How AI Agents Are Replacing Traditional SaaS

A few years ago, I was part of a team building a CRM. Not using one. Building one.

We spent months on things that had nothing to do with the actual product. Email sync. Calendar integrations. A pipeline view where reps could drag deals across stages. Notification logic. Role-based access. An activity feed that actually updated in real time. Each of these felt like a small feature on a roadmap. Each was a rabbit hole that ate weeks.

And after all of that, here's what the CRM still couldn't do: anything on its own.

It was a very expensive place to store data. Every insight it produced required a human to open it, read it, decide what it meant, and go do something. The software was passive. The human was the engine.

Fast forward to last year. Different company, similar problem space. A sales workflow product, but this time the target was agents, not dashboards.

The difference was hard to explain until I lived it.

What We Built Then vs What We Build Now

When we built the CRM, the entire architecture was built around the assumption that a human is always in the loop. The UI was the product. Every screen existed to help a person make a decision faster.

The agent-first product has almost no UI. The agent reads the emails. The agent scores the leads. The agent drafts the follow-ups, flags the at-risk deals, and pings the rep only when a human decision is actually needed. The rep reviews output. They don't operate software.

That single shift, from "give the human a better tool" to "let the agent do the work and loop in the human when needed", changes everything about how you build.

On the CRM, we had long arguments about button placement. On the agent product, we have long arguments about what the agent should be allowed to do without asking.

That is a fundamentally different problem. And honestly, a much more interesting one.

The Part Nobody Warned Us About

With the CRM, the hard parts were predictable. Integrations were painful. Data modelling was painful. Getting sales reps to actually log their calls was painful.

With the agent product, the hard parts surprised us.

The agent worked great in demos. It would handle a straightforward sales thread beautifully. Then it would hit an edge case, a prospect who replied in a mix of Hindi and English, a deal that had three decision makers CC'd with conflicting signals, and it would either hallucinate a confident response or do nothing and time out.

We spent more time on failure modes than on features. What does the agent do when it's not sure? How does it hand off to a human without making the human feel like they're babysitting software? How do you build trust with an end user who can't see what the agent is doing or why?

These questions don't exist in traditional SaaS. In SaaS, the human is always driving. With agents, you're designing for a system that drives itself most of the time, and that changes how users relate to the product entirely.

Where SaaS Still Wins

I want to be honest here because most of the "agents will replace everything" takes are written by people who haven't actually shipped one.

For structured, predictable workflows, SaaS is still better. A tool that generates invoices, or manages employee leave, or handles e-commerce returns works perfectly well as a form with logic behind it. Replacing that with an agent adds complexity without adding value.

SaaS also wins on auditability. Every action in a traditional system is a button click with a timestamp. Auditors love that. Regulated industries love that. An agent that "handled" something is a harder story to tell to a compliance team.

And SaaS wins on trust. People have spent 20 years learning to trust software that does exactly what they click. Agents ask them to trust software that does things they didn't click. That trust gap is real and it takes time to close.

Where Agents Win, Badly

But there is a category of work where agents don't just win, they make the old approach look absurd.

Anything that requires reading, synthesising, and acting across multiple systems. The CRM we built had an "insights" tab that nobody used because it required someone to go read it and then go do something about it. An agent doesn't have an insights tab. It reads everything, decides what matters, and goes does something about it.

Anything that runs on a schedule but needs judgment. Automated email sequences are a SaaS staple. They're also rigid and impersonal in ways that agents are not. An agent doesn't send email number 4 on day 7. It reads the last reply, decides if now is a good time, personalises the message, and sends only if the signal is right.

Anything that lives across too many tools. The CRM had 40 integrations. We built 12 of them. Each was a small product in itself. The agent doesn't need integrations in the same way. It reads and writes to systems the same way a human would, sometimes literally by using the same interface a human would.

What This Means If You're Building Something Now

If you're starting a SaaS product today in a category where agents can operate, you're not late. But you're building in a market that will look very different in three years.

The companies that will win are not the ones that add an AI feature to an existing SaaS. That's what we thought we were doing at first with the CRM. "Let's add AI summaries." It was cosmetic.

The companies that will win are rethinking the whole thing. Who is the primary user, the human or the agent? What does the human actually need to see? What should never require a human decision at all?

Those are hard questions to ask when you have an existing product and existing customers. They're much easier to ask when you're starting fresh.

If I were starting the CRM today, I wouldn't build a CRM. I'd build an agent that makes the CRM invisible.