Before you pick between "ratings vs. no ratings", ask what decision the score has to support. If a review must calibrate people across leads, defend a promotion, and inform up-or-out and staffing calls, you need a defensible rating. If it only drives development, continuous feedback without a number can be enough. For most project-based firms the honest answer is "both": a light, calibrated rating for the decisions that need differentiation, plus continuous feedback for growth.
Many big-name firms learned the hard way that dropping ratings leaves the underlying judgment in place and moves it somewhere less visible and less fair. When CEB studied companies that removed ratings, employee performance fell about 10% on average and the quality of manager conversations dropped 14% (CEB, 2016). Almost all employers still use a rating in some form (Mercer, 2025). So start from the decision and choose the format after.
Ratings, no ratings, or both: what does each model mean?
Three models, defined plainly. A ratings model attaches a score to the review: a number or a label ("meets / exceeds", a 1–5, a grade) that summarises performance against a bar and lets you compare people. A no-ratings model drops the score and keeps the substance: narrative feedback, frequent check-ins and documented conversations, with no summary number. A hybrid keeps a calibrated rating for the decisions that need it and runs continuous feedback alongside it for development.
Arguing about the format in the abstract gets you nowhere. A rating is an output that some decisions require and others don't. Deloitte made this move when it rebuilt its system: it separated the pay decision from the day-to-day performance conversation, and produced a quick per-project "snapshot" plus weekly check-ins (Buckingham & Goodall, Harvard Business Review, 2015). The same firm ran both mechanisms, one to differentiate and one to develop. That is the design question underneath "ratings or not".
Make it concrete with one person. Picture a senior designer staffed on three client projects across a year. Under a ratings model, each lead scores her against a shared bar and the firm calibrates those reads into a single number that feeds her promotion case. Under a no-ratings model, each lead leaves documented feedback and running check-in notes, and nobody produces a summary grade. Under a hybrid, she gets both: one calibrated rating for the promotion decision, plus a stream of feedback she can act on between projects. Same designer, same year, and the firm can do three different things with what her leads saw.
Why did so many firms drop ratings, and what happened next?
The annual rating had become expensive theatre. Deloitte found it was spending close to 2 million hours a year on performance management, and 58% of its executives said the approach drove neither engagement nor high performance (Buckingham & Goodall, HBR, 2015). Adobe abolished the annual review and stack-ranking in 2012, replacing them with lightweight "Check-ins" and reclaiming an estimated 80,000 manager-hours a year (Adobe, via Stanford GSB, 2016). Microsoft ended its forced "stack ranking" in 2013 because grading people on a curve pitted colleagues against each other (Microsoft, via CNN/SHRM, 2013). General Electric went further still: in 2015 it began retiring a decades-old five-point rating scale (from "role model" down to "unsatisfactory") in favour of an app-based continuous-feedback tool, PD@GE (GE, via Fortune, 2015). The wave looked like the end of ratings.
Then the evidence came back the other way. CEB studied organisations that had removed ratings and found the opposite of what was promised: employee performance dropped about 10% on average, the quality of employees' performance conversations with managers fell 14%, and the share of people who believed the firm connected performance to pay fell 8%, with high performers hit hardest (CEB, 2016). The judgment stayed. Managers still decided pay and promotions, with less structure and less visibility. Removing the number had left the decision in place, so most firms kept a rating: as of 2025, almost all employers still use one, and only 8% force a distribution (Mercer, 2025).
What do ratings do well, and where do they fail?
Ratings do one thing no narrative does: they let you compare and calibrate. When you have to decide who gets promoted, who's ready for partner, who to keep on the bench and how to split a raise pool, a shared score on a shared bar makes the decision defensible: you can show the person, and a tribunal, how you made the call. A number also forces a manager to commit to a judgment rather than hide behind warm, non-committal prose.
Ratings fail on precision. A rating carries a lot of the rater and not much of the performance. In a study of 4,492 managers, about 62% of the variance in ratings was idiosyncratic to the individual rater, while only 21% reflected actual performance (Scullen, Mount & Goff, Journal of Applied Psychology, 2000). A single reviewer is noisy: the mean interrater reliability of one supervisor's rating of overall performance is about .52 (Viswesvaran, Ones & Schmidt, Journal of Applied Psychology, 1996). Add recency and halo bias and rating anxiety, and a lone number can be both falsely precise and unfair. Keep the rating and calibrate it across raters, so the score reflects the work instead of who did the reviewing.
When does a no-ratings model work, and when does it break?
A no-ratings model works when the review's only job is development and three conditions hold. First, the team is small and high-trust, so feedback flows without a score forcing it. Second, the managers are strong: able to have direct, specific conversations and make pay calls fairly without a number to hide behind. Third, there is no hard differentiation to defend: no up-or-out, no partner gate, no tight raise pool that you have to split and justify. Strip the rating in that setting and you lose little. You keep the substance and drop the theatre. Gap Inc. is the textbook version at scale: it scrapped annual reviews and ratings for "Grow. Perform. Succeed.", frequent coaching conversations and short goal cycles whose whole job is development (Gap Inc., via e-reward case study). A ten-person design studio where the founder already knows everyone's work is the small-team version of the same bet.
It breaks the moment a real differentiation decision lands. Without a shared score, managers struggle to explain how someone performed or why a pay decision went the way it did. That is why conversation quality and the perceived pay-performance link fell when firms removed ratings (CEB, 2016). The judgment stays too: it goes underground, into a manager's private ranking that no one can see, calibrate or challenge. That is less fair than an explicit rating. If your firm has to promote, staff and defend up-or-out, pure no-ratings moves the hardest calls somewhere with no audit trail.
