The Calls Nobody Shows You

By the end of this newsletter, you'll be able to:

1. Run a bake-off: the test that settles tool arguments with data instead of opinions
2. Deliver "the plan needs to change" three weeks before a launch, and survive it
3. Recognize the moment when NOT rebuilding is the right engineering call
4. Ask the one question that keeps innovation from derailing execution
5. Judge any AI consultant (including me) by their conversations instead of their stack

TL;DR: Monday I showed you the math and the tools. Today: four real conversations from the last three months of client work, and what each one taught me. Because the stack was never the hard part.

Everyone shows you their stack. Nobody shows you their calls.

Scroll LinkedIn and you'll see a hundred diagrams of who's using which model, wired to which database, deployed on which platform. I posted one myself this week. What you almost never see is the conversation where a founder says "I'm stunned" three weeks before launch, and you have to decide in real time whether to hold your ground or fold.

On Monday I wrote about walking a client off Airtable and building custom. Today is the part that doesn't fit in an architecture diagram: four conversations from that season of work, quoted with permission and anonymized. Each one taught me something the tools never could.

Receipt 1: The bake-off

Three weeks before a client launch, something felt off about the document-processing tool we'd built the plan around. I couldn't prove it with a feeling. So I built the test: my AI-powered parser against the incumbent tool, on 105 real documents, field by field.

The AI parser held its own on accuracy. Then came the cost column: a fraction of a cent per document. On the call, the founder stopped me: "That's amazing. Like, that's not even a cent."

But price wasn't the real finding. The incumbent needed a human to update templates every time a document format changed. The AI parser handled formats it had never seen. One approach degrades as the world drifts; the other absorbs the drift.

His verdict, verbatim: "You've got some wizardry. I like it. Let's go with it."

The lesson: don't argue about tools. Nobody wins a tool debate with opinions, including consultants. Build the smallest honest test, run it on real data, and let the numbers do the talking. In 2026 the test costs an afternoon.

Receipt 2: The shock

That same bake-off surfaced something harder. Running the comparison forced a question nobody had thought to ask: how much of the plan was actually verified, and how much of it was assumed? When I checked, the answer wasn't what any of us expected. Three weeks from shipping, we were standing on an assumption instead of a measurement.

Nobody wants to be the consultant who blows up the plan late. It would have been far more comfortable to say nothing and let launch day find out.

The founder's first reaction was exactly what you'd fear: "This is shock mode. Like, at three weeks out, a whole plan is now like, hey, it doesn't work and we could do it better... I'm stunned."

I held my ground for one reason: 105 documents, field by field. When someone is stunned, your certainty doesn't help them. Your evidence does.

By the end of the call: "You're exposing stuff I need to know. So thanks."

The lesson: data beats deference. The models do the typing now. The consultant's job is the test nobody asked for and the sentence nobody wants to hear, said early enough to matter.

Receipt 3: The one I lost

Different client, different call. I pitched restructuring their data pipeline. Cleaner architecture, better foundations for the dashboards we'd want later. I had the full case ready.

The response: "Bro. I feel so good about where we have things right now. Breaking it all down to restructure... I just don't know if it's worth the squeeze currently."

He was right. The disruption cost of rebuilding a system his team already trusted was higher than the cost of the mess. We kept his structure and routed only new work through the cleaner pattern. The old "mess" is still running today, and it's fine.

The lesson: AI made proposing changes nearly free. I can produce a migration plan, scripts, and a slide deck in an hour, which means "can we?" is no longer the bottleneck. "Is it worth the squeeze?" is. And that question belongs to the person who lives in the system, not the person excited to rebuild it.

Receipt 4: The alignment talk

The conversation I think about most wasn't about technology at all.

A client leader sat me down mid-project: "Sometimes innovation can sort of derail execution... I just want to make sure you and I are aligned on what things need innovation right now and what needs execution."

He wasn't wrong. I have a bias for action, and AI has poured gasoline on it. When you can build anything in a day, everything looks like it should be built today. I owned that, and we redefined my role in one sentence: execute what we agreed, and raise a flag the moment something I see affects the work. Innovate on the flagged things, not on everything.

His close: "I think we're fine." We were.

The lesson: knowing what's possible stopped being the scarce skill; the models made almost everything possible. The scarce skill is knowing which of two modes this moment needs, innovation or execution. Ask it out loud. One question, and it saves whole quarters.

What the receipts add up to

Months after the messy middle of all this, one of those clients said something on a numbers call that stuck with me: "Once we got the architecture right in the beginning, all of these numbers and metrics became that much easier... you know that the numbers match the story."

That's the through-line. The unglamorous early work (the honest test, the hard call, the argument you lose on purpose, the alignment question) is what makes the impressive stuff possible later. The AI writes most of the code now. What clients are actually paying for is judgment under pressure, delivered in conversations exactly like these four.

So when you're evaluating anyone selling AI-era help, a consultant or an agency or me, don't ask for their stack. Ask what their last hard call sounded like.

I'm open for consulting and advising engagements. And I'd genuinely love to hear your version: hit reply and tell me about a call where the plan changed, the argument you lost, or the test that settled it. Best story gets featured (anonymized, with permission) in a future issue.

Know someone who only ever shows you their stack? Forward this to them. I write these so people can see what building with AI actually sounds like, not just what it looks like in a diagram.

Share on LinkedIn · Share on X

This is Automate With Rob, a newsletter about building with AI tools without losing your mind. If someone forwarded this to you, subscribe here. Monday's issue, the $135M argument for no-code plus the rent multiplier calculator, is where this week started.