DEEP DIVE
If Your AI Workflow Has No Exception Path, It's Probably a Demo
The happy path is the beginning of an automation, not the finished product.
There is a quick way I have started judging AI workflows: show me the exception path.
Not the clean version. Not Lead → AI → CRM → Email → Meeting. Show me what happens when the lead already exists, the data conflicts, the API fails, or the AI simply does not know.
The five-box workflow
You see these everywhere. A lead fills out a form. AI researches the company. AI qualifies them. The CRM gets updated. An email goes out. Maybe a meeting gets booked.
It looks great in a screenshot. It looks even better in a Loom. And if the goal is to demonstrate what is possible, there is nothing wrong with that.
The problem starts when we pretend it is production-ready.
THE DEMO
Lead → Enrich → Qualify → CRM → Email → Meeting
Everything is clean. Every API works. Every record is correct. Nobody does anything weird.
REALITY
Existing lead? Existing customer? Duplicate company? Wrong enrichment? Open deal? Opted out? Conflicting account owner? Low AI confidence? API timeout? Partial failure? Manual override?
The workflow is still the same business process. It just finally includes the business.
Businesses are mostly exceptions
This is something software engineers learn quickly. The first version is rarely the hard version. The hard part is everything you discover after real people start using it.
Someone enters data differently than expected. Two records represent the same customer. A third-party API goes down. A field that was supposedly required is blank. A customer has a special agreement. Somebody manually edits the record halfway through the process.
None of that is unusual. That is normal business.
What if the lead already exists?
Do we create another contact? Merge it? Update the existing record? What if the existing contact belongs to another rep? What if there is already an open deal? What if the existing record is a current customer instead of a prospect?
Suddenly one little box labeled “Create CRM record” becomes search, matching, account-state checks, ownership rules, merge logic, and escalation.
What if the systems disagree?
Your CRM says John works at Acme. LinkedIn says he moved to Beta Industries. Your enrichment provider still says Acme.
What is true? More importantly: who gets to decide what is true?
The hard part is not connecting those APIs. The hard part is deciding the business rule for conflicting information.
What if the AI doesn't know?
A lot of AI workflows assume every decision should produce an answer. Qualified or not qualified. Route to sales or don't. Send message A or B.
Sometimes the right answer is: I don't know.
High confidence can proceed. Medium confidence may proceed with restrictions. Low confidence should stop and ask a person.
The goal is not to make AI decide everything. The goal is to make the business run better.
The eight questions I want answered
STEAL THIS
The Production Automation Test
01 · What is the happy path?
02 · What are the exceptions?
03 · What authority does the system have?
04 · When does a human take over?
05 · What happens when part of the workflow fails?
06 · Who gets alerted?
07 · Can we reconstruct what happened?
08 · Who owns it six months from now?
A prototype is not production
Start small. Build the happy path. Put it in front of real users. See where it breaks. Then harden it.
Just be honest about what stage you are in. A prototype is a prototype. A proof of concept is a proof of concept. A production workflow is a production workflow.
If your automation diagram has no exception path, there is a good chance you are still looking at the demo.
BUILD SYSTEMS THAT SURVIVE REALITY
If you have a workflow that works on the whiteboard but keeps breaking inside the real business, this is exactly the kind of problem we work on at SMB Ops.
Drew Reynolds
The SMB Operator