I Gave My AI Agent a Budget. Here's What Happened.

I've been running an autonomous AI agent for several months now. It does research, monitors accounts, drafts outreach, synthesizes intelligence. It's useful, but it's been operating in a controlled environment — defined tasks, defined outputs.

So I ran an experiment. I gave it $50 in API credits and a simple instruction: generate pipeline.

No specific accounts to target. No outreach templates to follow. No defined process. Just: figure out who to reach, figure out what to say, and try to create conversations with potential clients.

Here's what happened.

What It Did Well

Research was genuinely impressive

Within 20 minutes of starting, the agent had identified 23 companies that matched my ICP — Series B SaaS companies with 50–200 employees, aggressive growth targets, and signals suggesting they were rebuilding or scaling their sales motion.

The research depth was excellent. For each company, it pulled recent funding announcements, job postings (which signal intent better than almost any other data source), executive LinkedIn activity, and relevant industry news. It synthesized this into a one-paragraph account brief for each target.

This is the [AI Account Researcher](/projects/account-researcher) workflow I've productized on this site — at scale, running autonomously, with no human prompting.

Personalization was above average

For each account, the agent drafted a first-touch outreach message. The personalization was real — it referenced the specific signals it had found, connected them to relevant pain points, and made a specific ask.

Were they perfect? No. Were they better than the average SDR cold email? Absolutely.

The volume was legitimately impressive

In 4 hours, the agent produced: 23 researched accounts, 23 personalized outreach drafts, a prioritization ranking by fit score, and a recommended contact for each account with engagement angle.

That's a solid week of SDR work, done while I was doing other things.

What It Couldn't Do

It couldn't actually send anything

This is by design — I haven't given my agent email sending credentials, and I won't yet. The liability of an autonomous system sending emails under my name without review is too high at this stage.

But it's worth noting: the last mile of outbound — actually contacting humans — is still a human step. For now.

It couldn't navigate ambiguity

When I gave it the open-ended "generate pipeline" instruction, it interpreted this narrowly: find accounts, draft messages. It didn't ask clarifying questions. It didn't explore whether there were faster paths to pipeline (like referrals, or activating a dormant relationship) that didn't involve cold outreach.

A human strategist would have started by asking: what's your current network? Who do you know? Where's the fastest path to a conversation?

The agent went to cold outreach because that's what it knows how to do. This is a real limitation — agents optimize for what they're trained to do, not for what the situation actually calls for.

It couldn't read relationship signals

Two of the companies it identified were ones I'd had prior contact with — one a successful client engagement 18 months ago, one a deal that went cold for reasons worth understanding before re-engaging.

It found them because they fit the ICP. It had no way to know that the approach to re-engaging them should be fundamentally different from a cold approach.

Context that lives in my head — relationship history, why things ended, who trusts me — is invisible to the agent.

What This Means for Sales Teams

The experiment taught me something important about how to think about AI agents in a sales context.

Agents are exceptional at research, synthesis, and generation at scale. They're weak at judgment, context, and the kind of relationship intelligence that lives in a rep's head.

The right model isn't "replace SDRs with agents." It's "give your best SDRs an agent that handles everything except the judgment calls."

Your rep should be reviewing the 23 accounts the agent found, adding relationship context the agent couldn't know, editing the outreach to reflect things that don't show up in public data, and making the actual send decision.

That version of an SDR can work 3–5× the pipeline of a SDR doing all of this manually. That's the unlock.

The $50 Result

Total spend: $47.23 in API costs. Time invested by me: about 20 minutes of setup and review.

Output: a prioritized list of 23 accounts with research briefs and personalized outreach ready to go.

Did it generate pipeline? Not directly — that depends on what you do with the output. But the input required to start 23 conversations cost me $47 and 20 minutes. The question every sales leader should be asking: what does that math look like at scale, integrated into your existing motion?

The tools to find out are [already on this site](/projects). The only thing missing is the decision to start.