Should you add an AI agent to your existing SaaS or rethink the workflow?

Your AI agent could save customers five clicks—and leave their biggest problem completely untouched.

That’s what worries me about starting a product conversation with: “Where can we add AI?”

I’d rather start with: “Walk me through what happens when someone actually needs to get this job done.”

Because somewhere between the product demo and a customer’s Tuesday afternoon, the workflow can take a very different route.

And that route is where I’d look for the opportunity.

Let me explain with a kitchen.

A faster oven won’t fix a backed-up kitchen.

Imagine I’m helping a restaurant owner figure out why orders take so long.

The owner wants a faster oven. Reasonable enough.

But when I follow an order, I discover that the food spends most of its time waiting. Waiting for someone to clarify a substitution. Waiting for a missing ingredient. Waiting for two stations to finish their parts.

A faster oven improves one step.

The customer might still wait just as long.

Now put that same situation inside a SaaS product.

A customer opens a request, checks another system, reads an email thread, messages a colleague, updates a spreadsheet, and comes back to your app.

You add an AI agent that summarizes the request.

Useful? Possibly.

But what if understanding the request was never the bottleneck?

What if the real delay was finding the information—and the person with the authority—to move it forward?

Before I choose the AI, I want to discover the work.

When I say “workflow discovery,” I mean following a real task from the moment something triggers it until someone considers it finished.

The product’s screens tell me what users can do.

I want to understand what they actually do.

My starting request would be simple:

“Show me the last time you completed this. Include the parts that happened outside the app.”

Then I’d follow the trail:

  • What started the task, and what did a successful outcome look like?
  • What information did you need, and where did you find it?
  • Where did you stop, wait, double-check, or ask someone?
  • Which decisions followed clear rules? Which required judgment?
  • What happened when the normal process broke down?

I’d want to see a routine case, an exception, and something that recently went wrong.

That spreadsheet on the second monitor might contain a missing product capability.

It might also exist because two departments disagree about who owns a decision.

Those discoveries call for very different responses.

Sometimes, augmentation is exactly the right move.

Suppose the overall workflow makes sense.

Customers know what to do. Responsibilities are clear. The painful part is a specific task: reviewing a long history, collecting context, or preparing a response.

I’d consider adding AI at that point.

Imagine someone handling a customer request. They still own the decision, but the product gathers relevant records, highlights missing information, and prepares a draft for review.

The familiar process stays intact. One demanding step gets easier.

That’s augmentation.

I’d favor it when the surrounding workflow works, the task has a clear boundary, and users can reasonably check the result.

But I’d measure the whole task.

If preparing a response becomes faster while checking and correcting it takes longer, I haven’t established an improvement.

And I wouldn’t assume every useful AI feature needs an autonomous agent. 

Anthropic’s guidance distinguishes predefined AI workflows from agents that decide how to proceed, and recommends starting with the simplest approach that works.

Sometimes a draft is enough. Sometimes ordinary automation can handle a predictable rule.

The product decision comes first.

Sometimes, the workflow itself needs to change.

Now imagine the customer’s job requires carrying information through six screens, three tools, and two handoffs.

Each step exists because a person has to move the work along.

Adding an assistant to every screen could make each step easier while preserving the entire obstacle course.

Here, I’d explore redesign.

Instead of asking customers to collect records, compare them, identify gaps, and route the request, the system could coordinate those activities within defined limits.

The person reviews unresolved issues and makes the decisions that require their authority.

That changes the shape of the work.

But redesign also raises harder questions:

What may the agent decide? What may it change? When must it stop? Who owns an incorrect action? How does someone recover?

I want those answers in the product requirements before they become surprises in production.

And redesign doesn’t automatically mean replacing the interface with a chat box.

Customers may still need a clear place to compare options, inspect evidence, track progress, and take control.

Then comes the part your existing customers have to live through.

Here’s the twist in the kitchen story.

Even if I design a better way to run the kitchen, I still have to introduce it while dinner service continues.

An existing SaaS product has its own version of that problem.

Customers have trained employees, built integrations, assigned responsibilities, and created routines around how it works today.

A familiar approval step may look like friction to the product team. To the customer, it may be how someone remains accountable.

Removing it changes more than the click count.

So I’d evaluate two things separately:

How much better could the workflow become? How difficult will the transition be?

A substantial improvement can justify substantial change. A modest convenience probably won’t.

My preferred starting point would be one bounded workflow with a small group of willing customers.

Let them see what the agent proposes before it acts. Keep a workable fallback. Make completed actions visible, and provide a recovery path where possible.

Then expand its authority as the evidence supports it.

I’d look for improvement beyond the demo.

For either augmentation or redesign, I’d establish a baseline and track:

  • Time from request to completed outcome, including waiting.
  • Human effort spent reviewing, correcting, and handling exceptions.
  • Errors, rework, and recovery effort.
  • Whether customers keep choosing the new workflow after the novelty fades.

An agent completing more actions is interesting.

Customers completing the right work with less effort is the result I’d want.

My decision rule would be:

Augment when a useful workflow has a painful step. Redesign when the sequence of steps creates the pain. Fix unclear rules or ownership before giving an agent more responsibility.

You can also start with augmentation and use what you learn to inform a later redesign. That first release can reveal which decisions customers are comfortable delegating—and which they need to keep.

That’s where I see workflow discovery earning its place in product strategy.

It helps me connect the customer’s actual work to the product roadmap, the technical requirements, and a transition customers can adopt.

If you’re deciding where AI agents belong in your existing SaaS, I help map the workflow, evaluate augmentation versus redesign, and turn that decision into a practical product roadmap—working with AI developers on implementation.

To reach out to me directly click here to discuss your product.

And I’m curious: what’s the one task your customers still finish in a spreadsheet, an email thread, or a message to a coworker—even though they’re paying for your SaaS?

Tell me what happens outside the app. That’s where I’d start asking questions.

#ProblemToProduct #WorkflowDiscovery

Bring the question into your business

If this article reflects a decision you are facing, share the context. We can explore whether a focused conversation or engagement would help.

Start a Conversation →