Choose workflow automation when the path can be specified, tested and maintained as rules. Use an AI agent when the route to a valid outcome changes with unstructured evidence, exceptions or context-sensitive judgement.
The strongest production design combines both. Deterministic code handles stable transport, validation and recorded actions. A model receives a bounded decision, approved evidence and explicit tools. Human approval protects consequences that remain expensive to reverse. This division gives the company flexibility where the work demands it without paying agent-level cost, latency and governance overhead across every step.
The practical difference is decision authority
A workflow can contain an LLM without becoming an agent. A fixed sequence that extracts fields from a document, validates the output and writes approved values to a system still follows a defined path. The model interprets content inside one step; it does not decide how the wider job should run.
OpenAI defines agents as systems where an LLM manages workflow execution, selects tools and determines when the task is complete. Its practical building guide recommends them for complex judgement, difficult-to-maintain rules and work dominated by unstructured data.
That authority creates the commercial distinction. A fixed workflow carries engineering and maintenance cost. An agent adds variable inference, tool use, evaluation, review and recovery. The added spend earns its place when adaptive judgement raises the value of accepted outcomes by more than the operating burden it introduces.
Stable paths favour deterministic automation
Use rules when the team can state the sequence and exceptions in advance. Reliable examples include moving an approved record between systems, checking required fields, calculating a known formula, sending a notification after a defined event or rejecting a duplicate operation ID.
Microsoft’s August 2026 documentation describes Copilot Studio agent flows as deterministic: the same input follows a rule-based path to the same output. These flows can still call models, connectors and other agents. Their defining property is controlled execution rather than the absence of AI.
Determinism protects several operating constraints:
- Testing: known branches produce expected states that can be asserted directly.
- Latency: the process avoids repeated planning and model calls where they add no value.
- Cost: Microsoft meters agent-flow capacity per action, while model-led loops can consume an uncertain number of steps.
- Recovery: the team can locate the failed branch and replay an idempotent action.
- Change control: a policy update maps to a named rule, owner and release.
A process does not become commercially stronger because a model chooses a route already known to the business. That choice replaces cheap certainty with paid variance.
Variable decisions create the case for an agent
An agent earns authority when the required path cannot be enumerated economically. The work may involve reading several documents, resolving ambiguity, requesting missing evidence, selecting a specialist tool and changing the plan after each result.
Anthropic separates predefined workflows from agents in its guidance on building effective agents. It recommends workflows for predictable, well-defined tasks and agents for open-ended problems where the number and order of steps cannot be hard-coded. Anthropic also highlights the trade: autonomy raises cost and allows errors to compound, which increases the need for sandbox testing and stopping conditions.
Three questions expose genuine variability:
- Does new evidence change the next step? A security questionnaire can require different sources and specialists according to the customer’s architecture and answers already supplied.
- Would a rules engine become a policy product of its own? Hundreds of brittle branches create maintenance pressure that model judgement can reduce.
- Can the business describe a valid outcome more clearly than a fixed route? Coding work has an acceptance state: tests pass and the reviewed change solves the issue. The files and edits required emerge during execution.
An unclear process is poor agent input. Variability should come from the case, not from unresolved ownership or contradictory policy.
Split the workflow at the judgement boundary
Treating the choice as “agent or automation” at whole-process level creates unnecessary complexity. Decompose the process into decisions, transitions and actions, then assign each part the cheapest execution mode that preserves quality.
Consider an inbound sales-enquiry workflow:
- A form submission creates a unique case ID and validates consent through deterministic code.
- Retrieval gathers the approved product, pricing and customer evidence available to that user.
- An agent interprets the enquiry, identifies missing context and drafts a qualification recommendation.
- Fixed rules block unsupported discounts, restricted claims and incomplete mandatory fields.
- A salesperson reviews consequential recommendations and accepts, edits or rejects the record.
- Deterministic actions write the approved state to the CRM and schedule the next event.
- Measurement joins the recommendation, review effort and eventual opportunity outcome.
