Your next AI agent should fix a broken workflow, not invent a new job

The invoice has arrived. The amount is wrong. Nobody knows why.

The purchasing system contains one figure. The supplier’s PDF contains another. Someone remembers approving an additional charge, but the approval is somewhere in an email thread. The employee responsible opens another tab.

Now imagine an AI agent doing the preliminary investigation: locating the agreement, comparing the charges, finding the relevant correspondence and preparing a recommendation.

This is an illustrative scenario, not a prediction of a fully autonomous finance department. Its appeal is more immediate. It addresses a recognisable problem with a potentially verifiable outcome.

It is also a better starting point for enterprise AI than asking which department could use a digital employee.

As 2026 draws to a close, organisations face two competing stories about agents. In one, they are already delivering returns. Anthropic’s 2026 State of AI Agents Report, based on insights from more than 500 technical leaders, reports that 80% of organisations in its research are seeing measurable ROI. That is a vendor-published survey finding, not an independently audited return across all enterprises.¹

In the other story, deployment is exposing weaknesses that demonstrations conceal. In May 2026, Gartner predicted that by 2027, 40% of enterprises would demote or decommission autonomous agents because of governance gaps discovered after production incidents. This is a forecast, not a measured failure rate.²

Those stories are not necessarily contradictory. An agent can produce useful work while the process around it remains poorly designed.

The opportunity for 2027 lies between enthusiasm and retreat: make agents responsible for a bounded contribution to a real workflow, and make people responsible for its outcome.

A good place to begin is the exception queue.

The work that begins when the normal process fails

An exception is something that falls outside the expected path.

A supplier invoice does not match an order. A delivery record is missing. An agency bills for a change nobody can trace. A report includes a claim without a reliable source.

These are not identical problems. They do, however, share a structure: something has arrived, it does not clearly meet expectations, and someone needs evidence before acting.

The useful agentic question is not simply whether a model can identify the discrepancy. It is whether a system can help move the case towards resolution.

An alert is not a resolution. Neither is a fluent explanation.

A reviewer still needs to know which record applies, whether the discrepancy is legitimate and who has authority to act. If an agent cannot make those questions easier to answer, it may have added another layer to the process rather than removed work.

This is why I would make resolution the organising principle for an enterprise agent programme.

Not conversations completed. Not agents deployed. Not documents processed.

Work brought to a correct, supported and authorised conclusion.

That sounds less spectacular than an autonomous organisation. It is also a more useful basis for an investment decision.

What has changed in 2026

The important change is not that organisations have suddenly discovered automation. It is that agents are being positioned to participate in longer sequences of work.

Anthropic’s 2026 report describes a move beyond single-step automation towards multi-stage workflows spanning teams. Its examples and survey findings suggest that the conversation is no longer confined to whether an assistant can produce a useful first draft.¹

Microsoft’s 2026 Work Trend Index offers a complementary perspective: the issue is how people integrate AI into their work and decide when to use it.

The research draws on a survey of 20,000 AI-using workers across 10 markets, alongside anonymised Microsoft 365 productivity signals. That population matters. Findings about people already using AI at work should not be presented as descriptions of every worker or organisation.³

Microsoft’s Australian account of the research describes a gap between employees’ adoption and organisations’ readiness to redesign work. It argues that as agents take on execution, people need to set direction, apply judgement and own outcomes.³

The resulting management problem is concrete.

Giving someone an assistant changes how they produce an answer. Giving a system access to tools and business processes changes what can happen after that answer.

The organisation must decide what the system can inspect, what it can recommend, what it can change and when it must stop.

Buying the capability does not settle those questions.

Do not give an agent a department

“An agent for procurement” is an ambition, not a specification.

So is “an agent for communications”. A department contains tasks with different evidence requirements, consequences and definitions of success.

Start with a sentence describing where work stalls:

When an invoice differs from the approved order, an employee must locate supporting records before deciding whether to accept or challenge it.

That sentence identifies a trigger, an investigation and a decision.

It also creates boundaries. The system does not need authority over every procurement activity. It needs a defined role in a particular case.

For a first implementation, that role might be to retrieve documents, identify conflicting information and prepare a review pack. The person responsible retains the decision.

This approach also helps answer an uncomfortable question: do you need an agent at all?

If the task is a straightforward comparison between structured fields, existing software or a rule may be sufficient. If the sequence is predictable, a fixed workflow may be preferable. If the real problem is that nobody maintains the approved records, the next purchase should not be expected to compensate for that neglect.

The agent should earn its place by doing something that a simpler approach does not do adequately.

