Choose how the work runs
AI agents vs automation: when does a workflow need judgement?
Automation follows rules that move work through agreed steps. AI agents add model-selected choices about what to do next. Many useful systems combine fixed rules with a model for a bounded task. Decide which parts of the route are predictable and which genuinely need flexible investigation.
Keep predictable steps explicit, use AI for a defined interpretation task and add agent choices where selecting the next step helps complete the job.
Compare three designs for the same work
| Design | Who determines the route? | Reporting example |
|---|---|---|
| Fixed automation | Software follows explicit rules. | Collect the agreed export, check its period and route the report for review. |
| AI-assisted workflow | The route stays fixed; a model performs a defined step. | Draft observations from the approved figures, then pass them to a reviewer. |
| Agent behaviour | A model chooses permitted next steps based on what it finds. | Select which authorised source to inspect next while investigating a change. |
These designs can coexist in one system. A fixed permission check can guard a model-selected tool call. A person can approve the resulting recommendation. Anthropic’s agent design explanation provides the technical distinction behind that comparison.
This is a choice about how the job runs. The separate agent-versus-chatbot guide explains how that behaviour can sit behind a conversation interface.
Keep predictable steps understandable
If a complete approved request always creates the same task, use an explicit rule. If an unapproved agreement always needs an account-owner decision, record that condition. These rules are easier to inspect when they stay visible in software.
Said describes our current build as requiring reviewable evidence through deterministic paths. That does not mean every part of the work is fixed. It means an agent’s selected route still needs checks that carry the supporting information into a human decision.
The onboarding sample shows a job that can run through fixed checks and actions. It has useful automation without needing a model to decide whether an approval exists.
Add choices when the next step is genuinely variable
An investigation may discover that a metric definition changed, a source is incomplete or another period is needed. A model can select a permitted next check based on those findings. Write down the choices and evidence required for each.
Bound time, tool calls and authority. Decide what happens when the source is unavailable or results conflict. A missing input should not send the agent into an unlimited search for an answer the system cannot substantiate.
Read-only investigation and changing a client system require different authority. Begin by inspecting recommendations and their traces. Add actions only when the approval and recovery rules have been tested.
Compare accepted results and upkeep
Give each design the same normal job and difficult variation. Measure completion, evidence accuracy, review effort and running usage. Include the work needed to maintain rules, sources and prompts.
A more flexible route may handle a useful variation, or it may add cost without improving the result. Let the observed job decide. Keep the evaluation record so changes can be compared against the same expectations.
Worked example
Three reporting designs
Invented teaching example. No client outcome is claimed.
FIXED
Collect approved monthly inputs and route them to a reviewer.
AI-ASSISTED
Add a bounded draft-observation step with source evidence.
AGENT
Permit selection of additional sources during a defined investigation.
BOUNDARY
All three preserve client permissions and human approval of the report.
The designs are invented; no comparative performance result is claimed.
Further reading
Written by Said Lopez, founder of LeverageAI and author of Lean AI. Our guides draw on the method in the book and our work with clients.