Employee Turnover Rate Formula and How to Predict It with AI
Get the employee turnover rate formula with worked examples, then use AI to predict attrition, check the model, and explain the numbers to leaders.
For HR and Executives

What is the employee turnover rate formula?
The employee turnover rate is the number of employees who left during a period, divided by the average number of employees during that period, multiplied by 100. Average headcount is usually (headcount at start + headcount at end) / 2. If 180 people left a company that averaged 1,180 employees, turnover is 180 / 1,180 × 100, or about 15.3%. Retention rate is the mirror image: 100% minus turnover.
Predicting attrition is a separate job. It means estimating who, or which groups, are likely to leave next, using the patterns in your past departures. Both are covered below, with prompts you can paste into an AI tool.
TL;DR
- The core formula is separations divided by average headcount, times 100. Use the same period (month, quarter, year) for both numbers.
- To compare a quarter with an annual figure, annualize it by multiplying a quarterly rate by 4 (or a monthly rate by 12), and label it as annualized.
- A single turnover number hides the story. Split it into voluntary vs. involuntary, regrettable vs. non-regrettable, and first-year vs. tenured.
- AI is good at writing the formulas, cleaning the export, and drafting the explanation. It is not safe for unchecked arithmetic, so make it show code or formulas and reconcile the totals yourself.
- A simple baseline (historical rate by tenure band, applied to current headcount) often gets you most of the way. Build a model only if it beats that baseline on data it has not seen.
- Use predictions at the team or segment level for planning, not to single out individuals. Involve HR, legal and privacy before you start.
How do you calculate employee turnover step by step?
- Pick the period. Month, quarter or year. Write it down, because every number in the report must use the same one.
- Count separations. Everyone who left in the period, whether they resigned, were dismissed, retired, or their contract ended. Decide up front what you exclude (see the pitfalls below) and keep that rule fixed.
- Get headcount at the start and end of the period. Use the same date logic each time, such as the last day of the month, and the same definition of "employee."
- Average the headcount. (Start + End) / 2. For a more accurate figure in a volatile year, average the 12 month-end headcounts instead.
- Divide and multiply. Separations / average headcount × 100.
- Annualize if needed. Quarterly rate × 4, monthly rate × 12.
Worked example
A business unit starts the year with 1,200 employees and ends with 1,160. During the year, 180 people leave: 130 resign and 50 are let go.
| Item | Calculation | Result |
|---|---|---|
| Average headcount | (1,200 + 1,160) / 2 | 1,180 |
| Total turnover | 180 / 1,180 | 15.3% |
| Voluntary turnover | 130 / 1,180 | 11.0% |
| Involuntary turnover | 50 / 1,180 | 4.2% |
| Retention rate | 100% − 15.3% | 84.7% |
Now a quarter. In Q1, 40 people leave and average headcount is 1,190. The quarterly rate is 40 / 1,190 = 3.4%. Annualized, that is about 13.4%. Say "annualized" on the slide, or someone will compare 3.4% to last year's 15.3% and panic.
The same formula in a spreadsheet
If start headcount is in B2, end headcount in C2 and separations in D2:
=D2 / AVERAGE(B2, C2)
Format the cell as a percentage. For annualizing a quarterly figure, multiply by 4.
Which turnover metrics should you report?
One company-wide rate rarely helps a leader decide anything. These cuts do.
| Metric | Formula | Why it matters |
|---|---|---|
| Voluntary turnover | Resignations / average headcount | Shows what the company could have influenced |
| Involuntary turnover | Dismissals and layoffs / average headcount | Shows hiring quality and restructuring effects |
| Regrettable turnover | Departures of people you wanted to keep / average headcount | Needs a consistent rule for "regrettable," such as a performance rating threshold |
| First-year turnover | Leavers who left within 12 months of hire / people hired in that cohort | Points to onboarding, role fit and hiring problems |
| Turnover by manager, team, location, level | Same formula on a segment | Finds the concentrated problem a company average hides |
| Retention of a cohort | People from a starting group still employed after N months / starting group | Cleaner than turnover for tracking a hiring class |
Note that first-year turnover uses the hired cohort as its denominator, not average headcount. If 200 people were hired and 36 left within a year, that is 18%.
Pitfalls that change the number
- Contractors and interns. Include or exclude them, but decide once and state it.
- Internal transfers. A move from Sales to Marketing is a separation for the Sales segment but not for the company. Count it only in segment views.
- Small segments. Two leavers out of a team of eight is 25%. That is a fact, not a trend. Show the count next to the rate.
- Seasonal businesses. Compare to the same quarter last year, not the previous quarter.
- Reorgs and divestitures. Headcount jumps distort the average. Note them or adjust.
- Definitions that drift. If HR counts retirements and Finance does not, the two numbers will never match. Write the definitions in the template below.
What template should you use for a quarterly turnover report?
Copy this into a sheet or document. Fill the definitions once and reuse them every quarter.
