Just because AI can run the conversation doesn't mean it should

There's a version of this argument that says agentic AI works brilliantly.

That's not quite right, and it lets the technology off too easily.

The real problem shows up on message one, not month four. An agent can write something grammatically flawless, personalised with the right data points, timed perfectly, and it still won't feel like a person wrote it.

Not because the model isn't good enough yet. Because there's no person there to feel like.

That's the trap with agentic AI in GTM right now.

Not that it degrades over time, though it does. It's that even at its technical best, it's optimising for the things a machine can measure, open rate, reply rate, fields personalised correctly, while missing the one thing a prospect actually responds to: whether someone on the other end gives a damn.

What's actually possible right now

It’s worth being specific here, because most of what gets written about this either oversells the magic or dismisses the whole category, and neither is accurate.

This is what agentic AI is genuinely doing inside GTM teams today, unsupervised, at scale:

• Account and signal research. Agents scanning firmographic data, hiring signals, funding news and intent data to build and rank account lists, the kind of research that used to take a Driver half a day now runs continuously in the background.
• First-touch message drafting. Pulling public signals about a company or a role and generating personalised opening messages without a human writing a word.
• Autonomous sending and sequencing. Full outbound sequences, email and LinkedIn, running end to end. Send, wait, follow up, adjust timing, no human touches it until a reply lands.
• Meeting booking. Once a prospect replies, the back-and-forth of finding a slot and confirming it happens without anyone on the sales side lifting a finger.
• CRM and data hygiene. Deduplicating records, enriching contact data, flagging stale fields, the unglamorous admin that used to eat hours nobody wanted to spend.
• Competitive and market monitoring. Tracking competitor pricing changes, positioning shifts and hiring patterns, surfacing what's changed rather than someone checking manually.

One operator we spoke to described their model as the agent doing over 90% of the work and a human closing the final stretch. That's the case being made for full autonomy on the human-facing side, and on the numbers alone, it's not a bad case.

The bit that doesn't make it into the case study

Here's what happens next, because there's always a next, and it's rarely in the same slide deck as the reply-rate chart.

GTM engines running AI-generated outbound at volume lose sender reputation fast, with sharp drops within 90 days of the volume starting. Reply rates that look strong in month one decay hard over the following year, because recipients learn to recognise the pattern.

Same structure. Same cadence. Same slightly-too-smooth phrasing that reads fine in isolation and instantly familiar in bulk. It works until people notice. Once they notice, it stops working for everyone running the same playbook, not just the business that started it.

The deeper issue isn't the writing quality. It's the absence of judgement.

A human running a sequence that isn't landing changes it, tests a different angle, reads the silence and asks why.

An autonomous agent optimises against whatever metric it was given, usually activity, sends volume, opens, replies, and keeps running the same failing sequence at scale because nothing in the system tells it to stop.

Has anyone built in a moment where the machine asks itself, "Is this actually working, or does it just look like it is."

That's not a technology problem you fix with a better model. It's a structural one. You've removed the one thing that used to catch a bad decision before it compounded, a person paying attention.

That's not an automation problem, it’s a judgement problem.

Most of the AI-in-sales debate argues about whether the tool is good enough yet. Wrong question.

The real one is which parts of this job should ever have been fully autonomous.

Research doesn't need a relationship. Enrichment doesn't need trust.

A conversation with a prospect, the specific moment they decide whether you're worth thirty more seconds of their attention, does.

Hand that moment to a system that can't tell the difference between a message that landed and a message that just got sent, and you're not automating sales. You're automating the erosion of the exact thing sales depends on.

This isn’t about being anti-agentic. It’s about using the tools in the right way. And ensuring your sales experience still gives that human feel.

Let's wrap this up

So where does agentic AI actually belong?

So here's the point of view, and it isn't "use less AI."

Put agentic AI to work everywhere the research list above lives: signal detection, enrichment, data hygiene, competitive monitoring, the pattern-spotting and process work that used to consume a Driver's whole week for no strategic gain.

That's where autonomy is a genuine asset, because none of it requires reading a room that isn't there to read.

Keep the human interaction human. The first message, the follow-up, the moment someone decides whether to reply, stays with a person who can feel when something isn't landing and change it before it scales into a problem. Not because that's more comfortable.

Because it's more accurate, and accuracy is the thing an unsupervised system structurally cannot self-correct for.

The businesses that get this right over the next year won't be the ones with the most automated pipeline.

They'll be the ones who worked out early which parts of the engine can genuinely run themselves, and which parts only work because a person is still paying attention to them.

Will the technology catch up? Probably yes. Does that mean we should use it that way? Not necessarily.

Automate the process. Never the relationship.