Blog · AI for small and midsize businesses
When to improve a process before adding AI
AI can produce an answer quickly while leaving the cause of delay untouched. Before adding it, check whether the task fails because information is missing, responsibilities are unclear or business rules conflict. A short repair experiment can reveal whether any AI work is still necessary.

In this article
When purchase requests arrive without quantities, an AI assistant may write polite clarification emails. A required quantity field may prevent the clarification altogether. Both address the same task, but only one prevents the omission: the required field. Before adding AI, identify whether the bottleneck comes from language variability or a broken process.
Generative AI produces plausible outputs with some variability. Adding that behavior to unclear rules can increase review work rather than reduce it. Deloitte's use-case framework explicitly asks why an agentic approach is preferable to traditional automation or rules. The same question is useful for ordinary AI assistance.
Four signs that process improvement should come first
| Symptom | Check | First intervention |
|---|---|---|
| Missing information | Do staff repeatedly ask for the same required field? | Improve intake instructions and validation. |
| Unclear ownership | Does a case wait because nobody knows who decides? | Name an owner and define escalation conditions. |
| Conflicting rules | Do experienced staff give different answers to the same facts? | Resolve and approve the rule before encoding it. |
| Weak master data | Do duplicates, old prices or inconsistent product codes drive corrections? | Repair the authoritative records and assign maintenance. |
Test a process improvement
Take a small, representative sample of failed or delayed cases. For each, identify the earliest preventable defect. Count missing data separately from judgment calls. The process owner chooses one correction that can be tested without a large system change, then observes a comparable period.
In a hypothetical packaging supplier, assume twenty of fifty quotation requests need clarification. Staff inspect the messages and find that fifteen lack either dimensions or quantity. They revise the intake form, explain units and keep a route for customers who need help defining requirements. Sales owns the change; operations checks that the captured fields are sufficient.
The team compares incomplete requests, clarification exchanges and total time to a usable requirement. It also checks whether customers abandon the form. These numbers are illustrative; the business must collect its own result. A stricter form that discourages legitimate inquiries can create another problem.
Decide what remains after repair
If structured intake removes most routine clarification, AI may be unnecessary there. If customers still send varied free-text descriptions, AI could help propose a draft intake card for staff to validate. That is a narrower, better-defined use case because required fields and review criteria now exist.
Other situations require different decisions. Stable calculations belong in conventional software. A disputed discount policy needs management agreement. An unusual technical decision needs qualified expertise. AI can help prepare information around these activities, but cannot supply authority that the process lacks.
Compare full workloads
Measure prevention effort, handling, review and corrections. Do not judge the repair by the number of emails generated or the AI by drafting speed alone. Include the downstream person who receives the output; a convenient front-end shortcut can simply transfer cleanup to another team.
OpenAI's impact-effort guidance supports deprioritizing low-value complexity. The NIST generative AI profile offers a broader evaluation reference when AI remains part of the redesign.
Document the defect, repair result and unresolved work. Bring those findings to an AI opportunity review to decide whether assistance adds enough value after the process is clearer.