Workflow, Agent, or Both? Choosing How Work Gets Done

Define the outcome before choosing the mechanism

Start with whose result matters, which cases are eligible and who can accept the work. Identify the information and permissions the operation needs. Specify what happens when those conditions are absent. These choices establish a fair comparison between designs.

Before adding a mechanism, examine the reason for the difficulty. An unclear policy needs clarification. Duplicate requests may be removable. A missing record may need a better source or a dependable handoff. Existing automation may already support the necessary sequence. Giving an agent more freedom will not establish which policy is authoritative or fund an understaffed review team. Redesigning Work Around Outcomes provides the broader method; Modernize the Automation You Already Have examines the existing estate.

Hypothetical operating scenario

The following design is illustrative, not a reported implementation or result. An insurer wants to prepare claim evidence for substantive review. The eligible work is a defined class of claims, with an agreed evidence checklist and access to approved records. Claims outside that class follow the existing specialist route.

The outcome is an accurate, source-linked evidence file accepted as ready for review by a designated claims specialist, acting for the Outcome owner. The claims operations head holds that accountability and sets the acceptance criteria and resourcing commitment. Preparing the file does not decide coverage or settle the claim.

• The file matches the correct claim and contains the required records for the defined case class.
• Material facts can be traced to their sources, and relevant policy-record versions have been confirmed.
• Material omissions and record conflicts are resolved; a flagged but unresolved file remains incomplete.
• The receiving specialist can access the file, checks it against the criteria and records acceptance.

These criteria apply to every mechanism. An exception may change the route and the people involved; it cannot quietly lower the standard for completion. The intended benefit must also survive this handoff: a faster preparation step has limited value if it leaves the specialist with more checking or correction.

Separate interpretation from control of the sequence

A rules-based workflow follows predefined steps and routing conditions. Its branches can be extensive, but the permitted sequence is specified in advance.

AI assistance introduces interpretation, extraction, drafting or recommendations within work directed by a person or a predefined workflow. A model may summarize a document while the workflow determines exactly what happens next.

With one bounded agent, the system has discretion over specified steps or tool choices. Findings can influence which permitted action it takes next. The assignment, sources, actions and stopping conditions still constrain that choice.

Multiple agents give distinct responsibilities to more than one agent. Their contributions must be coordinated and combined into one accepted result. Several predetermined model calls are still a workflow if none has discretion over how its assigned work proceeds.

These are operating alternatives. Engineering guidance distinguishes predefined workflow paths from agents directing their process and describes useful combinations of the two. It also cautions that added flexibility can increase cost and latency.^[1]

Compare the four arrangements

The comparison below holds the outcome and acceptance criteria constant. It identifies the operating work each choice introduces; it does not assign a score or rank the alternatives.

Match discretion to the case

Routine records favor a rules-based workflow

A routine claim arrives with structured information, valid record identifiers and the required documents. The checklist and retrieval route are known. A workflow checks identity and completeness, retrieves the specified records and routes the file for acceptance.

Here, a rules-based workflow is preferable: there is little useful discretion to delegate. Its defined path makes expected behavior easier to inspect and failures easier to locate. People still maintain the checklist, source mappings and exception route. If those inputs deteriorate, the workflow needs attention; predictable execution cannot compensate for incorrect rules or records.

A known interpretation task needs assistance

Another claim contains a narrative inspection report. The required task is to extract dates and reported damage into the same evidence file. The sources, checks and next step remain known.

AI assistance is sufficient for this design. A prescribed model step can draft the extraction with source references; a person checks material facts before acceptance. The workflow directs the sequence. Giving an agent freedom to choose further searches would need a separate reason.

The benefit depends on review effort. If checking and correcting the extraction takes as much work as preparing it manually, the model has moved effort rather than released useful capacity. The comparison should include an improved manual route and conventional document processing where they are plausible.

An unfolding search can justify one bounded agent

In the variable case, correspondence refers to an earlier inspection, which points to a subsequent amendment. The useful next lookup depends on the previous finding. Writing a branch for every plausible sequence may create substantial maintenance work.

A bounded agent is justified as a candidate because its discretion addresses this changing sequence. It can select searches within approved repositories, compare retrieved records and assemble a draft with source references. It has no permission to determine which conflicting policy applies, seek unrelated personal information or keep searching indefinitely.

The Operating owner sets the search and effort limits before use, along with conditions for escalation. Within those bounds, the agent must show that adaptive retrieval improves accepted files or reduces total human effort compared with an assisted workflow. If the same sequence recurs reliably, it may be better captured as a predefined route.

Several agents need an additional operating case

A team could assign record collection, chronology and evidence checking to separate agents. That may help when the tasks can proceed independently and their contributions have clear interfaces. It also creates work: maintaining a common case and source version, reconciling disagreements, avoiding duplicate searches and deciding when the combined file is ready.

An agent checking another agent is not automatically an independent check. Shared missing records or mistaken assumptions can affect both. Completion still requires evidence against the common acceptance criteria.

