Does it need AI, or a script?
Save this for your next AI use case meeting. Half the backlog will sort itself with one question.
AI use case meetings collect ideas fast and sort them slowly. Everything sounds like AI now, including things software has done reliably and cheaply for twenty years. The sorting question:
If yes, that is the AI part. If no, that is a script.
Seven patterns, with what to look for in your own operation and the trap where a script still wins.
1. Work that starts with messy input
AICustomer orders arrive as emails, texts, and PDFs and have to land in the ERP. Software can now start with the email, the scan, the voice note. Any layout. Look for work that starts before the data reaches the ERP, CRM, or claims system.
ScriptThe file arrives in the same shape every time. Nothing reads a CSV better than a parser.
2. Comparing meaning
AIAn invoice against a purchase order, a contract against its last version. The software says what changed. Look for two documents side by side and someone eyeballing the difference.
ScriptBoth sides sit in the same system with the same field names. That's a join.
3. Language as a workflow step
AIOne support email contains three different requests. The software reads the message, works out the requests, prepares the next action. Look for anyone who reads text in one system and creates work by hand in another.
ScriptThe sender already picks the right category on a form. Routing is a rule.
4. The long tail of exceptions
AIInvoices that fail matching sit in a queue. AI handles more of the variation; uncertain cases still go to a person. Look for where automation stops and a manual exception queue begins.
ScriptThe exceptions are rare and always the same three. Write the three rules.
5. Small custom software
AIThe weekly account brief nobody ever built. One person builds the first version in hours and deletes it next week if nobody uses it. Look for the useful tool that never survived the IT backlog.
ScriptReports on clean data run on a schedule. Scripts have done that forever.
6. The interface made for the moment
AIThe same five numbers, rebuilt three times: a board memo, a customer briefing, a morning view, from the same information. Look for people exporting data to rebuild the presentation for every audience.
ScriptEveryone actually wants the same five numbers. That's a dashboard, and dashboards are fine.
7. Search across formats
AI"Why did we agree that discount?" Meetings, reports, emails, and decisions connect into one answer. Look for questions that need someone to read half the company first.
ScriptThe answer lives in one field, in one system. Query it.
The good meeting is short
Run the backlog through the reading-and-deciding question and it splits into three piles: real AI use cases, scripts someone should have written years ago, and ideas that were neither. The first pile gets a metric and an owner. The second goes to whoever writes scripts. The third dies happy.
Frequently asked questions
How do I know if a task needs AI or simple automation?
Does someone have to read the input and decide what it means? Yes points to AI. No points to a script or a rule, which is cheaper and deterministic.
What should stay a script?
Same-shape files, same-system joins, pre-categorized routing, scheduled reports on clean data, and dashboards. Scripts win all five.
Where do AI automation projects usually fail?
In the exception queue. Decide early who owns the cases the model gets wrong, and keep the person who notices strange documents.
Sort your backlog against 80 real cases
The AI Use Case Library documents 80 implementations with tools, timelines, and honest numbers, so you can see which of your ideas already worked somewhere. Leave your email to open it.
Open the full library Book a 30-min call