It is a tempting feature request: "We already have all the CRM and ERP activity. Can we flag the people who are about to leave?" Logins dropping, fewer updates, slower replies. It sounds reasonable. I think it is one of the worst things you can build on top of business data, and there is a more useful thing to build instead.

Some context. I posted a list in Spanish on LinkedIn and expanded it on EGA Futura Discover: ten signs that it is time to leave your job. Stopped learning, a toxic atmosphere, work nobody sees, constant exhaustion, a boss who controls everything, good ideas that annoy people. The illustration said it in one line: your dashboard shows everything except who is about to quit. People can hit every target, update Salesforce every day and smile in meetings while they have already made up their minds.

Why per-person attrition scores go wrong

The proxies are weak. Fewer logins can mean someone is checked out. It can also mean they automated their own work, moved to a different tool, or are on vacation. The model has no way to tell, and the person never gets to explain.

It turns metrics into surveillance. The moment people suspect activity is scored, activity becomes theater. You get more clicks and fewer honest signals, which is Goodhart's law in its most predictable form.

It points at the wrong layer. Most of the ten signs are not about the employee at all. "My boss needs to approve everything" and "every improvement idea gets 'we have always done it this way'" are properties of the system and of management. Scoring individuals hides that.

What to measure instead: team-level friction

The data you already have can show where the process wears people down, as long as you aggregate and set a floor on group size. Three metrics I find honest:

  1. Approval load for small decisions. How many approval requests per team fall under a low threshold, and how long they wait.
  2. Unanswered improvement proposals. Ideas older than 30 days with no response, per area.
  3. Manual rework. Records edited by hand after an automated step, per process.

Here is the second one in plain SQL, with a minimum group size so nothing can be traced back to one person:

SELECT area, COUNT(*) AS unanswered_ideas, COUNT(DISTINCT author_id) AS authors FROM improvement_proposals WHERE status = 'new' AND created_at < CURRENT_DATE - INTERVAL '30 days' GROUP BY area HAVING COUNT(DISTINCT author_id) >= 5; 

The HAVING clause matters more than it looks. An area with two people and four unanswered ideas is not a metric, it is a name. Suppress it and handle it in a conversation.

A few design rules I would apply to any version of this:

  • Group by area or process, never by employee.
  • Publish the metric definitions to the people being measured.
  • Report to the manager who owns the process, and make the manager the one accountable for moving the number.
  • No prediction column. The output is "this process creates friction", not "this person might leave".

The part no query covers

None of this tells you that someone stopped enjoying their work six months ago. A dashboard full of green can coexist with a team that is halfway out the door. What the metrics can do is remove the excuses on the system side, so the conversation with each person is about them and not about a broken approval chain.

I run EGA Futura, a native ERP on Salesforce, and co-founded Vantegrate, an AI agent company, with Francisco Morales. If your people data lives in Salesforce, EGA Futura People (Spanish) keeps each employee's record in the same org as the rest of the operation, and our notes on management for mid-sized companies (Spanish) cover the management side.

The query above returns areas and counts. The names stay out of it, and the conversation stays with a person.