Why does "both" fit most agencies and consulting boutiques?
A project-based firm runs both jobs at once, and a single model can't do both well. Your people are staffed per engagement and rated by different leads, and those reads have to roll up into promotion, raises, bench decisions and up-or-out on a partner track. That is a differentiation problem. It needs a calibrated rating, or one lead's generosity and another's severity decide someone's career. The same people also need to grow between cycles, and a once-a-year grade is useless for that. Growth needs continuous feedback close to the work.
So the honest answer for most boutiques and agencies is a hybrid: a light, calibrated rating on a shared bar for the decisions that require differentiation, plus ongoing feedback and check-ins that carry development. Keep the two separate. The rating is the calibrated judgment of the past, and the feedback is the forward-looking growth conversation. Kept apart, the number can't poison candour and nobody can renegotiate the coaching into a grade. Deloitte's split of the pay decision from the weekly conversation is the same idea (Buckingham & Goodall, HBR, 2015), and CIPD frames performance management as a continuous cycle you fit to your own context rather than a single annual event (CIPD, Performance Management). Going fully no-ratings removes signal your partner track still needs. Making the rating the whole story punishes the candour your feedback culture depends on, so most firms do best with both, kept apart.
In practice, that can be modest. A 25-person data-consulting boutique might keep a simple 1–4 rating, calibrated across the leads who staffed each consultant, only to make partner-track and raise-pool calls defensible. It would run fortnightly check-ins that never touch that number, so coaching stays honest. A creative agency might do the same with a "below / at / above bar" label reconciled in a 30-minute cross-lead calibration, plus a running feedback doc per person. In both, the rating stays light, development runs all year, and neither one turns into the other.
How do you choose a model for your firm?
Work down from the decisions, and leave the format for last.
- List the decisions the review must support. Promotion, partner/up-or-out, raise-pool splits, staffing and bench calls, development. Write them down and let this list pick your model, whatever the fashion.
- Ask if any decision needs differentiation. If you must rank, calibrate across leads, or defend a promotion or exit, you need a rating. If the only output is growth, you don't.
- Check your calibration honestly. A rating is only as fair as the calibration behind it. One lead scoring alone is noise (Scullen, Mount & Goff, 2000). If you can't run a cross-lead calibration, fix that before you trust any score.
- Judge your managers and your trust level. No-ratings leans hard on strong managers and high trust. If either is thin, a shared bar protects your people from inconsistent judgment.
- If you have both jobs, run a hybrid and keep the halves apart. Use a calibrated rating for the decisions that need it and continuous feedback for development. Never let one silently become the other.
- Name where the judgment lives. Whatever you choose, make the differentiation decision explicit and calibrated. If your model pushes it into a manager's head, you have hidden the bias and left it in place.
A quick decision self-check
Score your current model: one point per "yes". Six or more means your model fits your decisions. Four or fewer means you're likely arguing format instead of purpose.
- You've written down the actual decisions your reviews must support.
- You know whether any of those decisions require differentiation (promote, pay, up-or-out).
- If you rate, the score is calibrated across the leads who staffed the person, with no lead scoring alone.
- If you don't rate, no hard differentiation decision secretly depends on a hidden ranking.
- Development feedback runs continuously, not once a year.
- Your rating (if any) is kept separate from the forward-looking growth conversation.
- People can see the bar they're measured against before the cycle, not after.
- The person making a promotion or pay call can explain it against shared evidence.
FAQ
Is it better to have performance ratings or not?
Neither is better in the abstract. It depends on what the review has to decide. If it must differentiate people for promotion, pay or up-or-out, a calibrated rating makes the decision fair and defensible. If it only drives development, continuous feedback without a number can be enough. When companies removed ratings, performance and conversation quality fell (CEB, 2016), because the judgment stayed and lost its structure.
Why did companies like Adobe and Microsoft drop ratings?
The annual rating had become slow and expensive: Deloitte alone spent close to 2 million hours a year on it (Buckingham & Goodall, HBR, 2015). Forced-curve ranking also damaged collaboration, which is why Microsoft ended stack ranking in 2013, GE retired its five-point scale for the PD@GE app in 2015, and Adobe replaced annual reviews with lightweight check-ins in 2012. Others, like Gap Inc., dropped ratings entirely in favour of frequent coaching conversations (Gap Inc., via e-reward). Most firms kept a rating in some form, though. As of 2025 almost all employers still use one (Mercer, 2025).
Doesn't a rating just add bias?
A single rating is noisy. About 62% of rating variance is down to the individual rater rather than the performance (Scullen, Mount & Goff, 2000), and one supervisor's reliability is only around .52 (Viswesvaran, Ones & Schmidt, 1996). Dropping the rating leaves that judgment in place and hides it in a manager's head, where no one can calibrate it. Calibrating the score across multiple leads is what reduces the bias, whether or not you show a number.
What does a hybrid model look like for a consulting boutique?
A light, calibrated rating on a shared bar, built from per-engagement inputs and reconciled across the leads who staffed the person, that feeds promotion, raises, staffing and the partner track. Alongside it, continuous feedback and check-ins carry development between cycles. You keep the two separate so the rating stays honest and the growth conversation stays open.
How do I choose between the models?
Start from the decisions your reviews must support, and set aside what other firms do. If any decision needs differentiation, you need a defensible, calibrated rating. If the only output is growth, you may not. If you have both jobs (most project-based firms do), run a hybrid and keep the rating separate from the development conversation (CIPD, Performance Management).