Notes from the build.

"Cleaning up the CRM" sounds like a one-time chore.

It isn't. The agentic move installs an operator that keeps it honest. Reads what reps skip. Fixes broken tools. Sends safely. Refuses mass writes without you.

Anything else is automation that ran once.

Move 5. Write the rules down before the session ends.

Script fix in git. 90-second stagger as the default. Contact rule in the agent's skill file.

Next session inherits all of it. Mess gets shallower every week, not reset every Monday.

Move 4. Mass writes go through a hard gate.

Agent drafts every message, surfaces them, refuses to send until you say go. Coded refusal, not a UI prompt.

The line between an operator and a runaway script.

Move 3. Send slower than you think you need to.

35 texts at once = carrier flags your number for "high-volume commercial" suspicion. Deliverability drops for weeks. You never see the flag.

A 90-second stagger reads as a human texting.

Move 2. When the tool is the bug, fix the tool.

CRM helper script creating tasks assigned to the wrong partner? Write a failing test, add the missing assigned_to, run it green, commit. Every call inherits the fix. The bug is permanently dead, not papered over.

The fix: attribute every message back to the contact it was actually sent to, pulled from the API, not the UI.

Then your automation can pick the contact who's actually been talking with you. Naive automation cools live deals here every day.

Most CRMs flatten multi-contact leads into one timeline. It looks like one conversation. It isn't.

If a lead has two contacts and the "primary" one has gone cold, your "send a re-touch" automation hits the wrong person. The engaged seller gets nothing. The UI lies by omission.

Move 1. Read the threads, don't count them.

Counts hide the work. The work is in the actual conversations. Hundreds of KB of transcripts, ranked the way a sales manager would rank them. Three buckets: ready to sell, routine touch, close out.

The bucket sets the action.

When people say "AI cleaned up my CRM," they picture a bulk action.

That's not what an agentic operator does. An operator picks the CRM up, looks under it, and finds the tool itself is part of what's broken.

Five moves that matter.

Most people think 'track your business' means track your leads.

That's maybe a fifth of it.

A business is a stream of events. And the ones that actually run it happen in the back office, where almost nobody is looking.

Everyone thinks the hard part of a voice AI is making it sound human.

It isn't. The voice got solved a while ago. The hard part is speed. Here is the wall nobody talks about.

Direct mail is about to get strange.

An AI can pull a list of homeowners, render what a pool or renovation would look like on each specific house, and mail a personalized postcard. Generic mailers stop working when the one next to it shows your house, upgraded.

The hard part of a voice AI was never the voice. It was the round trip.

Speech to text, the model thinks, then text back to speech. If that loop runs longer than about a second, the conversation feels broken. Winning voice AI is really just winning latency.

Prompt caching is the cheapest speed-up in AI and almost nobody uses it.

If an agent reuses the same long instructions every call, cache the unchanging part instead of reprocessing it. With Claude, cached tokens read at about 1/10th the price of fresh ones.

Anthropic just shipped Claude for Small Business. 15 ready to run workflows wired into QuickBooks, PayPal, HubSpot, Google Workspace and Slack. Free on Pro, Max and Team.

Everyone is reading it as a product update. It's actually a useful signal about where AI creates value.

Most teams stop at AI chatbots. We build the layer that runs the business.

xovionlabs.com