EMPLOYEE TURNOVER REPORT: [Quarter, Year]
Scope: [Entity / business unit]. Includes: [e.g., regular full-time and part-time]. Excludes: [e.g., contractors, interns, internal transfers at company level]
DEFINITIONS
- Separation: [what counts, what doesn't]
- Average headcount: (start + end) / 2
- Voluntary: [resignation, retirement]. Involuntary: [dismissal, layoff, end of contract]
- Regrettable: [your rule]
- Annualized: quarterly rate × 4
HEADLINE NUMBERS
| Metric | This quarter | Annualized | Same quarter last year | Change |
| Total turnover | | | | |
| Voluntary | | | | |
| Involuntary | | | | |
| Regrettable | | | | |
| First-year (cohort) | | | | |
BY SEGMENT (show leaver count AND rate)
| Segment | Avg headcount | Leavers | Rate | Annualized | Vs. last year |
WHAT CHANGED (3 to 5 bullets, each with a number)
-
LIKELY DRIVERS (label as hypothesis or evidence)
-
OUTLOOK (next 2 quarters: expected leavers and range)
-
ASKS / DECISIONS NEEDED
-
DATA NOTES (reorgs, definition changes, data gaps)
-
How do you use AI to calculate turnover without errors?
AI tools are useful here for three things: writing the spreadsheet formulas, cleaning and reshaping an export, and writing the summary. They are unreliable when you ask them to add up hundreds of rows "in their head." A language model predicts text, so it can state a wrong total confidently. Ask for formulas or code, run them yourself, and reconcile.
Protect the data first. Employee-level data is sensitive. Use only a tool your company has approved for HR data. If you must use anything else, remove names and IDs and aggregate first (counts by month and segment, not rows by person).
Paste this prompt, along with an aggregated table or a column layout description:
I'm an analyst calculating employee turnover. I will give you a table with
these columns: [employee_id (anonymized), hire_date, termination_date (blank if
active), termination_type (voluntary/involuntary), department, location, level].
Do the following:
1. Write the spreadsheet formulas (or Python pandas code) to calculate, for
each quarter from [start] to [end]: start headcount, end headcount, average
headcount, separations, total turnover rate, voluntary rate, involuntary
rate, and the annualized rate.
2. Calculate first-year turnover as a cohort: of people hired in each quarter,
what share left within 12 months.
3. List any assumptions you made and any data problems you would check
(duplicates, termination dates before hire dates, missing types, rehires).
4. Give me 5 reconciliation checks I can run to prove the output is right,
for example: start headcount + hires − separations = end headcount.
Do not give me final numbers from reading the data directly. Give me code or
formulas I can run myself.
What to check in the answer:
- The reconciliation identity holds: start + hires − separations = end headcount. If it does not, you have rehires, transfers or bad dates.
- Total separations from the formula match the count HR or payroll reports.
- A segment's leavers add up to the company total.
- Rehired employees are handled the way your definitions say.
How do you predict employee attrition?
There are three levels, and you should go in this order.
Level 1: A baseline with no model
Calculate historical turnover by segment, such as tenure band (0 to 1 years, 1 to 3, 3 to 5, 5+) and level. Apply those rates to today's headcount. If you have 300 people in the 0 to 1 year band and that band historically loses 20% a year, expect about 60 leavers. This is explainable in one sentence, and it is the number any fancier model must beat.
Level 2: A statistical model on historical data
The standard approach is a classification model: for each employee at a snapshot date, predict whether they leave in the next 12 months (or 6 or 3).
Build the dataset. Take a snapshot of every active employee on several past dates (for example, each January 1 and July 1 over three to five years). For each row, record what you knew on that date, then add a label: did they leave within the next 12 months, yes or no.
Common features (inputs):
- Tenure, time in current role, time since last promotion
- Pay relative to the band midpoint, and recent pay change
- Manager change in the last year, or manager's own team turnover
- Performance rating and its trend
- Team size, level, location, remote or on-site
- Engagement or pulse survey results, if you have them at a level where they are legitimately usable
- Internal mobility, such as applications for other internal roles
Models that work: logistic regression is easy to explain and a good start. Gradient boosted trees (such as those in scikit-learn, XGBoost or LightGBM) often fit better and can show which features matter. If you want to predict when someone leaves, not just whether, look at survival analysis methods.
Here is a prompt that gets an AI assistant to do the technical work in a way you can inspect:
Act as a careful people-analytics data scientist. I have a CSV of employee
snapshots. Each row is one employee on one snapshot date, with columns:
[list columns]. The target column "left_within_12m" is 1 if they left within
12 months of the snapshot date, else 0.
Write Python code that:
1. Splits data by TIME, not randomly: train on snapshots before [date], test on
snapshots after it.
2. Trains a logistic regression and a gradient boosting model.
3. Compares both to a baseline that predicts the historical rate by tenure band.
4. Reports ROC AUC, precision and recall in the top 10% of risk scores, and a
calibration check (predicted vs. actual leaving rate by risk decile).
5. Shows the top features for each model in plain English.
6. Excludes protected attributes (age, gender, race, disability, etc.) and warns
me about any remaining feature that could act as a proxy for them.
Explain each step in plain language I can repeat to HR leadership.
