Monitor your agents: AI and human

Jason Lemkin published a story about SaaStr receiving a scorched-earth email from a PR firm, blocking all future contact on behalf of an AI startup executive. When SaaStr reached out to that executive directly, he was baffled. He had a great experience with SaaStr and had no idea his agency had fired off that message on his behalf. The PR firm had been acting unilaterally, burning a relationship their client valued, while their client stayed completely in the dark.

The two-class agent problem

Most B2B founders think about agent oversight as a purely technical question, something for the engineering team to handle in the AI layer. That framing misses half the problem.

You have two classes of agents operating in parallel right now:

Human agents. PR firms. BDR teams. Fractional SDRs. Agencies managing your LinkedIn DMs, your partner communications, your media relationships. They have your brand voice guide, your talking points, and their own judgment about what to do when a situation falls outside the script.

AI agents. AI SDRs running outreach sequences. Chatbots handling inbound leads. Automated follow-up tools. LinkedIn engagement tools. Any system that generates and sends communications on your behalf.

Both classes share one failure mode: they can operate for weeks in ways that contradict your intent, and you only find out when someone complains, or stops responding, or fires off a scorched-earth email to a community you actually care about.

The SaaStr example is a human-agent failure. But Lemkin also notes their own AI SDR invited a prospect to "meet next week at SaaStr" when SaaStr was happening that same week. Another vendor's AI agent pitched SaaStr their product while SaaStr was already a paying customer. These are AI-agent failures. In every case, no human was watching.

What "auditing" actually means

Founders hear "audit your agents" and picture a quarterly review. That is too slow and too thin.

A real audit has three components:

Sample the actual output, not the template. Pull ten emails your AI SDR sent this week and read them verbatim. Not the sequence template, not the A/B test subject lines, the actual emails that went out with actual prospect names inserted. Check whether the personalization logic produced anything embarrassing. "Congrats on the Series B you announced in 2023" sent in 2026 is a real failure class, and it happens because enrichment data goes stale.

Check for context drift in your human agents. Ask your PR firm or BDR team to forward you the last three cold emails or pitches they sent on your behalf. Not for approval, just to read. If the messaging has drifted from what you'd say in the room, that gap is costing you.

Map who can speak for you without approval. This is the governance question most founders skip. Make an explicit list: which agents (human and AI) can send communications autonomously, which require a human review step, and which require your direct sign-off. If you have not made that list, you are operating on implicit assumptions, and implicit assumptions are where the SaaStr situation lives.

Why this compounds in the LinkedIn layer

The SaaStr story involves email and PR, but the same dynamic plays out in LinkedIn outreach at smaller scale with higher frequency. If you are using any tool to automate LinkedIn connection requests, DM sequences, or comment engagement on your behalf, every one of those interactions carries your name and profile photo.

We have covered the difference between engagement-led inbound and broadcast posting in depth at /blog/engaging-vs-posting-linkedin, and the core argument holds here too: LinkedIn is a relationship channel. A badly timed or contextually wrong automated message does not just get ignored, it gets screenshotted and shared.

The failure modes in automated LinkedIn outreach look like this: connection requests that reference a company's funding round from two years ago as if it just happened, DMs that pitch a product the recipient already uses, comments that are technically on-topic but clearly templated. Each of these is recoverable in isolation. At scale, they define your reputation.

The operators we work with who run the cleanest LinkedIn engagement do one thing consistently: they read a sample of their own outbound each week. Not every message, a sample. Five DMs. Ten comments generated by a tool. The act of reading them as a recipient would read them catches most problems before they compound.

The governance gap most founders have

Here is the real pattern we see across B2B founders who have started deploying AI agents in their GTM stack. They approve the initial configuration. They review the first batch of outputs. Then the attention moves to the next project and the agents run unsupervised for months.

The AI SDR gets smarter. The sequences evolve. The enrichment data goes stale. The PR firm changes account managers and the new person has a different interpretation of your positioning. Nobody flags any of it because nobody is looking.

This is a governance gap, not a technology gap. The tools exist to monitor what your agents are doing. What most teams lack is a rhythm: a recurring calendar item where someone with context actually reads the outputs and compares them to current positioning.

For AI agents, most platforms offer logs. Use them. For human agents, ask for reports that include actual copy, not just metrics. "We sent 47 emails this week with a 42% open rate" tells you nothing about whether those 47 emails represented your company well.

If you want a sharper read on where AI agent stacks are heading for GTM specifically, /blog/ai-agent-stacks-saastr-owner-klaviyo covers what operators at SaaStr were actually building and where human oversight was showing up as the differentiating variable.

The rule is simple

If an agent is speaking for you, you are responsible for what they say. The person on the receiving end does not ask whether the message came from a human or a system, an approved script or a rogue judgment call. They got the message from you.

Lemkin's team got lucky. The executive came back to them directly, which is rare. Most of the time a burned relationship just goes quiet, and you never know it happened.

The fix is not to stop using agents. Human and AI agents are how you scale outreach without scaling headcount. The fix is to close the monitoring gap: sample real outputs weekly, map who can speak autonomously versus who needs a review step, and treat stale context in your agents the same way you would treat stale inventory. It expires, and expired inventory costs you.

Build the rhythm now, before a relationship you wanted to keep gets burned by a message you never approved.