Start with the failure mode
Delay, rework, error, and inconsistency are symptoms. They may arise from unclear policy, missing ownership, poor inputs, duplicate entry, broken integrations, unstable demand, insufficient capacity, or a genuinely variable judgment task. Selecting AI before distinguishing those causes can automate waste or make the source of failure harder to see.
Signals that the process—not intelligence—is the primary problem
The answer already exists
People search multiple locations because information architecture, access, or ownership is weak. Improve the source and retrieval path first.
The rule is stable
The same explicit condition produces the same action. Use a rules engine, validation service, or workflow control rather than probabilistic generation.
The input is defective
Missing fields, inconsistent identifiers, stale records, and weak data capture should be corrected at source before adding a model.
The bottleneck is downstream
Faster upstream output merely creates a larger review, approval, production, or service queue.
Ownership is ambiguous
No technology can make an end-to-end process improve when no one has authority over its handoffs and performance.
Volume is too low
A one-time or low-frequency task may not justify integration, evaluation, monitoring, and support cost.
Separate deterministic, assistive, and action-taking work
| Work pattern | Usually appropriate mechanism |
|---|---|
| Exact lookup, arithmetic, eligibility, entitlement, or reconciliation | Governed query, calculator, decision table, or validated code. |
| Several authorized sources must be reconciled or explained | Permission-aware retrieval with citations, and generation only when synthesis adds value. |
| Prediction informs a recurring allocation decision | A calibrated predictive system with an incumbent comparator and decision-owner oversight. |
| Multi-step action across tools | A bounded workflow with external policy enforcement, approvals, transaction limits, verification, and rollback. |
| Policy, ownership, or data is unstable | Process remediation and instrumentation before any machine-intelligence pilot. |
The no-AI test
- Can a clearer policy or form remove the ambiguity?
- Can one authoritative source replace repeated search and reconciliation?
- Can integration eliminate duplicate entry?
- Can a deterministic rule handle the standard path and isolate true exceptions?
- Can better staffing, training, scheduling, or ownership solve the actual constraint?
- Would the total cost of evaluation, review, monitoring, and correction exceed the value of the task?
Decision output
The analysis should end with a named intervention, the evidence supporting it, the assumptions still unresolved, the owner, the expected measure, and a reason to revisit the decision. “No AI now” is not a refusal to modernize; it is a specific operating choice that preserves the option to add intelligence when the process can support it.
