An admission of guilt

A few weeks ago I walked my team through where Trellace is headed over the next few years. Partway through, I had to stop and sheepishly admit that I thought the deck was garbage.

The ideas in it were mine, and they were all there. I’d done the thinking, but I was short on time. I was in back-to-back calls, so I handed AI everything after that: the structure, the wording and the slides. Then I grabbed what came back and jumped in with the team. It was long and full of AI phrasing, and you had to dig to find my point. Even with all our templates and writing instructions, it completely whiffed.

My inbox has the same problem from the other side. It’s full of outreach that was obviously written by AI and clearly doesn’t apply to me. One recent email wanted to sell me exotic fish. And if I see one more “quick question,” I’ll scream.

I think my garbage deck and that fish email are the same mistake at two different scales, one in how a person works and one in how a company sells. In both, someone left the ends of the work to AI.

I call the fix the barbell method. I’m supposed to be the person at Trellace who’s best at it, which is what made the deck so embarrassing.

My soapbox: the barbell method

My team is probably sick of hearing me talk about this, but this is how I picture the proper use of AI. Think of a barbell. The weight sits at the two ends, and the bar between them is long and light.

The first 20% needs to be human. Before you open a chat window, work out what you’re trying to do and what a good result looks like. Figure out what information you need and what context matters. Don’t jump straight to the tool.

The middle 60% is where AI can be really helpful. It’s very good at sifting through a lot of data, gathering it, structuring it and filling in the gaps.

The last 20% is human again. Read the output line by line and question it. Did it capture what you were actually trying to do? Did it pull in the right context? Is the explanation structured well, and does it suit the audience and the medium? Never copy something out of a chat window and send it to anyone, inside the company or out, without scrutinizing it first.

The same applies whether it’s a chat window, an outreach campaign or an automation. You need humans on either end, or you’re going to produce errors, put off your teammates, clients and prospects, and generally add to the slop problem plaguing the world.

The percentages are rough, and a two-line email doesn’t need a planning session. The part to keep is people at both ends and AI in the middle.

The average answer

When people skip the first 20%, they rarely notice. They type a broad request, and the tool hands back something that looks finished.

A language model tends to produce the most likely response to what you asked, based on what it learned and what you told it. For a lot of the middle of a job, such as formatting a table, the likely response is what you want.

It works less well for strategy. Ask AI who your company should sell to with a one-line prompt, and you’ll get roughly what a competitor would get from the same line. Give it real context, like which customers stayed and which deals died, and the response gets more precise. A likely answer also saves you from reinventing the wheel or re-solving a problem someone already solved.

That’s not always what you need, though. When you need a truly new idea, AI isn’t going to one-shot it for you. If we want uncommon results for our clients, we need new, specific ideas, and we can’t let AI do that thinking for us.

Set up well, AI can be a good sparring partner. I’ve built custom skills for that, with different modes for different kinds of decisions, including one for go-to-market. They lay out options with their trade-offs and push back on my assumptions, and then I make the decision.

The first 20% is also where my deck went wrong. AI can produce a good presentation if you’re specific, give it real context, and hand it your branding and templates. It had our templates, but I hadn’t told it what I needed.

A sales team has its own first 20%: who you’re going after, what you’re going to say and why anyone should care. Every later step depends on those decisions, and people should make them.

I’d guess nobody made that call for my exotic fish email, and that a tool picked who to contact and then wrote the message too. Buyers notice, and they hold it against you. Gartner found that 73% of B2B buyers actively avoid suppliers who send irrelevant outreach.

The quick look

In my experience, the last 20% is the part people skip most, and skipping it costs you in two ways.

The first is that AI can be wrong in a very confident voice. Because I think it’s funny, I often describe it as a very drunk college professor. It’ll tell you a lot of things with total confidence, and most of them will be right. Every now and again it says something completely wrong, and you can’t tell, because everything around it was right and it all came with the same confidence and the same facts and figures.

Recently, we were asked to do a deep analysis of a large set of a client’s sales calls using AI. The model concluded that one rep’s long sales cycle came from having more complex accounts than his colleague’s. It sounded reasonable. The Trellace team member on the project didn’t buy it, so she dug back in and had the model go through the full CRM history. His accounts looked a lot like his colleague’s, and the gap turned out to be in how he followed up, which a skim would have missed.

The second is trust. AI can be very verbose. Even when you steer it well, it hands back a lot, and it’s easy to stop reading carefully. A single error calls the rest of the document into question, the same as it would with a person’s work, and you start wondering how carefully the whole thing was reviewed. Don’t mistake the confidence AI speaks with for accuracy.

