You can't automate a feeling

"We waste 20 to 30 percent of our time on CRM data management."

I've heard a version of that sentence in almost every scoping call I've ever sat through. It's usually true. And it's almost always useless.

Useless because of what happens next. You ask one follow-up question, and the room goes quiet.

"What, specifically?"

Is it cleaning properties? Moving deals between stages by hand? Pasting AI summaries into deal notes? Fixing duplicate companies after every trade-show import? Reconciling two pipelines that define "Qualified" differently?

Nobody knows exactly. They just know it feels like a lot.

A feeling is not a spec

Here's the trap, and solution partners walk into it constantly. We hear "30% on data" and our brains start building. By the time the prospect finishes the sentence, we're already three steps into an imaginary solution, nodding, mentally scoping the agent.

But "30% on data" isn't a problem. It's a sensation. A weight people carry without ever putting it on a scale.

You can't automate a sensation. You can only automate a task – something with a trigger, an input, a rule, and a definition of done. The distance between "we waste a lot of time on data" and a task you can actually hand to an agent is the entire job. And it's the part everyone wants to skip, because it's slower and less impressive than a demo.

What "what specifically" actually surfaces

When you keep asking until the answer gets concrete, the vague 30% almost always splits into a handful of very different problems wearing the same coat:

1. Entry errors. Reps type company names three ways. Forms accept "acme corp", "ACME", and "Acme Corp." as three companies. This is a validation problem, fixed at intake.

2. Reconciliation work. Someone manually merges duplicates after every import, or hand-moves deals because a workflow doesn't fire. This is an automation problem.

3. Translation work. Pulling three reports and stitching them in a spreadsheet because no single view answers the question. This is a data-model problem.

4. Judgment work. Deciding whether two records are the same human, or what a freeform job title really means. This is the part you don't fully automate – you give it to an agent that knows when to stop and ask.

Four different problems. Four different fixes. One feeling. If you'd started building after the first sentence, you'd have automated the wrong one.

Discovery is the product

The real work happens before the demo, before the proposal, before anyone says the word "agent." It happens when you keep asking "what specifically" until the answer is sharp enough to write down.

That's not a sales tactic to look thorough. It's the only way the thing you build survives contact with the actual portal. An agent pointed at a vague problem produces a confident, plausible, wrong result – at machine speed, across your whole database. Precision up front is what keeps that from happening.

So the first job was never to send a proposal faster or cheaper than the next vendor. The first job is to ask better questions than they did. If you genuinely want to help someone, you earn the right to build by understanding the problem better than they've managed to articulate it themselves.

A demo shows you what's possible. Discovery tells you what's true. Only one of them is safe to automate.