capital, customer churn, and compliance posture. The decision to watch is where cycle time, not headcount, is the hidden P&L lever.

📌

The hidden prize. The strongest early cases often come from reducing waiting time between systems, teams, and approvals.

Agents earn ROI where they compress the handoffs between teams, systems, and approvals.

The workforce metaphor is useful and dangerous

The bet is that treating agents as a workforce makes investment and governance easier to understand. The stakes are that the metaphor can also mislead. Agents do not have judgment, loyalty, context, or accountability in the human sense. They have capabilities, permissions, instructions, memory, evaluations, and logs. Managing them like people is a category error; managing them like ordinary software is also too narrow.

A useful comparison is that traditional enterprise software requires a person to drive the process, while an agent can take ownership of a bounded outcome. In a retailer, that may mean monitoring slow-moving inventory, proposing markdowns, checking margin rules, drafting supplier messages, and escalating only exceptions. In a bank, it may mean assembling a fraud case file, identifying missing evidence, and preparing a recommended action. The agent is not the accountable decision-maker; it is an operating unit inside a governed process.

This reframes workforce planning. The relevant questions are not only how many people are needed, but which decisions require scarce human judgment, which tasks become supervisory, and which processes should be redesigned rather than automated in place. The decision to weigh is whether agents are being used to patch old operating models or to create new ones.

Control becomes the product architecture

The bet is that control will determine which agent programs scale. The stakes are trust, regulatory exposure, customer harm, and operational resilience. As agents gain tool use, the ability to act through APIs and enterprise systems, the risk is no longer only a wrong answer. It is a wrong action, taken quickly, with plausible language around it.

Control has to be designed in layers. Identity determines what the agent is allowed to access. Permissions determine what it can change. Policy determines when approval is required. Evaluation determines whether it is fit for the work. Observability determines whether failures can be reconstructed. Without these layers, a pilot may look impressive while remaining unfit for production.

A public agency using agents to triage caseloads illustrates the point. Summarising applications is lower risk than changing eligibility status. Drafting a letter is lower risk than sending it. Prioritising a case queue may be acceptable if the criteria are auditable; deprioritising vulnerable applicants without review may not be. The decision to decide is the acceptable autonomy envelope for each class of action, not a single enterprise-wide yes or no.

⚠️

Autonomy is not binary. The practical choice is not agent or no agent; it is which actions require evidence, approval, reversibility, and review.

Control is designed in layers: identity, permissions, policy, evaluation, and observability.

Unit economics will separate pilots from platforms

The bet is that agent economics will improve as models become cheaper, orchestration improves, and reusable patterns emerge. The stakes are capital discipline. A demo can absorb expensive inference, manual review, and bespoke integration. A scaled workforce of agents cannot. The cost base includes model usage, data access, integration maintenance, evaluation, monitoring, exception handling, and the human time needed to supervise the work.

The important metric is cost per completed outcome at an acceptable quality level. For a bank, that may be cost per resolved fraud alert. For a retailer, cost per profitable markdown decision. For a hospital, cost per avoided readmission workflow completed. For a manufacturer, cost per downtime incident prevented or shortened. Comparing agents only to labor hours misses the broader equation; comparing them to the full cost of delay, errors, rework, and lost opportunity is more realistic.

There is also a scale trap. If every business unit builds separate agents with separate prompts, data paths, controls, and evaluations, early speed turns into long-term complexity. The economic advantage comes from shared foundations with local variation at the edge. The decision to weigh is where standardization creates leverage without freezing the business into one brittle operating model.

Cost per completed outcome, not cost per interaction
Quality adjusted for reversals, rework, complaints, and exceptions
Cycle-time impact where speed changes revenue, risk, or service levels
Reuse of controls, evaluations, and integrations across multiple workflows

A demo can absorb cost; a scaled workforce of agents must earn a low cost per completed outcome.

The competitive response will be operational, not theatrical

The bet is that competitive advantage will come from compounding small operational gains, not from announcing an agent strategy. The stakes are timing and learning velocity. Early movers may not win because they chose the best model; they may win because they accumulate process data, evaluation cases, exception patterns, and trust in where autonomy is safe.

The S-curve will likely be uneven. Low-risk knowledge work will spread quickly but may be easy to copy. More valuable agent work will sit closer to systems of record, regulated decisions, customer commitments, and supply chain execution. That work will move slower because controls matter, yet it may be more defensible once embedded.

Second-order effects will matter. If a logistics company reduces exception resolution time, customers may expect tighter delivery promises. If an insurer accelerates claims, fraud tactics may adapt. If a retailer automates markdown recommendations, suppliers may change negotiation behavior. Agents do not merely reduce internal cost; they can change market expectations and adversary behavior. The decision to watch is not only what agents can do, but how counterparties respond once the operating tempo changes.

The practical choice is where to place the first durable bets

The bet is that the first durable agent programs should sit where value, evidence, and control can meet. The stakes are avoiding two failures: scattering capital across disconnected experiments, or waiting so long that learning starts too late. The right starting points are rarely the most glamorous. They are workflows with clear outcomes, accessible data, measurable quality, tolerable downside, and enough volume to learn.

A strong candidate has a defined business owner, a repeatable decision or process, known escalation paths, and a way to compare agent-assisted outcomes against a baseline. A weak candidate has ambiguous authority, contested data, high reputational downside, and no clear measure of success. Customer-facing autonomy should usually lag internal preparation unless the action is reversible and the service benefit is obvious.

The portfolio should include three horizons: assistive agents that improve current work, delegated agents that complete bounded processes, and transformative agents that make a different operating model possible. The decision to make is how much capital and management attention to allocate to each horizon, and what evidence is required before moving from assistance to delegation.

✅

A good first bet. The best early workflow is valuable enough to matter, bounded enough to govern, and frequent enough to generate learning.

Key takeaways

Agents matter because they enable delegated digital work, not because they are better chatbots.
The earliest value often comes from reducing coordination cost and cycle time across fragmented processes.
Control is the scaling constraint: identity, permissions, policy, evaluation, and observability must be designed together.
Unit economics should be judged by cost per completed outcome at acceptable quality, not by model cost or labor substitution alone.
Competitive advantage will compound through process evidence, exception learning, and faster operating tempo.
The first durable bets should be bounded, measurable, reversible where possible, and connected to material business outcomes.

Glossary

Agent: Software that can interpret a goal, plan steps, use tools, check progress, and escalate when it reaches a boundary.
Autonomy envelope: The agreed boundary for what an agent may decide or do without human approval.
Tool use: An agent’s ability to act through approved systems, APIs, data sources, or applications rather than only producing text.
Orchestration: The coordination layer that manages how models, tools, data, policies, and approvals work together in an agent process.
Evaluation: The structured testing of agent performance against expected outcomes, failure cases, policies, and acceptable risk.