People have already learned to be wary. Across 13 experiments published in 2025, people who disclosed that they’d used AI were trusted less than people who didn’t. A team at BetterUp Labs and the Stanford Social Media Lab even has a name for the stuff that circulates inside companies: workslop, AI-generated work that looks polished but doesn’t move the task forward. In their survey of 1,150 US desk workers, 40% had received some in the past month, and about half of them thought less of the colleague who sent it.

I think that wariness is earned. AI lets you appear to do the same amount of work in a lot less time, at lower quality. It also lets you do much better work in the same time, or better work in slightly less time, or more work at higher quality in the same time. Which one you get depends on how you choose to use it.

On a sales team, the last 20% needs an incredibly human touch. Say you spend real time finding exactly the right buyer, and then you let AI write the email. AI doesn’t know that you took a vacation to the town where that buyer lives a few summers ago, had the best time and still think about the coffee shop downtown. Finding a real, personal connection with someone takes time, and it shows when you reach out.

I think it comes back to the law of reciprocity. People can sense how much time and effort went into reaching out to them, or how well you prepared for a meeting. When you respect their time, attention and inbox, they tend to give you the same respect back.

The middle, done well

The middle is where AI does its best work, and Vercel has a good example.

On Lenny’s Podcast, Vercel’s COO, Jeanne DeWitt Grosser, walked through the agent that now handles their inbound leads, meaning people who contacted Vercel first. A go-to-market engineer started by shadowing their best-performing SDR to see what that rep actually did with a new lead. About six weeks later, the agent was ready.

The agent decides whether a lead looks qualified, using the process they wrote down from watching that rep. It researches the lead and drafts a reply. Then a person reviews every message and hits send.

Vercel went from ten inbound SDRs to one, and lead-to-opportunity conversion held flat. The other nine moved to outbound.

I know “ten to one” lands hard right now. At Vercel, people moved into different work. I’m less sure it goes that way everywhere. There’s a genuine, valid concern about AI taking people’s jobs, and I share it. If this technology keeps progressing at this pace, we’re going to need a serious conversation as a society about what work looks like and how we take care of one another, so that everyone feels they contribute and benefit. That’s a whole other article, and I don’t have the answers.

What I do believe is that AI is a tool, and for a business that’s trying to have a positive impact and grow, it can free people up for more creative, interesting work. Grosser’s stated goal is to get salespeople spending 70% of their time with customers, up from the 30% to 40% she’s seen for most of her career. I think that’s the point of automating the middle. It gives people more time for the ends, which on a sales team means time with customers.

Where I’m less sure

Models keep getting better, including at persuasion. In a 2025 study in Nature Human Behaviour, GPT-4 given a few basic facts about the person it was debating was more persuasive than human debaters in 64% of the matchups that had a clear winner. That was GPT-4, and we’re on GPT-6 now. Even so, as AI saturates outbound, I’ve become suspicious of even the most personalized cold message. I find myself drawn to in-person connections, or at least video calls, running into people and hearing them speak. My short-term view is that better models will push us further toward channels that are hard to fake. If deepfakes become indistinguishable, we’re all in the Matrix, and I don’t know what happens then.

The last 20% can also get quicker. The more work you do up front, the closer the output gets to what you need, especially on simpler tasks where you’ve built templates and skills. It’s like training a new employee. Give them clear process steps and templates, and you get what you expect back in fewer rounds. Just don’t let your guard down.

Then there’s time. The barbell takes longer than most people expect AI work to take, because a vague prompt, a walk away from the desk and a paste is quicker. It’s still much faster than doing the work without AI, and you can usually get more done than you could before.

And not everything needs AI. For really creative work, I often sit in a chair with a pad of paper and a pen to hash out a complicated idea I’m wrestling with. If I’m working out how something should look, it’s paper and pen, or a wireframe I build by hand in Figma. Sometimes the best tool is a phone call and a real sparring session with another person.

When I talk with investors about diligence, one question I suggest is how a company uses AI in its go-to-market. More isn’t automatically better. From experience, though, I’m usually more skeptical of the company using none, because it’s really hard to use AI well.

Everyone gets the same middle

The tools that research accounts and draft copy are available to every company at roughly the same price, and they improve for everyone at once. The middle of the work is becoming something every competitor has.

That leaves the part before the tool and the part after it, meaning what you decided before you opened the chat window and what you did with the output before anyone else saw it. As AI improves, I expect those two ends to matter more.

That deck I called garbage had all the right ideas in it. I rushed the setup and didn’t read what came back.