The human is still the point.

Everyone has the same tools now.

The same models, the same agents, the same prompts recycled through a thousand LinkedIn carousels. Somebody posts "the AI playbook that 10x'd our pipeline," you screenshot it, you paste it into Claude or your AI of choice, and you ask how to do it for your company. Ten minutes later you have a plan that looks a lot like everyone else's plan.

That's a photocopy, not a transformation.

I spend most of my time at the intersection of sales and AI. In my various sales roles, whether for my own work at HubSpot or consulting gigs or Foundry, I've built agents that do real work, and I've helped teams stand up the unglamorous foundational stuff that makes AI actually useful. The longer I do this, the more convinced I get of something that sounds almost backwards: the more powerful the AI becomes, the more the human matters.

The call still gets closed by a person

Take sales call prep, because it's the cleanest example I know.

We run an agent like this ourselves. It reads the pipeline, researches the prospect, pulls every prior email and call transcript, and hands over a briefing better than anything a sales manager could have assembled a few years ago with a full week of lead time. After the call, it goes back through the transcript and flags what we missed against our own discovery checklist. It works. It's genuinely great.

Then the call starts.

The agent doesn't run the call. The rep does. The rep has to read the room when the CFO shows up unannounced. Field the question that wasn't in the brief. Decide, in the moment, whether to push or let the silence sit. Choose a tone. Choose the exact words. The brief made that rep more prepared than any rep in history, and it decided exactly none of the moments that determine whether the deal closes.

Give two reps the same brief. One closes and one doesn't. The difference was never the AI.

Context layers are a mirror, not a shortcut

Second example. Knowledge graphs, context layers, whatever you want to call giving AI real institutional memory instead of a blank slate.

The concept sells itself. Feed your meetings, docs, CRM, and Slack into a system, and suddenly your AI answers with your company's actual context instead of the average of the internet. Every meeting, every deal, every internal debate flows into a “brain” that's approaching ten thousand connected facts, and it rebuilds itself fresh every night. Any model will happily walk you through the architecture. The tutorials are everywhere.

Here's what the tutorials skip. Someone has to figure out what the real bottleneck is versus what just feels annoying. Someone has to decide what's sensitive, and I don't mean sensitive by the compliance-checklist definition. The candid note from a client call. The pricing exception you gave one customer that would cause a problem if another customer ever saw it. The internal debate that hasn't been settled yet. None of that appears in a data classification policy. All of it matters.

Someone has to own governance. What updates, when, and who approves it. What the system is allowed to propose versus what it's allowed to actually do. Our own graph now grades its own answers and proposes its own fixes, and not one of those fixes ships without a human signing off. That isn't a technical limitation. That's a decision made on purpose.

Another system we worked on could technically send letters to a client's customers and change account statuses, because the API allowed it. We scoped it read-only anyway and let it earn write permissions one at a time. No model made that call. A person did, because a person understood what one wrong letter costs.

I keep a running list of pitfalls from these builds, and almost none of them are technical. Basically anything can be built at this point. The technical work is the fast part. The judgment calls are the real work, and there is no prompt for them.

Using AI is not the same as becoming AI native

The gap between a company that uses AI and a company becoming AI native comes down to the humans who interpret the output and decide what to do with it.

It's the operator who reads the agent's confident recommendation and says "that's wrong, and here's why," because they've lived the process the agent is only pattern-matching against. It's the person who notices the answer is confidently missing the one constraint that actually matters. It's the leader who understands that a workflow can be technically automatable and still be the wrong thing to automate, because the relationship inside that workflow is the whole point.

I've assessed companies where a third of the staff was pasting client work into personal ChatGPT accounts. Not because anyone was reckless. Because nobody had decided anything. That's not a tooling gap. That's a leadership gap, and no amount of software fixes it.

To be clear, this is the opposite of an anti-AI position. I'm about as deep in this stuff as anyone I know. I use agents daily. My default is to automate. That's exactly why I believe this. The people getting the most out of AI are the ones bringing the most of themselves to it. Foundational knowledge. Critical thinking. Taste. A moral compass, honestly, because a lot of what AI makes easy is stuff you shouldn't do in the first place.

Slop is a choice

AI slop gets blamed on the models. Wrong target.

Slop is what happens when a person removes themselves from the loop. Copy the playbook, paste the prompt, publish the output, skip the thinking. It feels fast in the moment and it's slow everywhere that counts, because nobody trusts it, nobody remembers it, and it compounds into nothing. Meanwhile the person down the street took the same tool, filtered it through actual expertise, and shipped something people act on.

Same technology. Completely different outcome. The variable is the human.

So yes, run in lockstep with AI. Let it prep your calls, draft your first passes, surface the insight you would have missed. Then read between the lines. Question the confident answer. Make the call yourself.

Be a human. That part was never optional.