A 2026 controlled benchmark study held prompts, tools and compute budgets constant while comparing coordination structures. The benefit of collaboration depended on the tested conditions and the strength of the single-agent baseline. Its findings support testing the additional coordination; they do not establish a universal selection threshold or savings in a production operation. The study also notes limits in benchmark coverage and prompt optimization.^[2]

For this hypothetical assignment, I would first compare the workflow, assistance and single-agent designs. Several agents deserve inclusion only if useful separation or parallel work has a credible benefit that one agent or a predefined sequence cannot provide economically.

Combine mechanisms while keeping authority explicit

The operation need not adopt one mechanism for every claim. A workflow can check eligibility, route routine files through known checks and assign variable evidence assembly to a bounded agent. Both routes return to the same verification and receiving specialist. AI assistance can also sit within a fixed step on either route.

This arrangement keeps responsibility for the service with people. The Outcome owner remains accountable for the result and resources. The Operating owner runs the queues, exceptions and continuity arrangements. The Knowledge owner maintains the authority and applicability of the sources. The Technology owner supports availability and recovery; the Control lead checks that the intended limits are enforced.

These are the authority arrangements for this hypothetical design. A technically available action is not necessarily permitted, and adding agents does not confer wider authority. Bounded Autonomy explains the delegation boundaries; Before an AI Agent Acts: Where Authority Is Enforced examines how those permissions hold at execution.

The boundary also affects reversibility. A draft evidence entry can be corrected with its history retained. Sending a customer message, changing a claim decision or authorizing payment has different consequences. The selection decision must account for the checks and receiving authority those actions require, even if a system can technically perform them.

Follow a difficult case through recovery

During evidence assembly, the agent finds an endorsement in correspondence that conflicts with the effective version shown in the policy record. It must not choose whichever version makes the file appear complete.

The workflow places the case on hold and withholds ready-for-review status. It preserves both source references, their versions, the conflict description and the checks already attempted. The designated claims operations lead, acting as Operating owner, receives the case and responsibility for keeping it visible. A named backup and agreed response arrangement cover absence; an unattended alert is not a handoff.

The lead asks the policy-records specialist to resolve the source conflict. The Knowledge owner coordinates any required source correction. Resolution must identify the authoritative version and its applicability, with evidence from the authorized source or an authorized correction. The agent's confidence in its own interpretation is insufficient.

Once that evidence is recorded, the operations lead releases the affected preparation work to resume. The summary is rebuilt or corrected, the affected checks are rerun, and the receiving claims specialist decides whether the file meets the original acceptance criteria. The agents cannot approve their own resumption or acceptance.

If the necessary record remains unavailable, the case stays owned and incomplete while the lead arranges specialist investigation or a staffed manual route. Exhausted search limits trigger that route rather than another open-ended attempt.

If a save times out, check the actual case record before retrying or declaring success. If its state remains uncertain, the Operating owner holds the case while the Technology owner investigates. Acceptance requires evidence that the intended file is available to the receiving specialist. Agent-evaluation guidance similarly distinguishes what an agent says from the resulting state of the environment.^[3]

This is also a staffing question. NIST's generative AI guidance recommends evaluating human–AI configurations and involving practitioners and operators in testing difficult scenarios.^[4] For this operation, that means testing whether the receiving person has the evidence, time and authority to resolve the case. Detailed transaction and multi-agent coordination designs follow from that requirement; they cannot be assumed away by selecting an agent.

Compare the recurring cost of accepted work

Start with a baseline for the current operation, including its manual preparation and exception work. Compare improved rules and information alongside AI options. If correcting a source removes repeated searches, the benefit belongs to that correction rather than automatically to the agent using it.

Use the same unit of work and the same acceptance standard across the alternatives:

Cost per accepted outcome = Total recurring cost of running the capability in a period ÷ outcomes accepted by the owner in that period.

Recurring cost includes technology and usage, human preparation, review and correction, failed and repeated attempts, exceptions, support, recovery, and ongoing rule and knowledge maintenance. Multiple-agent designs also incur coordination and reconciliation costs. Count a file once when the authorized specialist accepts it for the Outcome owner. Repeated attempts do not create additional accepted outcomes.

Keep one-time design and build cost separate in the investment case. Match scope, case mix, period, acceptance criteria and follow-up. Declare the allocation of shared support costs and account for the lag between preparation and acceptance; do not combine unmatched costs and completions. With zero accepted outcomes, Cost per accepted outcome is not calculable.

Compare routine and variable cases separately before combining them in a service view. A workflow handling mostly routine files and an agent handling difficult files do not provide a fair headline cost comparison. Record missing information and rejected or reopened files so exclusions cannot make an option appear successful.

The measure also needs quality and service evidence. Track time to acceptance, active human effort, correction, unresolved cases and the receiving team's workload. Faster throughput can matter, but a lower cost is not sufficient if the operation misses required evidence or exceeds permitted authority.

Released staff capacity can support other valuable work. It becomes realized cash savings only when expenditure actually falls. Where the case depends on redeployment, name the work that will use the capacity and who will make that change. From Idea to Operating Case: Building the Business Case for Automation-to-Autonomy develops that investment distinction.