The model handles ambiguity in the buyer’s language and context. Code protects state transitions. A person owns commercial judgement where the cost of a wrong promise exceeds the saving from unattended execution.
This architecture also produces cleaner evaluation. The team can test extraction, reasoning, policy checks, approval and write-back separately. When conversion or handling time changes, the trace shows which mechanism moved.
Use six tests before granting model authority
Score each proposed decision rather than labelling the department “agentic”.
| Test | Prefer deterministic automation | Consider bounded agent judgement |
|---|---|---|
| Path variability | Steps and branches remain stable | Evidence changes the required route |
| Input structure | Fields and formats are controlled | Meaning sits across documents or conversation |
| Acceptance | Exact output can be asserted | A rubric or expert judgement defines quality |
| Consequence | Actions are reversible and low-cost | High-impact action stays behind approval |
| Latency | Response must be immediate or fixed | Extra reasoning time is acceptable |
| Economics | Rules remain cheap to maintain | Adaptive handling improves accepted outcome value |
The table is a design filter rather than a maturity ladder. A deterministic component can be the more advanced choice because it concentrates model judgement where evaluation can support it.
AWS frames agents and automation as an autonomy spectrum in its executive decision guide. Its guidance favours the simplest solution that works and evaluates autonomy, task complexity, latency and full cost. Those dimensions belong at step level. One process can contain a rules engine, an LLM classifier, a planning agent and an approval gate without forcing every component into the same autonomy category.
Compare full cost per accepted outcome
A cheaper build can produce an expensive operating model. Agent loops add model calls, tool consumption, traces, evaluations and intervention paths. Deterministic systems accumulate maintenance cost when rules multiply and policy changes repeatedly break branches.
Use one denominator for both options: full cost per accepted workflow outcome. Include implementation, licences, model and tool usage, maintenance, active review, exceptions, rework and recovery. Verify the final outcome in the CRM, ERP, support platform or other operating system rather than relying on the agent’s completion message.
The AI agent unit economics guide explains how to connect execution cost, human effort and recovery to that verified state. The comparison becomes defensible when the same case segment and acceptance rule apply to both designs.
Run a shadow test before moving the boundary. Let the agent recommend a route while the existing process remains authoritative. Compare accepted recommendations, reviewer time, missed exceptions, latency and cost. Expand authority only for the segments where evidence shows a gain.
Keep human review attached to consequence
Human involvement has several jobs: supplying missing context, judging a recommendation, approving an action and investigating a failure. Collapsing them into one generic review queue hides where the system needs stronger design.
Place approval immediately before the expensive or irreversible transition. Show the reviewer the proposed action, evidence used, policy basis and consequence. Capture the decision and correction reason so repeated interventions can change a rule, source, tool or evaluation set.
The earlier Model Operator note on human review loops explains why review needs a named owner and structured decisions. The workflow boundary adds a further discipline: every repeated approval should justify its cost or trigger a design change.
Pilot the boundary, then decide
A useful pilot tests one decision boundary inside a live workflow. Define the eligible cases, existing baseline, approved evidence, model authority, prohibited actions and acceptance owner before development. The AI agent pilot success criteria guide provides the scorecard for a scale, revise or stop decision.
After representative cases have run, inspect four results:
- Did agent judgement improve acceptance or handling of cases that resisted rules?
- Which deterministic steps prevented variance from leaking into system state?
- Where did reviewers reconstruct context that the decision packet should have supplied?
- Did full cost per accepted outcome beat the baseline for the selected segment?
A failed agent comparison can still produce a valuable architecture. If rules win, the company has removed unnecessary inference and review. If the agent wins only for one exception class, it can serve that class while the remaining volume stays on the cheaper path.
Model Operator maps recurring workflows at this decision level: authoritative evidence, accepted company truth, model judgement, approvals, recorded actions and measurement. For teams deciding where agent autonomy earns its operating cost, start a build conversation or email alexander@modeloperator.io.