When 1:1s Aren't Enough: The Notion-Airtable System That Detects Flight Risks 90 Days Ahead

This morning, your best ML engineer accepted another offer. Two months ago, they told you they were happy, and just three weeks prior, they mentioned wanting to learn more about LLMs. However, five days ago, they stopped committing code at 11 PM like they always did. The signs were there. You just didn't catch them in time.

Photo: Marvin Meyer on Unsplash

The problem with talent retention in AI isn't that people leave. It's that they exit quietly, and by the time you find out, they've already signed with someone else. Startups often try to solve this issue with more 1:1s, better compensation, or additional equity. But in my experience, the reality is more complex. You need an early detection system that captures faint signals before they grow into resignation letters. This article dissects how to build that system, using Notion as a layer of human intelligence and Airtable as a motor for quantifiable signals.

The Anatomy of a Silent Departure: What to Actually Measure

The most common mistake is thinking that retention is solely about satisfaction. It’s not. It’s about progressive engagement and cumulative disconnection signals. A happy but bored engineer will leave. Conversely, a challenged engineer who feels unrecognized will also exit.

The three categories of signals that matter:

Technical Signals (What Airtable Captures Automatically)

• Commit velocity: This doesn't just count how many commits they've made; it also tracks the consistency of their timing. A change in pattern (from late-night to only working hours) is a sign of emotional disengagement.
• Code review participation: When they stop commenting on others' PRs, it indicates they're mentally checked out.
• Slack response time: The average response time in technical channels. A sustained standard deviation of +40% over three weeks is critical.
• Learning budget usage: If they haven’t used their budget for courses or conferences in two quarters, it means they might be learning elsewhere.

Relational Signals (What Notion Documents Qualitatively)

• 1:1 depth: It’s not just about frequency; it’s about depth. Are they sharing real frustrations or just project updates?
• Peer mentions: In retrospectives, do they mention colleagues, or do they only talk about their own work?
• Initiative ownership: Do they propose new ideas, or do they just execute tickets?

External Context Signals (What You Sync Manually)

• LinkedIn activity changes: Have they updated their profile, changed their photo, or added new skills?
• Conference attendance patterns: You notice they are attending more networking events.
• Slack in external communities: If you're part of the same, you’ll see their activity increase.

The Architecture of the System: Airtable as the Quantitative Brain

Photo: Annie Spratt on Unsplash

Airtable will be your relational database for signals. Honestly, don’t use Google Sheets. You need relationships between tables, advanced formulas, and automations that Sheets can’t handle at this scale.

Core Table Structure:

Table 1: Team Members

Essential fields:

• Name, Role, Team, Start Date
• Risk Score (calculated number, 0-100)
• Last High Risk Alert (date)
• Retention Status (dropdown: Green/Yellow/Orange/Red)

Table 2: Technical Signals

Fields:

• Member (linked record to Team Members)
• Signal Type (dropdown: Commit Velocity, PR Activity, Slack Response, etc.)
• Value (number)
• Baseline Average (number - average over the last 90 days)
• Deviation % (formula)
• Date Recorded
• Anomaly? (checkbox - checked if Deviation > 30%)

Deviation % Formula:

IF(
  {Baseline Average},
  ((({Value} - {Baseline Average}) / {Baseline Average}) * 100),
  0
)

Table 3: Qualitative Signals

Fields:

• Member (linked record)
• Signal Category (1:1 Quality, Initiative Level, Peer Engagement)
• Notes (long text - link from Notion here)
• Sentiment Score (1-5)
• Date
• Logged By (manager's name)

Table 4: External Context

Fields:

• Member (linked record)
• Source (LinkedIn, GitHub public activity, conference attendance)
• Activity Description
• Risk Weight (low/medium/high)
• Date Detected

Crucial Automation:

Use Airtable Automations (not Zapier for this; it's native and more reliable) to:

1. Calculate Risk Score weekly: a weighted average of technical anomalies (50%), qualitative sentiment (30%), and external context (20%).
2. If Risk Score > 60, change Retention Status to Orange and send a notification to Slack.
3. If Risk Score > 75, change to Red and create a task in Notion for immediate intervention.

Notion as the Intelligence Layer: Turning Data into Decisions

Airtable gives you the numbers, while Notion provides the context for action. This is where you document the complete story of each individual and plan specific interventions.

Notion Page Structure:

Main Database: People Profiles

Each person has a page with:

Section 1: Current State

• Linked database view from Airtable (using native integration) showing their current Risk Score.
• Visual timeline of signals over the past 90 days.
• Status: Green/Yellow/Orange/Red (synchronized from Airtable).

Section 2: 1:1 Archive

• Sub-page for each 1:1 with structure:
  • Mood check (emoji + brief note).
  • Key topics discussed.
  • Concerns raised (tagged).
  • Action items agreed.
  • Energy level (1-5).
  • Manager observations (private).

Section 3: Growth & Challenges

• Current projects with level of challenge (1-5).
• Skills they want to develop (list).
• Learning investments made (paid courses, conferences).
• Mentorship relationships (who mentors them, whom they mentor).

Section 4: Intervention Log

• When the Risk Score rises, document here:
  • Signals detected (bulleted with links to Airtable records).
  • Hypothesis about the cause.
  • Planned intervention.
  • Date executed.
  • Observed outcome.