Level 3: Forecasting for planning
For headcount planning, you rarely need individual scores. Add up the predicted probabilities across a team to get an expected number of leavers. A team of 40 whose probabilities sum to 6.2 should expect about six departures, with a range. That is a planning number you can defend, and it avoids pointing at named people.
How do you check whether the prediction model is any good?
A model can look impressive and be useless. Check these before showing anyone.
- Time-based test. The model must be tested on later periods than it trained on. A random split lets the future leak into the past and inflates results.
- Beat the baseline. Compare against the tenure-band baseline. If the model is only slightly better, use the baseline.
- Top-decile capture. If the overall rate is 15%, and the highest-risk 10% of employees actually leave at 30% or more, that is useful lift. If it is 17%, it is not.
- Calibration. When the model says 30% risk, about 30% of those people should leave. If not, do not use the probabilities as headcount numbers.
- Leakage. Watch for features that only exist because someone is leaving, such as a final-month pay change, a resignation-related status, or an exit interview date.
- Stability. Retrain on a different year. If feature importance changes completely, the pattern is fragile.
- Fairness check. Compare error rates and flag rates across groups your legal team defines. A model can be accurate on average and still systematically over-flag one group.
What should you never do with attrition predictions?
- Do not use them to make decisions about individuals. Pay, promotion, project and layoff choices based on a leaving score are risky, unfair and, in some jurisdictions, regulated. Employment-related AI is treated as high risk under the EU AI Act, and privacy law such as GDPR limits how employee data can be processed. Ask legal and privacy before you start.
- Do not include protected characteristics, and test for proxies such as postcode or career gaps.
- Do not share individual scores with managers. A manager who hears "Priya is a flight risk" may treat her differently, which can create the outcome you feared.
- Do not confuse correlation with cause. A model that shows people leave after a manager change does not tell you the change caused it.
- Do not trust an AI tool's confident story about "why people leave." Treat any driver as a hypothesis, then check it with exit interview themes, survey comments and conversations.
Use the output where it helps: choosing which teams to run stay conversations in, where to review pay bands, where onboarding needs work, and how many hires to plan for.
How do you explain turnover numbers to leaders?
Leaders ask for the number. Analysts end up defending it. A short, consistent structure helps. This prompt turns your table into a draft summary:
Here is our quarterly turnover table: [paste aggregated table].
Write a one-page summary for senior leadership with:
1. A two-sentence headline: the overall number, annualized, and how it
compares to the same quarter last year.
2. The three segments with the biggest change, each with leaver count and rate.
3. What we know (evidence from the data) versus what we are guessing.
4. Expected leavers next two quarters, with a range, based on [baseline method].
5. Two or three specific decisions or actions to consider.
Use plain language. Do not invent causes. If a segment has fewer than 10 leavers,
say the rate is based on a small number.
Then read it against your own table. Check every number, remove any cause you cannot support, and make sure the definitions are stated. When someone asks "why is it different from the number Finance showed?", you can point to the definitions block.
Can you build your own turnover dashboard instead of rebuilding the spreadsheet every quarter?
Many teams recalculate turnover by hand each quarter, then paste it into slides. A small internal app could do it for you: upload or enter monthly headcount and separations, calculate every rate in the template above, and show segment views. With Remy, you describe that tool in plain English and it builds and deploys a full-stack app, including a database and sign-in with role-based access, so a manager could see only their own team's numbers. Because HR data is sensitive, involve IT and privacy before putting any employee-level data into a new tool, and consider starting with aggregated counts only.
Frequently Asked Questions
What is a good employee turnover rate?
It depends heavily on industry, role type and location, so any single "good" number is misleading. Retail, hospitality and call centers typically run far higher than professional services or utilities. Compare yourself to your own history and to published benchmarks for your sector and region, using the same definition of turnover.
Should I use average headcount or ending headcount in the formula?
Use average headcount. Ending headcount understates turnover in a year when you shrank, and overstates it when you grew. For volatile years, average the 12 month-end headcounts instead of only the start and end points.
What is the difference between turnover and attrition?
People often use them interchangeably, and the formula is the same. Some organizations use "attrition" for departures that are not backfilled (roles that disappear) and "turnover" for departures where the role is refilled. Whichever you choose, write the definition in your report so readers do not assume otherwise.
How much data do I need to predict employee attrition?
More than most people expect. As a rough guide, you want several years of history (the body suggests snapshots over three to five years) and enough leavers per period that the model can learn patterns, not memorize individuals. If the model is only slightly better than a segment baseline, use the baseline.
Can I paste employee data into ChatGPT or another AI tool?
Only if your company has approved that tool for HR data and its terms cover how the data is stored and used. Otherwise, work with aggregated, anonymized tables, or ask the AI to write the code and run it yourself on your own systems. When unsure, ask your security or privacy team before pasting anything.
Useful for
How we source, write, and correct our articles. Editorial standards
Related articles

Will AI Replace Jobs? What Hormozi Says Workers Should Do
Performance Improvement Plan Template (Plus an AI Prompt)

Are managed database services worth it for small teams?
