Blog · AI for small and midsize businesses
From process to use case: map a workflow before automating it
A process map should show what happens when information is missing, a request changes and an approval is delayed. Following one ordinary request and one difficult exception reveals better places for AI assistance than starting with a list of tasks someone hopes to automate.

In this article
A proposal to automate quotations may hide six different activities: understanding the request, checking product identifiers, looking up prices, calculating totals, approving exceptions and sending an offer. Each has different information needs and consequences. Map the full request before assigning any step to AI.
A useful map shows how work actually moves. It includes people, systems, handoffs and routes for incomplete information. OpenAI's workflow-mapping guidance encourages breaking a workflow into individual tasks. For an SME, make responsibility at each transition explicit.
Walk one request from trigger to completion
Select a recent representative case that you may inspect. Ask the people involved to reconstruct events using its actual messages and records. Begin with the trigger and stop only when the result reaches the person who needs it and the final record is stored.
For each step, record the actor, input, output, system, waiting time and decision rule. Write handoffs as events: sales submits a complete requirement to pricing, rather than sales talks to pricing. A handoff is ready only when the receiving person has the information needed to continue.
An illustrative quotation workflow
| Step | Work and evidence | Suitable responsibility |
|---|---|---|
| Receive inquiry | Read an email and identify requested items, quantities and delivery location. | AI may propose extracted fields; a person resolves ambiguity. |
| Validate requirements | Check required fields and product IDs against the approved list. | Fixed validation rules; return incomplete requests. |
| Retrieve facts | Read current stock and authorized customer pricing. | Controlled system lookup using the employee's permissions. |
| Calculate | Apply approved quantities, rates and calculation rules. | Deterministic business software; preserve the calculation record. |
| Review exception | Assess an unusual discount or unavailable product. | Authorized manager or technical specialist. |
| Prepare and send | Create a readable offer from verified facts and approved conditions. | AI may draft; sales reviews commitments and sends. |
These allocations are a hypothetical design, not a description of a completed client project. The AI drafting step should receive verified values, not be expected to reconstruct prices from old emails. The calculation software should handle arithmetic under agreed rules. A polished paragraph cannot make an uncertain delivery commitment reliable.
Map the exceptions separately
Now follow an awkward case. Suppose a customer writes the wrong model number, requests a discontinued item and asks for a delivery date. Where does the workflow pause? Who can clarify the model, approve a substitute and confirm supply? Show separate routes instead of drawing a single arrow labelled handle exceptions.
Also map repeated messages and revisions. If a customer changes quantity after pricing approval, the process needs to invalidate the old approval where appropriate and recalculate. If the same message arrives twice, staff must recognize whether it is a reminder or a new request. Adding AI without these rules can accelerate duplication.
Find the actual bottleneck
Measure active handling and elapsed time separately for the sample cases. If preparing the text takes ten minutes but a pricing approval waits two days, faster drafting addresses only a small part of the delay. Ask whether clearer approval limits or better queue visibility would matter more.
Identify repeated information entry, unclear ownership and steps that exist only because two systems do not share data. These are candidates for process repair or conventional integration. AI is useful where interpreting variable language or preparing a reviewable draft adds value; it is one component in the map.
Make a small redesign testable
Choose one bounded step, such as extracting requirements into a draft intake card. Define required fields, permitted sources, reviewer and return route for ambiguity. Test ordinary messages and the difficult cases already mapped. Compare field accuracy, reviewer corrections and total time until the card is accepted.
The process owner approves the revised workflow, while the systems owner checks access and record handling. Use the NIST generative AI profile as a reference for documenting and evaluating the change. Stop expansion if exception work grows or responsibilities become unclear.
A good process map lets a colleague answer what starts the work, where facts come from, who decides and what happens when the normal route fails. Bring that map and two representative cases to an AI strategy review to identify a sensible first intervention.