The Bottleneck Was Never Response Time
The brief was obvious. Leads come in, the sales team is drowning, replies go out late or not at all. Build an agent that answers leads so the reps get their hands back.
So that's what I built. It worked. It wrote good replies, in their voice, pulled from their knowledge base, fast.
And it didn't move the needle nearly as much as it should have. Because answering leads was never what was eating the team's day.
The interview that told me nothing
Before I built it, I did what everyone does. I sat down with the sales team and I asked them to walk me through their process.
They gave me a good answer. Clear, articulate, sensible:
"A lead comes in, we review it, we reply, we try to get them on a call, we move them along."
Every word of that is true. And I understood nothing.
What I got was the org chart version of the job: the tidy narrative a person constructs when someone asks them to describe their work in the abstract. It's not a lie. It's a summary, and summaries drop exactly the things I needed. Nobody's mental model of their own job includes the twelve minutes they spend every morning in a tab they'd never think to mention.
I took that clean answer and built exactly what it described. An agent that reviews and replies.
Then I sat behind them
What finally fixed it was boring. I asked to sit with the reps while they worked and just watch. No questions, mostly. Just watching.
Within an hour it was obvious, and it was embarrassing.
A lead would come in. The rep would open it, and then not reply. They'd open another tab. Look the company up. Check the CRM for whether anyone had touched this account before. Find an old deal that went nowhere and read through it to figure out why. Check whether this person was even the right kind of buyer. Ask a colleague on Slack if they remembered this name.
Then they'd write the reply. The reply took two minutes. Everything before it took forty.
The time wasn't going into answering. It was going into qualification — building enough context to know whether this lead was worth answering at all, and what to say if it was. I had automated the two-minute part of a forty-two-minute job and wondered why the team still felt buried.
That's the thing you can only see by watching. When I asked, they said "we review it." That's the whole forty minutes, compressed into two words, because to them it's not work — it's just what you do before the work. The stuff people don't mention is never the stuff they're hiding. It's the stuff that's so automatic they no longer see it.
The things you only catch by watching
Once I started doing this on every project, the same categories kept showing up. None of them ever came up in an interview:
- The extra tab. There's always one. A tool nobody listed in the stack, open all day, load-bearing. Often a spreadsheet.
- The hesitation. The moment someone stops, squints, and checks something manually. That pause is a missing gate in the system, and they've been personally absorbing it for a year.
- "That's just how we've always done it." Said cheerfully, usually about the single most expensive step in the process. It's a fossil of a constraint that stopped existing a long time ago.
- The silent triage. Work that gets skipped so consistently nobody counts it as work. Leads that quietly never get touched because everyone knows they're not worth it. That rule is nowhere in your spec, and your agent will happily violate it on day one.
So I changed the questions
The fix wasn't just watching more. It was that the questions I'd been asking were structurally incapable of getting me a real answer.
"How do you work?" produces a summary. It invites the person to zoom out and narrate, and the narration is where the truth gets lost. The whole trick is to ask questions that make zooming out impossible.
Now I ask two things instead:
"Show me." Not tell me — show me. Open the thing. Do it while I watch. This isn't a politeness upgrade on the same question; it's a different question entirely, and it's the highest-leverage change I've made to how I run discovery.
"How did you close this deal?" One named deal, a real one, start to finish. The specific case can't be summarized, because they have to actually remember it — and the memory drags in all the mess the summary sanded off. Where they got stuck. Who they had to chase. What they looked up. The week it sat.
Ask ten people to describe their process and you'll get ten clean, contradictory stories. Watch ten people work and you'll see the same three problems, over and over.
The part nobody warned me about
There's a second effect here that I did not expect, and I now think it's worth as much as the agent.
When you sit with someone and ask why they do the thing they just did, and then ask again about the next step, you are not only sharpening your own model of the process. You're sharpening theirs.
I've watched people talk themselves out of steps mid-sentence — start explaining a habit, hear it out loud for the first time, and stop halfway with "…actually, I don't know why we still do that." Nobody had ever made them say it out loud. The step didn't need an agent. It needed a witness.
And when the process crosses teams — sales into CS, marketing into sales — the effect compounds. Building the system forces the spec, the spec forces the argument, and the argument is the first time in months anyone has agreed out loud on what actually happens at the handoff. The organization gets better from the act of being mapped, before a single agent ships.
That's a real deliverable. It just doesn't look like one on an invoice.
The principle
People are reliable narrators of what they're supposed to do and unreliable narrators of what they actually do. Not because they're evasive — because expertise makes work invisible to the expert. So don't ask what the process is. Ask to watch it, and ask about one real case that already happened.
Before you scope your next automation, go sit behind whoever does the job for two hours and write down every moment they pause. That list is your actual spec. The one they told you in the kickoff is a summary of a job nobody has.
About to automate the wrong half of a job?
I run a mapping sprint before I build anything. Let's talk about what's really eating your team's week.
Book a 30-min Meeting