Its name, interface and apparent initiative are not evidence of value.

The most useful architecture may be deliberately modest

It is tempting to treat autonomy as a ladder: observe, recommend, act, then operate without intervention.

But moving upwards is not automatically progress. More authority is useful only when it improves the process enough to justify the additional risk.

Gartner’s May 2026 guidance distinguishes four levels: observe, advise, act with approval and act autonomously. Each crosses a different trust boundary and requires different controls.²

For an exception-handling workflow, those levels could look like this:

This is an illustrative application of the framework, not a claim that every process should reach the final level.

A read-only investigation assistant may be the right long-term design for a sensitive workflow. A tightly bounded autonomous action may be appropriate elsewhere.

The mistake is to grant broad access because the system performs well on a narrow task.

Gartner identifies precisely this mismatch between an agent’s ability to act and the scope of access it receives as a source of enterprise failure.²

For 2027, a useful design goal is not maximum autonomy. It is appropriate authority.

Four practical possibilities for 2027

The following are proposed applications, not verified deployments or promised returns. They illustrate where the exception-handling pattern could be tested.

1. Supplier-invoice investigation

The agent retrieves the invoice, order, receipt and applicable agreement. It identifies the disputed item and distinguishes confirmed differences from missing information.

Its output is a case for review, not an instruction to withhold payment.

A useful system might say: “The invoice includes an additional service fee. I found no approval in the records available to me.”

It should not silently turn that into: “The fee was never approved.”

2. Agency and professional-services billing

The agent compares an invoice with the statement of work, milestones, agreed rates and recorded changes.

The purpose is not to generate disputes. It is to establish whether the charge has adequate support.

A well-scoped pilot could cover one supplier category and a small set of recurring discrepancies, with a named contract owner reviewing the findings.

3. Editorial and communications assurance

The agent checks a draft against an approved brief and source material. It flags unsupported factual claims, missing approvals and quotations that cannot be traced.

The system should help an editor inspect the work. It should not confuse the existence of a citation with proof that the source supports the sentence.

Editorial quality, strategic relevance and tone remain broader judgements.

4. Service-case preparation

The agent assembles the relevant history, policy and outstanding information before a person resolves a complicated request.

Its contribution is a better hand-off: fewer missing facts, clearer uncertainty and a visible decision path.

In each example, the useful question is the same: has the agent reduced the effort needed to reach a defensible outcome?

That is more demanding than asking whether it produced something plausible.

A source of truth needs an owner

Connecting an agent to documents does not make those documents authoritative.

The latest file may be unsigned. The signed agreement may have an amendment. An approval may apply to one transaction rather than the entire contract.

Before a system can reconcile conflicting information, the organisation needs rules for deciding what counts.

For a pilot, I would require explicit answers to five questions:

• Which records may the system use?
• How does it identify an approved version?
• What happens when sources conflict?
• Which missing evidence prevents a conclusion?
• Who can resolve an ambiguity?

These are proposed readiness requirements, not claims about what every existing deployment does.

They are also revealing. If nobody can answer them, the immediate bottleneck may be document ownership rather than model capability.

That is useful to discover before granting an agent more access.

There is no shame in learning that the first phase of an AI project should be a better records process. The embarrassment comes from scaling a system that conceals the problem beneath confident prose.

“Human in the loop” is not a complete operating model

A human approval step can look reassuring on a diagram.

In practice, the reviewer may receive a vague recommendation, inaccessible sources and a deadline. Clicking approve becomes the easiest way to keep the queue moving.

Gartner’s 2026 guidance warns that approval fatigue and time pressure can weaken human review. It recommends clear approval workflows, audit trails, security testing and agent-specific incident response when systems can act after approval.²

That warning should shape the interface.

A reviewer needs the disputed detail, relevant evidence, proposed action and known uncertainty. They also need enough authority to reject the recommendation or request further investigation.

Responsibility must be specific.

The person checking the evidence is not necessarily the person authorised to change a financial record. The process owner is not necessarily the technical owner. The employee affected by an incorrect action needs a route to challenge it.

An agentic workflow should make these responsibilities visible, not collapse them into an anonymous human checkpoint.

For organisations planning 2027 deployments, this is where communication becomes part of control design.

People need to understand what changed, what did not and what remains theirs to decide.

The enablement programme has to change too

A short course on prompting will not prepare someone to supervise a system with operational permissions.

They need to understand the difference between a retrieved fact, an interpretation and a proposed action.

They need examples of appropriate reliance: when a recommendation is usable, when it requires checking and when missing evidence should stop the process.

They need to know how to report an error without inventing a workaround.

