We Have 20+ AI Agents and Just 3 Humans. But Even So, We Still Need “Real” B2B Software.

We hear this a lot: “Once you have AI agents the work, you don’t need CRM. Just give them a Postgres database and let them rip.”

It sounds clean. It’s also wrong, I think. At least for 99% of us. At least for now.

At SaaStr AI, we’ve gone from ~30 humans down to 3 humans + 20+ AI agents running in production. Revenue is up 47% YoY, from -19%. The agents have closed $1M+ in revenue. They run win-back campaigns at 72% open rates. Artisan sends 15K+ outbound messages a month. Qualified drives seven figures in pipeline. They’re real. They work.

And every single one of them still needs proper B2B software to operate in.

Here’s what the “just use Postgres” crowd is overlooking:

#1. 99% of humans aren’t going to log their leads in Postgres. They won’t even log them at all as it is.

Not today. Not in 5 years. Maybe not ever.

Your AE isn’t writing SQL. Your VP of Sales isn’t querying tables to build a forecast. Your CFO isn’t running joins to understand pipeline coverage. And frankly, neither are you.

Even in 2026, even with the best AI copilots sitting on top of raw data, humans need interfaces that match how they actually think about their work. Pipelines. Stages. Territories. Quotas. Accounts. Those aren’t database concepts. They’re workflow concepts.

Strip them away and you don’t have a leaner stack. You have a company where half the team can’t do their jobs.

#2. The agents need shared systems too. Maybe even more than humans.

This is the part folks get wrong most often. They assume agents are infinitely flexible, so they can just work directly on raw data.

In practice, it’s the opposite.

If you have 20 agents writing updates to a raw Postgres table, you have 20 different interpretations of what a “qualified lead” is. 20 different stage-transition logics. 20 different ways to handle duplicates. No shared forecasting rules. No consistent ownership semantics. No audit trail anyone trusts.

Our agents at SaaStr — Artisan, Qualified, Agentforce, Monaco, QBee, 10K — all work on top of the same system of record. Not because Salesforce is pretty. Because they need the same shared logic humans do: what’s an opportunity, who owns it, what stage it’s in, what the comp plan says about it, how it rolls into forecast.

The shared substrate isn’t a nice-to-have. It’s the thing that keeps 20+ agents from turning your data into noise.

#3. Every other system in your stack needs a source of truth.

Your marketing automation tool trusts your CRM. Your billing system trusts your CRM. Your BI stack trusts your CRM. Your CS platform trusts your CRM. Your comp tools, forecasting tools, exec dashboards — they all trust your CRM.

Rip that out and replace it with a Postgres DB that 20 agents are writing to in 20 different ways, and suddenly nothing downstream works right.

The integration graph is 25 years deep, and it’s built around the idea that there’s a canonical system of record.

That doesn’t disappear because you hired some agents.

#4. And Salesforce is Safer

Agents make mistakes. A lot of mistakes. We’ve watched ours hallucinate deal amounts, loop on bad prompts, misclassify leads, and in one case, try to update 400 records in a way that would’ve been really bad if it had gone through.

When that happens inside Salesforce, guardrails catch it. Permissions limit blast radius. Validation rules reject bad writes. Audit trails tell you exactly what happened. Workflows require approvals above certain thresholds. 25 years of enterprise-grade safety nets are sitting there, for free, whether you asked for them or not.

When that happens inside a raw Postgres DB you wired up yourself? You find out on Monday morning when the forecast is broken and nobody knows why.

Add to that: SOX compliance, GDPR, FedRAMP, HIPAA, SOC 2, data residency, encryption at rest, backup and recovery, penetration testing. You can build all of that yourself. Or you can use a platform where it’s already built, already certified, and already approved by every enterprise procurement team on Earth.

Giving 20 agents direct DB access doesn’t make you faster. It makes you the one personally responsible every time one of them does something dumb.

Headless is Great, Though (For Us At Least)

Benioff’s Headless 360 announcement the other day is the right read: the CRM becomes an API and MCP substrate that both humans (through UIs) and agents (through tools) operate on. Slack becomes a surface. Voice becomes a surface. Custom UIs become a surface. The system of record stays.

The winning architecture isn’t “replace CRM with Postgres.”

It’s “expose the CRM as a platform humans AND agents can work on together.”

That’s a very different bet.

Our Agents Outnumber Humans 7:1. But We Still Need Software That Works Seamlessly With … Both of Them.

At SaaStr, agents outnumber humans almost 7:1. We’re as bought in on AI agents as anyone. I’ve shipped 12+ production apps myself, Amelia built QBee (our AI VP of CS) on Replit, I built 10K (our AI VP of Marketing) at 14,230 lines of code.

And we still run on Salesforce. Not because we have to. Because when you actually operate this way, you realize fast: the agents need the shared system of record even more than the humans do.

“Just use Postgres” sounds like the future.

In practice, it’s the fastest way to 20 agents producing data nobody — human or agent — can trust.