The agent market will narrow before it scales
Over the next two years, the most valuable AI agents will not be general-purpose digital employees. They will be bounded systems that can complete a defined workflow across a small set of tools, under explicit permissions and with measurable outcomes. A procurement agent might compare approved suppliers, draft a purchase order and route exceptions to finance. A service agent might classify a complaint, retrieve account history, offer one of five authorised remedies and update the customer record. This is narrower than the vision of autonomous colleagues, but far more useful because businesses can specify where the agent starts, what it may change and how success is judged.
The shift reflects an uncomfortable engineering reality: reliability tends to fall as task length, tool count and environmental uncertainty rise. An agent that performs each step correctly 95 per cent of the time has only a 60 per cent chance of completing a ten-step process without error if those probabilities compound. Better models and orchestration will improve those figures, but leaders should resist treating benchmark gains as permission for unlimited autonomy. The winning deployments will compress workflows, add deterministic checks and escalate ambiguity rather than asking a model to improvise indefinitely.
This will also change investment priorities. Instead of funding a single enterprise-wide ‘agent platform’ and searching for uses, organisations should build a portfolio of constrained agents tied to high-volume processes. Claims triage, invoice matching, sales research, compliance evidence collection and IT access requests are credible candidates because inputs, policies and downstream systems are identifiable. The trade-off is less theatrical autonomy in exchange for lower operational risk, faster deployment and clearer returns. In practice, a system that resolves 40 per cent of routine cases safely may create more value than one that attempts 100 per cent and fails unpredictably.
Tool ecosystems will become the real competitive layer
Models will remain important, but the decisive advantage will increasingly sit around them: connectors, permissioning, identity, workflow state and reliable interfaces to business software. An agent cannot complete useful work merely by generating a persuasive answer. It must locate the correct customer, query current inventory, apply the right discount rule, write an auditable update and avoid exposing data to an unauthorised user. That requires robust tools and contracts between systems, not just better prompting.
Software vendors will therefore redesign products for machine users as well as human users. APIs will need clearer schemas, idempotent actions, granular scopes and explicit error messages. Tool catalogues will need ownership, version control and retirement policies. Emerging interoperability standards can reduce integration effort, but standardisation will not eliminate the hard work of deciding which actions an agent may take. A connector that technically permits ‘update account’ is insufficient if the business needs separate rights for changing an address, altering credit terms and closing an account.
Leaders should treat the agent tool layer as shared infrastructure. A governed capability for retrieving a contract, checking a customer’s entitlement or creating a support ticket can be reused across dozens of workflows. Without that discipline, departments will build duplicate connectors with inconsistent permissions and hidden maintenance costs. The strategic question will move from ‘Which model are we using?’ to ‘Which verified actions can our agents perform, on whose authority, and with what evidence?’ Organisations that answer it well will switch models more easily while retaining their operational advantage.
Evaluation will move from demos to continuous assurance
Agent demonstrations are unusually flattering. The path is selected, the data is clean and a human is ready to steer the system away from trouble. Production exposes the opposite conditions: incomplete records, conflicting policies, unavailable tools, prompt injection, stale knowledge and users who phrase requests in unexpected ways. Over the next two years, serious adopters will build evaluation systems that resemble software testing, risk management and operational analytics combined.
A useful evaluation programme will include offline task suites, adversarial tests, simulated tool failures and live monitoring. Measures should go beyond answer accuracy. Businesses need to track task completion, unnecessary actions, escalation quality, policy compliance, latency, cost and the reversibility of errors. For a billing agent, a 90 per cent completion rate is meaningless if 3 per cent of completed cases contain an incorrect credit. For an internal research agent, modest factual error may be tolerable if every claim is cited and employees review the output before publication.
The strongest teams will maintain ‘golden’ workflow sets and replay them whenever a model, prompt, policy or connector changes. They will also sample real traces, label failures and feed those cases back into testing. This creates a practical release gate: a new version ships only if it improves target outcomes without breaching risk thresholds. The cost is material; evaluation requires domain experts, representative data and ongoing maintenance. Yet it is cheaper than discovering through customers or regulators that an apparently capable agent behaves differently after a routine model update.
Governance will become embedded in execution
Agent governance will shift from policy documents to controls enforced at runtime. Traditional AI guidance often tells employees not to enter sensitive data or to verify generated content. That approach is inadequate when software can initiate payments, modify records or communicate externally. Controls must sit between the model and the action: identity checks, least-privilege access, transaction limits, approval requirements, data-loss prevention and immutable logs.