Section 5: Retention Lever Inventory

This is key. For each person, document which levers work:

• Compensation: Do they care? How much?
• Equity upside: Do they understand the value? Do they prioritize it?
• Technical growth: What specifically motivates them to learn?
• Impact visibility: Do they need public or private recognition?
• Autonomy: Do they thrive with freedom or need structure?
• Team dynamics: Who do they work best with?

Secondary Database: Retention Playbooks

Create playbooks for scenarios:

Playbook: "Senior Engineer Showing Technical Disconnection Signals"

• Signals pattern: Commit velocity -35%, PR comments -50%, 1:1 depth declining.
• Common causes: Technical boredom, lack of challenge, external opportunities.
• Actions:
  1. Deep 1:1: "I've noticed you're less active in code reviews. What's going on?"
  2. Pitch a new, challenging project (always have 2-3 in the pipeline).
  3. Offer time for an internal side project (20% time).
  4. Connect with an external mentor if they seek growth outside the current scope.
• Timeline: Act within 5 days of detecting the pattern.

Playbook: "ML Researcher Showing LinkedIn Activity Signals"

• Pattern: Updated profile + attending conferences + less internal initiative.
• Causes: They are being actively recruited, evaluating options.
• Actions:
  1. Don’t confront directly (they’ll shut down).
  2. Create an opportunity for them to speak: "How do you see your career in the next 2 years?"
  3. Show the company’s technical roadmap, emphasizing their critical role.
  4. Accelerate equity cliff if they are close (bring vesting forward).
  5. Propose a research paper collaboration with their name as the first author.
• Timeline: Act within 3 days; they are likely already in late stages.

The Weekly Workflow: Turning the System into Habit

A system only works if you use it consistently. This is the ritual that truly makes a difference:

Monday 9:00 AM - Review Risk Scores (15 min)

• Open filtered view in Airtable: Risk Score > 50.
• Review who rose this week and why (specific signals).
• Note in Notion who needs a check-in this week.

Wednesday - Data Capture (30 min)

• Update Technical Signals with data from the last week (automate where possible via API).
• GitHub commits: use GitHub CLI + Python script that writes directly to Airtable.
• Slack response time: Slack Analytics or custom script.
• PR activity: GitHub API as well.

Friday Post-1:1s - Qualitative Update (20 min)

• After each 1:1, immediately update that person's Notion page.
• Summarize sentiment, concerns, energy.
• If you detected something out of the ordinary, create an entry in the Intervention Log.
• Update Qualitative Signals in Airtable with sentiment score.

Monthly - Retention Lever Review (1 hour)

• Review each person's Retention Lever Inventory.
• Ask yourself: Did anything change? Did I discover a new lever?
• Update playbooks if you found a new pattern.

When Signals Turn Red: The Intervention Protocol

You've detected it in time. Now, what specifically do you do?

Phase 1: Confirmation (first 48 hours) Don't act solely on data. Confirm your hypothesis:

• Casual but direct 1:1: "I noticed X and Y. Is everything okay?"
• Listen without defending or justifying.
• Take detailed notes in Notion, including direct quotes.

Phase 2: Root Cause (Days 3-7) Dig deeper if they confirm that something is wrong:

• "What would need to happen for this to improve?"
• "If you had a magic wand, what would you change?"
• Look for whether it’s situational (current project) or structural (role, team, company).

Phase 3: Custom Intervention (Days 8-14) Based on that person's Retention Levers, design a specific plan:

• If it’s growth: a new project or role expansion.
• If it’s recognition: visibility with founders, public win.
• If it’s compensation: equity acceleration or off-cycle raise.
• If it’s team dynamics: a change of squad or reporting line.

Phase 4: Follow-through (Days 15-90) The fatal error is to intervene and forget:

• Weekly check-ins for the first 30 days.
• Measure if the Risk Score drops.
• If it hasn’t improved in 30 days, the intervention didn’t work; you need a different approach.
• Document everything in the Intervention Log.

The Metrics That Truly Matter: What to Track

Your system is only as valuable as its predictions. Track this monthly:

Leading Indicators (Prediction):

• % of individuals with Risk Score > 60 who left within the following 90 days (if it works, it should be ~70%+).
• Average days between the first red signal and the intervention (target: < 7 days).
• % of interventions that reduced Risk Score > 20 points in 30 days.

Lagging Indicators (Outcome):

• Voluntary turnover rate (especially among top performers).
• Time to detect (days between actual turnover and the first retrospectively detectable signal).
• Retention rate 12 months post-intervention.

If your system detects but you don’t intervene, or you intervene late, or you intervene poorly, none of these numbers will improve. Early detection without action is just organized anxiety.

What This System Doesn’t Solve (And Why It Matters to Know)

Be honest with yourself: there are departures that no system can prevent.

If your compensation is 40% below market, if your tech stack is legacy while the industry has moved on, if your culture is toxic, or if there’s no real career path, no dashboard will save you. This system detects and gives you time to act. But it will only work if you have something to offer when you intervene.

There are also individuals who simply need to move on. They are looking for something your startup can’t provide today: more seniority, a different expertise, or international remote work.