Microsoft’s 2026 research places intentional delegation and human judgement at the centre of effective AI use. Its Canadian account highlights the practice of pausing before a task to decide what belongs with AI and what belongs with a person.⁴

An enterprise can turn that habit into a practical briefing.

For each workflow, explain:

• What the system is meant to accomplish.
• Which sources and tools it can access.
• Which actions it may take.
• What users must verify.
• When it must escalate.
• How a mistake is corrected.

This is less glamorous than announcing a new digital workforce. It is more likely to help the actual workforce use the system responsibly.

It also creates a feedback mechanism. Reviewers should be able to identify recurring failures, not merely fix individual outputs.

The organisation needs to learn from the queue, not just empty it.

Measure the work that remains

The easiest metrics are often the least informative.

Documents processed. Agent runs completed. Recommendations generated.

These show activity. They do not establish improvement.

For a pilot, measure the whole case, including the work people still perform:

This is a proposed evaluation framework, not a summary of the survey methods cited earlier.

The distinction between potential and confirmed benefit is particularly important. A disputed charge is not recovered money. A suggested correction is not a completed resolution.

Similarly, a faster first draft is not necessarily a faster approval process.

Anthropic’s reported ROI findings make agents worth investigating. They do not remove the need to establish the economics of your own workflow.¹

Nor should Gartner’s forecasts become an excuse to avoid experimentation. They identify risks to manage, not an inevitable outcome.

The point is to make the investment falsifiable: establish what would count as success, and what would persuade you not to scale.

What 2027 could make possible

There is a more interesting possibility than adding a separate agent to every task.

An organisation could build a governed exception service: a consistent way to collect evidence, prepare cases, route decisions and learn from recurring problems.

Different workflows would still require different permissions and expertise. But they could share standards for source visibility, ownership, escalation and measurement.

That is a proposed direction, not a forecast of universal adoption.

The value would come from improving the connections between tasks rather than merely accelerating each task in isolation.

A purchasing discrepancy might reveal a recurring contract ambiguity. Repeated missing approvals might expose a broken hand-off. Editorial flags might show that teams need better source records before drafting begins.

In that model, agents help surface process weaknesses. People decide whether to fix the underlying system.

A second possibility is more selective autonomy. Once a narrowly defined category of cases has demonstrated reliable performance, an organisation could permit limited action within explicit boundaries, while continuing to route ambiguous cases to people.

Gartner’s framework supports this differentiated approach. It recommends stronger monitoring, rollback mechanisms, circuit breakers and clear ownership for autonomous agents—not the same controls applied indiscriminately to every system.²

The ambition for 2027 should therefore be better judgement about where autonomy belongs.

Not autonomy everywhere.

Begin with one queue

The practical next step is deliberately small.

Choose one process with recurring exceptions, accessible evidence and a named decision-maker. Test historical cases. Include contradictory records and legitimate changes that initially resemble mistakes.

Then run the system alongside the existing process without allowing consequential actions.

Compare outcomes. Inspect disagreements. Count human effort.

If a simpler workflow performs as well, use it. If the records are inadequate, repair them. If the agent improves investigation but should not act, keep that boundary.

As 2026 closes, enterprises do not need another reason to be excited about agents. They need a better basis for deciding what to entrust to them.

The exception queue offers one.

Find the work that repeatedly stalls. Establish the evidence required to move it forward. Give the system a bounded role and the people around it clear responsibilities.

Then judge the result by whether the problem was resolved—not whether the organisation looked more futuristic while handling it.

Footnotes and research

1. Anthropic, The 2026 State of AI Agents Report. Based on insights from more than 500 technical leaders. The publisher reports measurable ROI among 80% of organisations represented in the research; this should not be read as a population-wide, independently audited ROI result.
2. Gartner, Gartner Says Applying Uniform Governance Across AI Agents Will Lead to Enterprise AI Agent Failure, 26 May 2026. Introduces differentiated governance for observe, advise, approval-based action and autonomous action. Its statement about 40% of enterprises demoting or decommissioning autonomous agents by 2027 is a prediction, not an observed failure rate.
3. Microsoft, Aussie Workers Charge Ahead with AI, While Leadership Falls Behind, 16 June 2026. A regional account of the 2026 Work Trend Index, describing the global research base and the gap between individual adoption and organisational workflow redesign. Regional findings should not be substituted for global estimates.
4. Microsoft, AI, human agency, and the opportunity for every Canadian organization, 28 May 2026. Discusses intentional delegation, human oversight and responsibility within the 2026 Work Trend Index. Its regional framing should be retained when using specific findings.