A practical autonomy ladder can make governance concrete. Level one agents retrieve and draft; level two agents act after human approval; level three agents execute low-risk actions within predefined limits; level four agents manage broader processes with sampled oversight. Different workflows should occupy different levels. An agent may autonomously reschedule a delivery within a two-day window, while refunds above £250 require approval and changes to bank details remain human-only. These thresholds should reflect potential loss, reversibility, customer impact and regulatory obligations rather than enthusiasm for the technology.
Governance will also become more dynamic. Permissions may depend on the user, jurisdiction, data classification and current risk signals. A sales agent could access public company information for any prospect but retrieve contract pricing only for accounts owned by its requesting employee. Security teams will need to monitor tool calls and unusual action sequences, not merely model inputs and outputs. Boards, meanwhile, should demand an inventory of production agents, named accountable owners and evidence of testing. The objective is not to prevent autonomy; it is to make autonomy legible, bounded and revocable.
Human roles will be redesigned around exceptions
The largest productivity gains will come from changing operating models, not adding an agent to an unchanged process. If an agent drafts every claim assessment but employees must copy each result between systems and recheck every field, the organisation has created more supervision rather than less work. Processes must be divided deliberately between machine-speed handling of routine cases and human judgement for ambiguity, empathy, negotiation or high-impact decisions.
This redesign will create exception-driven teams. Instead of processing 200 similar items each day, an employee might review 30 cases flagged for missing evidence, policy conflict or unusual customer circumstances. That can improve job quality, but it can also increase cognitive load because the remaining work is consistently difficult. Managers will need to adjust staffing ratios, queue design and performance measures. Average handling time may rise even as total labour per transaction falls, because humans see only the complex tail. Measuring employees against old throughput targets would punish precisely the behaviour the new system requires.
New roles will emerge around agent operations: workflow owners, tool administrators, evaluation leads and domain specialists responsible for failure taxonomies. These need not become a large new bureaucracy. In many firms, existing product, operations, risk and engineering staff will take on the responsibilities through cross-functional teams. What matters is clear accountability. Every production agent should have someone who owns the business outcome, someone who owns technical reliability and someone empowered to stop the system when risk exceeds tolerance.
Economics will favour orchestration over maximum intelligence
As model capabilities converge, businesses will optimise agents for total workflow economics rather than raw intelligence. The most capable model may cost more, respond more slowly and still be unnecessary for routine classification or extraction. A well-designed system will route simple tasks to smaller models, invoke premium reasoning only for difficult cases and use deterministic software wherever rules are sufficient. It may also cache stable results, batch non-urgent work and impose budgets on repeated tool calls.
The relevant calculation is cost per successfully completed, policy-compliant task. Suppose an agent attempt costs £0.20 but succeeds safely only 60 per cent of the time and creates £1 of review effort for failures. Its headline inference cost obscures the true economics. Conversely, a £1 attempt that succeeds 95 per cent of the time may be cheaper for a process previously costing £8 in labour. Leaders should include integration, evaluation, oversight, incident handling and vendor dependence in the business case, not merely token prices.
This discipline will temper extravagant claims while accelerating credible deployments. Early projects should establish baseline volumes, handling times, error rates and downstream costs before automation begins. Benefits can then be attributed to fewer touches, faster cycle times, improved conversion or avoided loss. Some agents will not reduce headcount; they will absorb growth, extend service hours or let specialists focus on higher-value work. Those outcomes are legitimate, but they must be named and measured rather than translated into vague promises of productivity.
Leaders should build optionality now
The next two years will reward organisations that make reversible bets. Model performance, pricing and vendor strategies will change quickly, while core business controls and data structures move slowly. Architectures should therefore separate models from tools, policies, memory and workflow state wherever practical. Contracts should address data retention, service continuity, model substitution and access to operational logs. Sensitive use cases should have fallback procedures for model outages or sudden quality regressions.
A sensible 90-day agenda is concrete. Select three workflows with meaningful volume and bounded risk; document their current cost and failure modes; define autonomy levels; build a representative evaluation set; and expose only the minimum tools required. Run agents in observation or recommendation mode before allowing actions, then expand authority when evidence supports it. At the same time, establish a central register of agents and reusable tools so experimentation does not become invisible production infrastructure.
The strategic opportunity is substantial, but it will not be captured by waiting for a universally reliable autonomous worker. It will come from hundreds of carefully designed transfers of work: a search completed automatically, an exception identified earlier, a record updated without copying, a decision assembled with better evidence. Narrow autonomy, strong tools, continuous evaluation, executable governance and redesigned teams are not constraints on the agent era. They are the conditions that will allow it to enter everyday business systems and stay there.
Comments (0)
Discussion is opening soon. Be the first to comment.