A hiring manager asks which feedback model to use and expects a short answer: SBI, or the sandwich, or one of the half-dozen acronyms that circulate in management training. The models are real and one of them is genuinely good. They also sit downstream of the thing that actually goes wrong, which is that the panel walked into four interviews without agreeing what they were judging, and no amount of formatting rescues an opinion that was never anchored to anything.
Worth knowing before you pick one: feedback is not reliably helpful. Kluger and DeNisi’s meta-analysis in Psychological Bulletin (607 effect sizes across 23,663 observations, published in 1996 and still the reference point) found an average improvement of d = .41, and found that over a third of feedback interventions made performance worse. Their explanation is the useful part. Effectiveness drops as the recipient’s attention moves up from the task toward the self. Feedback that stays on what was done helps; feedback that arrives as a verdict on who someone is does not, and often does damage.
That single mechanism decides which model to use and how to prepare for it.
The problem is the record, not the phrasing
Run the sequence in a typical panel. Four people interview a candidate across a week. Nobody wrote down what “good” meant for this role beyond the job ad. Each interviewer forms an impression, and by the debrief three days later what survives is a feeling with a reason attached after the fact. Then someone asks for structured feedback, and the model gets applied to material that has no structure in it.
You can watch this in the output. “Strong communicator, but I wasn’t sure about the technical depth” is not underspecified because the interviewer picked the wrong template. It is underspecified because nothing was ever specified. Reformat it into any acronym you like and you get the same sentence in three parts.
So the order runs backwards from how it is usually taught. Decide the criteria, capture against them during the interview, and the model becomes a formatting step on material that is already concrete, which is also, per Kluger and DeNisi, exactly the condition that keeps attention on the task rather than the person.
The method, in the order it has to happen
Write the criteria before the first interview. Four or five, specific to the role, phrased as things you could observe: “explains a technical trade-off to a non-technical listener”, not “communication”. Every interviewer rates the same ones. This is the step teams skip, and skipping it is what makes the rest cosmetic.
Capture during the call, not after. Notes written three days later are reconstructions, and a reconstruction reliably drifts toward the overall impression the writer already holds. One line per criterion, written while the candidate is still talking, is worth more than a paragraph written on Friday.
Then format with SBI. Situation, behaviour, impact — developed by CCL, the US leadership-development non-profit — is the structure worth learning: name the situation, describe the observable behaviour, state the impact it had. It works for the same reason the meta-analysis predicts it should — each of its three parts is a fact about an event, so the sentence has nowhere to drift toward character. CCL’s own extension adds intent as a fourth part, which turns the delivery into a question rather than a verdict and is worth having when the feedback goes to a colleague rather than a candidate.
An example of the difference. “You were a bit disorganised in the interview” is a judgement with no event in it. “In the system-design round you started the schema before the requirements, and I couldn’t tell which constraints you were designing for” names a situation, a behaviour and an impact, and the candidate can do something with it.
The sandwich, and an honest reading of the evidence
The praise-criticism-praise sandwich is the model most often recommended and most often attacked. The case against it is that the framing puts attention on how the recipient is being assessed rather than on what happened, which is the failure mode the 1996 meta-analysis identifies, and that the surrounding praise gets read as a delivery vehicle rather than as praise.
The case against it is not as settled as the internet suggests. A 2020 experimental study in Learning and Instruction found better subsequent task performance after sandwich feedback than after corrective feedback alone or none at all. The honest summary is that the empirical base is thin either way, and that the sandwich’s real problem is less that it fails than that it is used as a substitute for having anything specific to say.
If you want a rule that holds: use SBI for anything a person will act on, and stop constructing sandwiches around feedback you have not yet made concrete.
What it costs
Twenty minutes per role to agree the criteria, and a few minutes per interview to write against them. The recurring cost is the argument in the first debrief, when someone’s confident impression turns out not to map onto any criterion the panel agreed. That argument is the point. It is cheaper than a mis-hire and it only happens once per team.
Disclosure, since it is relevant: at Join we sell an ATS with interview scorecards in it, so weigh what follows accordingly. The software does one thing here: it holds the criteria before the interview, gives every interviewer the same ones with notes per criterion, and aggregates the panel’s scores on the candidate profile so disagreement is visible rather than averaged away in a meeting. It does not currently give you a grid across candidates; comparing two people means clicking between two profiles.
None of that is the part that matters. The criteria on a shared document, agreed before anyone interviews, get you most of the benefit for the price of a meeting. If your feedback is vague today, a tool will render the vagueness more neatly. Decide what you are judging first, and the model you pick afterwards will matter much less than the argument suggests.
Once the record exists, the harder half is the other direction: taking the feedback your own process generates without arguing with it.
