The ROI of AI Coaching: What to Measure and What to Expect

Read ten comparisons of AI coaching tools and you will learn a great deal about features: session formats, integrations, avatar quality, pricing tiers. You will learn almost nothing about the question your CFO will actually ask: what does this return?

It is a strange gap. Coaching is one of the better-studied interventions in workplace psychology, and the arrival of AI changes its economics more than it changes anything else. So instead of another feature list, let us do the math — starting from the research, not the brochures.

What does coaching return before AI enters the picture?

Two meta-analyses anchor the field.

Theeboom, Beersma and van Vianen (2014) pooled workplace coaching studies and found significant positive effects across five outcome categories — performance and skills, well-being, coping, work attitudes, and goal-directed self-regulation — with effect sizes ranging from g = 0.43 to 0.74. In plain terms: moderate to large effects, across the board.

Jones, Woods and Guillaume (2016) found an overall effect of δ = 0.36 for workplace coaching — respectable — but the headline sits one level deeper: on individual-level outcomes, the effect climbs to δ = 1.24. That is a very large effect by any behavioral-science standard. Coaching most reliably moves the coached person's own outcomes: their goals, their skills, their behavior. Effects on team and organizational metrics exist but are smaller and noisier, because more variables sit between a manager's behavior and a P&L line.

One caution before we go further. You will meet claims like "coaching returns five to seven times its cost." Treat them with suspicion: such multiples usually trace back to self-reported estimates from people who commissioned the coaching, not to controlled studies. The defensible position is narrower and stronger: coaching produces robust, well-documented behavioral effects at the individual level. The dollar multiple depends on whose behavior changes and what that behavior touches.

What does AI change in the equation?

Think of coaching ROI as a fraction: (value of the behavior change × people reached × probability it actually happens) divided by cost. AI moves all four terms.

Cost. A human executive coach bills by the hour, and the meter shapes everything — session count, who qualifies, how long programs last. With an AI coach, the marginal cost of one more session is close to zero. That single fact rewrites the denominator.

Reach. Traditional coaching is rationed: a handful of executives get it, the managers who shape most employees' daily experience do not. When cost per person collapses, the entire management layer becomes coachable — and the individual-level δ = 1.24 starts applying to fifty people instead of five. This is where most of the untapped ROI actually sits, because a first-time manager with three raw skills gaps has more room to improve than a polished executive.

Frequency and the space between sessions. Behavior change happens between sessions, not during them — and between sessions is precisely where a human coach cannot be. An AI coach is available Tuesday at 7 a.m., twenty minutes before the difficult conversation, for exactly as long as needed — we have written about why session length works differently with an AI coach.

Effectiveness. The equation only improves if the effect holds. The best evidence so far: Terblanche and colleagues (2022) ran two longitudinal randomized controlled trials and found an AI coach matched human coaches on goal attainment, with both beating the control group. The scope matters — this is structured, goal-focused coaching, not the full range of executive work; Graßmann and Schermuly (2021) map what AI coaching can and cannot yet do. We review the whole evidence base in does AI coaching actually work.

And one multiplier sits on top of everything: usage. An unused subscription has an ROI of exactly zero. Which is why measurement starts there.

What should you actually measure?

Four levels, from fastest signal to strongest proof.

1. Engagement. Weekly active usage, session depth, retention at 90 days. Necessary but not sufficient — engagement is the multiplier, not the outcome.

2. Goal attainment. The best-validated outcome measure in coaching research, and the one the RCTs used. Do it properly: each participant defines two or three concrete goals at the start, with a baseline, and scores progress on a simple scale at 90 days. If you measure only one thing, measure this.

3. Observed behavior change. Ask the people around the coachee, not the coachee. Short 360-style pulses, one question to the team — "what has this person started doing differently?" A warning from the research here: Kluger and DeNisi (1996), analyzing 607 effect sizes, found that more than a third of feedback interventions actually degrade performance. Volume of feedback is not a success metric; quality and follow-through are.

4. Business metrics. Team retention under coached managers, internal mobility, win rates in sales. Real, but slowest and most confounded — treat attribution with humility.

The four levels work as a chain, and each link tests the next. No engagement means nothing downstream will move, so stop and fix adoption first. Engagement without goal progress suggests pleasant conversations rather than coaching — tighten the goals. Goal progress that colleagues cannot see usually means the goals were too internal; make the next round behavioral. Only when the first three links hold is it worth arguing about business attribution at all.

Which traps make coaching ROI numbers lie?

In our coaching work we see the same five mistakes recur.

Satisfaction as outcome. People like coaching. Liking it is not changing.

No baseline. "Rate your progress" asked only at the end invites retrospective inflation — people rewrite their starting point to flatter their journey. Define the baseline before the first session, in writing, while nobody has anything to prove yet.

Attribution greed. Coaching rarely arrives alone — there is also a reorg, a new comp plan, a new boss. Claim the individual-level effects the research supports; be modest about the rest.

Activity dressed as outcome. "1,200 sessions completed" is a cost line, not a benefit. Sessions are input; changed behavior is output.

The wrong window. Behavior shifts show in weeks; business metrics need quarters. Judging a program on revenue after six weeks guarantees a false negative — and running it two years without interim behavioral measures guarantees drift.

What does sales coaching show about ROI?

Sales is the cleanest laboratory for coaching ROI, because outcomes are counted weekly and nobody debates what a win rate is.

The reference numbers come from Korn Ferry: companies with consistent sales coaching and impact measurement see 32% higher win rates and 28% higher quota attainment. The catch has never been whether sales coaching works — it is that frontline managers do not do it. Coaching is the first activity a busy quarter kills.

That is exactly the constraint AI relaxes: rehearsing discovery calls, pre-call preparation, post-call review — at zero marginal cost, on the rep's schedule. We detail the use cases in what an AI sales coach actually does. And the arithmetic is unusually honest here: take a team closing €10M a year at a 20% win rate — you do not need anything close to Korn Ferry's full delta for the program to pay for itself many times over. The deeper point: win rate is measurable, so for once the ROI debate can be settled with your own data instead of a vendor's slide.

What should you realistically expect from a 90-day pilot?

A design we recommend, small enough to run without a committee:

1. Ten to thirty volunteers — opt-in, never conscripted; forced coaching measures compliance, not change.
2. Two goals per participant, defined in week one, with a baseline.
3. Weekly usage, tracked but not policed.
4. At day 90: goal attainment scoring, one question to each participant's colleagues, and the engagement curve.

What to expect if it works: most engaged users report measurable progress on at least one goal; colleagues can name a specific behavior that changed; the usage curve flattens but does not collapse. What not to expect: clean revenue attribution in one quarter — anyone promising that is selling something. For senior leaders the outcome criteria differ enough that we treat them separately in AI coaching for executives.

The feature lists will keep coming. But coaching ROI has always lived in the same place: one person, behaving measurably differently, in situations that matter. Stop reacting. Start seeing.

The cheapest way to estimate the ROI of AI coaching is to run your own 90-day experiment with real goals — practice it with Lumia, your AI coach.