We delivered an AI CRM Administrator for a twenty-three-person North Charleston sign company where every employee had already picked a different personal AI — one that let each of them keep the ChatGPT, Grok Bot, Claude Cowork or Spark they were already fast on, writes every site visit, install and change order into Salesforce out of Slack with nobody typing into a CRM, and stops short of the price field on purpose.

She bought the company in 2011 from the man who started it in a garage off Dorchester Road in 1994. It makes signs — channel letters for restaurants on upper King, wayfinding packages for two hotels, monument signs for the office parks on Daniel Island, and around ninety vehicle wraps a year out of a shop on Clements Ferry Road. Twenty-three employees: three designers, two estimators, four project managers, a shop of nine, and five people who are in trucks by six in the morning. Salesforce is the system of record, and on any given Tuesday it was wrong.

She knew the size of it because she had counted. Of 148 open opportunities in March, 61 carried no logged activity inside fourteen days. It was not laziness, and she never thought it was. An installer who finishes a monument sign at seven in the evening does not open a laptop.

An estimator with six site visits in a day ends it with six voice notes and no typing time. So the record got entered in bursts, on Thursdays, from memory, by people reconstructing a week. Twice that spring the shop fabricated to a proof that had been superseded on site and never written down, and the second one cost $4,200 in remade aluminum. The other thing she knew was that every one of the twenty-three already had a personal AI they used all day and liked.

Nine were on ChatGPT, five on Grok Bot, five on Claude Cowork, four on Spark inside Gemini. Nobody had been told to. It had happened the way phones happened. She had twice been advised to standardize — pick one, buy seats, retrain everybody.

That means taking away the tool each person chose and got fast on, and it ends eight weeks later with half the company quietly back in the old habit. We built the recipe on the company's side instead, so that nobody's AI had to be replaced. Slack is the system of action. Every job already had a channel, so the recipe defines which channels carry which events, what a complete record looks like in Salesforce, and a set of write rules stated as rules about the business rather than as instructions for any one product — which is why the same recipe works for four different AIs.

Then each employee's own AI was brought onto it, one person at a time, on that person's schedule rather than ours. It took eleven weeks. The installers came in on two Saturdays. One estimator came three times.

One man waited five weeks and turned up in August. What runs now has three triggers and none of them is a person opening Salesforce. A message landing in a job channel from its own person is the first. The installer posts a photo and eight words.

His AI, on his laptop under his account, writes the install-complete date to the opportunity, attaches the photo to the job record, closes the open task, sets the warranty start and queues the invoice. A calendar event ending is the second: when a site visit closes, the estimator's AI takes the voice note he recorded on the drive back, writes the survey against the opportunity, moves the stage and flags any dimension that disagrees with the drawing on file. The third is a cadence — at 6:10 each morning every AI reads its own person's book and sends them a short account of it. The boundary is drawn in two places.

Each AI writes only to records its own person owns, never to anyone else's. And no AI moves a deal to Closed Won or writes a price. It writes everything around the number and stops at it. On a Friday in August an installer finished a monument sign at Kiawah at 7:40 in the evening, put one photo and a line of text in the channel and drove home.

By the time he reached Main Road the opportunity was closed to install-complete, the photo was on the record, the warranty date was set and the invoice was queued for Monday morning. Nobody at the office was awake. He found out on Monday that it had already gone out. She gets one message at 6:30 each morning — an account of what was written overnight and the one thing that is actually stuck.

It is not a queue of things waiting on her. In August, 3 of 151 open opportunities were stale instead of 61 of 148. Median last-activity went from nine days to same day. Time from install to invoice went from eleven days to one.

Nobody learned a new piece of software. "They all kept the AI they already liked," she said. "The only thing that changed is that the CRM is true now, and I stopped being the person who finds out on Thursday."