The PaintScratch Color Help agent started with one customer question:
“What paint color should I order?”
That sounds simple until you touch the actual work.
Sometimes the customer has a paint code. Sometimes they have a VIN label. Sometimes they send a photo of the vehicle. Sometimes they picked a color from the site and want to know if it is right.
Those are not the same case.
A clear factory paint code can be strong evidence. A VIN label usually is not enough. A vehicle photo may look helpful and still prove nothing. A selected color can be close without being safe to recommend.
If an AI system treats all of those the same, it is not helpful.
It is just fast at being risky.
The first version was narrow on purpose
The first version of the Color Help agent was not fully autonomous.
That would have been the wrong goal.
The first job was smaller:
Did the customer provide enough evidence to recommend a color?
If yes, help prepare the reply.
If no, ask for the smallest missing item.
Most of the time, that meant a readable photo of the paint-code tag.
Not flashy. But useful.
That is where good AI systems usually start: one repeated business decision with a clear next step.
The brain got built from real cases
The real value was not that AI could write a nicer reply.
The value was that every case taught the system something.
What did the customer send?
What evidence mattered?
Which PaintScratch page needed to be checked?
What reply went out?
Where did the system need to stop?
That became the business brain.
Not a folder of documents. Not a clever prompt. Not “company knowledge” as a buzzword.
A working memory layer that changed how the next case got handled.
Over time, corrections became rules. Edge cases became memory. Repeated patterns became reusable context.
That is when the system started getting better.
Autonomy came later
The agent now does more of the job.
It checks open Color Help requests. It prepares replies. It verifies product pages. It sends safe answers when the evidence is strong enough. It checks Gmail follow-ups. It matches replies back to the right case. It archives handled messages. It routes blockers away from my inbox.
But it did not get there by being told, “Go be autonomous.”
It got there because the workflow became clear.
The system learned what counts as evidence, what does not, when to answer, when to ask, and when to hand off.
That restraint matters.
A business does not need AI that confidently guesses through risky cases.
It needs AI that handles repeatable work and leaves the right decisions to people.
AI takes the first pass. People stay in control.
This is the useful part of “Agent SaaS”
Greg Eisenberg has been talking about “Agent SaaS,” and I think the useful part is simple:
Businesses do not want another AI tool.
They want the work handled.
A customer question answered correctly. A ticket routed. A lead followed up with. A buyer pointed to the right product. A messy inbox cleaned up.
That is the part worth paying for.
But the wrapper is not the work.
The dashboard, logs, and control room matter because they build trust. A business owner needs to see what happened, what was sent, and where the system stopped.
But a control room wrapped around weak judgment is still weak judgment.
For Color Help, the real system is the evidence rules, catalog checks, reply logic, Gmail follow-up, handoff rules, audit trail, and memory from past cases.
That is the work behind the work.
The useful framework
For this kind of AI system, I keep coming back to four parts:
Evidence: What does the customer provide, and is it strong enough?
Action: Should the system answer, draft, ask, check, send, or route?
Handoff: Where should the system stop?
Memory: What should this case teach the system for next time?
That is a better starting point than asking, “What agent should we build?”
Start with the job.
Define what “done” means.
Decide what the system can do safely.
Decide where people stay in control.
Then improve it from real cases.
That is how a useful agent gets built.
The chatbot is just the box.
The brain is what makes it useful.
And when that brain is tied to a real workflow, the conversation changes.
You are not selling access to software.
You are selling the work getting done.