Define the work before choosing the technology.
If the same input should produce the same approved action every time, use clear rules, standard automation and a defined owner. When the input genuinely requires interpretation, AI may assist inside a bounded step.
The goal is not AI adoption. The goal is a reliable operating result the client can understand, test and own.
Map what actually happens
Begin with the real trigger, inputs, decisions, handoffs, exceptions and completion condition. The process people describe in a meeting is often cleaner than the one the company actually operates.
Observe representative work. Compare the documented path with messages, spreadsheets, system history and manual interventions. If the trigger or definition of done changes by person, the first job is workflow design.
RAND's research on failed AI projects found that teams often begin with the technology instead of the end user's real problem. The order matters: define the work first, then choose the mechanism.
Choose rules, bounded AI or human authority
| Work should happen the same way every time | Use rules, scripts, integrations, validations and standard automation. |
|---|---|
| Input is unstructured and requires interpretation | Use bounded AI for research, extraction, classification, summarization or drafting. |
| Action affects customers, money, people or access | Keep final authority with a named person. AI may support the decision. |
| The process cannot be clearly explained | Map and repair the process before automating it. |
OpenAI's practical guide to agents recommends agents for work involving complex decisions, difficult rules or unstructured data, and says a deterministic solution may be enough when those conditions are absent. Microsoft's workflow guidance makes the same architecture distinction: workflows provide structure while agents handle ambiguity.
Make authority visible
Every consequential decision needs an accountable owner, an evidence standard and an escalation path. A model should not quietly inherit authority that no person was willing to define.
- What may the system execute without review?
- What may it recommend but never execute?
- Which data may it access, create or change?
- Who approves exceptions and high-consequence actions?
- How can a person stop the workflow and restore control?
Use the free Automation Authority Matrix to record these boundaries for each workflow decision. OWASP describes excessive agency as risk created by excessive functionality, permissions or autonomy. Limiting authority is part of the operating design, not a final security add-on.
Define evidence, data and review boundaries
List the evidence required for each decision, where it comes from, how fresh it must be and what happens when sources disagree. Separate public information, internal company data, customer information and regulated or highly sensitive records.
A workflow is not ready because a model produced a convincing example. It is ready when the team can test representative inputs, expected outputs, prohibited actions and review conditions across normal, unusual and adversarial cases.
The NIST AI Risk Management Framework treats testing, evaluation, verification and ongoing monitoring as part of responsible AI risk management.
Design exception, observability and rollback paths
Good automation does not hide uncertainty. It preserves context, routes exceptions and makes failure visible. Before launch, define:
- evidence thresholds that force human review;
- timeouts, duplicate detection and retry limits;
- logs showing what happened and what rule or model contributed;
- an owner for failed or disputed outcomes;
- a kill switch and a tested rollback procedure.
The exception path is part of the product. Treating it as an edge case is how a small operating ambiguity becomes a scaled failure.
Price the full lifecycle
Compare the current cost and error rate with build, inference, review, monitoring, evaluation, maintenance and failure costs. Include the time required to investigate model changes, update tests and handle edge cases.
The FinOps Foundation's token economics guidance connects AI consumption to business value instead of treating token price as the whole cost. A simpler workflow, fewer tools or clearer ownership may produce a better operating result with lower maintenance.
Test one real example before scaling
Imagine inbound lead routing. The predictable path can validate required fields, check duplicates, match existing accounts and apply territory rules. AI might assist only when a free-text description requires classification.
Strategic accounts, unclear identity, privacy requests and low-confidence classifications should route to a person. The model interprets one bounded input. It does not own assignment policy, customer priority or permission to change protected records.
Build for client ownership
The finished system should be understandable without dependence on the person who built it. Document triggers, rules, model use, permissions, tests, logs, exception owners, operating costs and rollback. Train the internal owner and verify that the team can operate the normal and exception paths.
Review the workflow automation and AI systems service to see how diagnosis, implementation, testing and handoff stay connected, or use the automation decision model for a shorter version of this framework.