How to Calculate the ROI of AI-Supported Mental Health Tools

If you are building a business case for Finance, the honest answer is that there is no single ROI number you can drop into a slide and expect it to survive review. Every credible study on mental health investment, including studies Unmind itself has run, returns a different multiple because the inputs are different: workforce size, baseline stress levels, existing benefit spend, geography, and what "support" actually consists of. A number imported from someone else's workforce is not evidence. It is a guess wearing a decimal point.

What does survive a Finance review is a range built from two things: your own cost stack, and impact data from studies that show their methodology. This piece walks through both halves, using verified outcomes from case studies Unmind has published, so you can build a model Finance will actually sign off rather than push back on.

Unmind has published its own ROI position at unmind.com/roi, and the position there is not "here is the number." It is that ROI depends on what you combine. Self-help content and Therapy & Coaching used together return a different multiple than either used alone, and that multiple only holds for organizations with a comparable starting point.

The same logic applies to any vendor's benchmark, including this one. A multiple calculated across a six-month study with a defined high-stress cohort is not a universal constant. It is a data point from a specific population, doing specific things, over a specific window. Treat every published multiple, including the ones below, as evidence you can adapt rather than a number you can borrow.

A companion piece from Unmind on the typical per-employee cost of turnover and absenteeism from poor mental health makes the same argument from the cost side. Building a defensible range starts there, with your own inputs rather than an imported number.

Before any impact figure means anything, you need your own baseline. Finance will ask for this before they ask for the ROI multiple, so build it first.

The inputs that matter:

• Voluntary turnover linked to burnout or mental health strain. Pull exit interview data and manager attrition notes for a mental-health-attributable share, then cost it against your own recruitment, onboarding and lost-productivity figures per role.
• Absenteeism and short-term disability days coded to mental health conditions. Your own occupational health or disability carrier data will have this broken out, or can be requested broken out.
• Presenteeism, meaning people at work but underperforming due to stress, low mood or burnout. This is harder to isolate but is typically the largest line in the stack once you have a productivity-impairment proxy.
• Healthcare claims tied to mental health diagnoses, pulled from your carrier or PBM data where available. This is the line most CHROs underuse in a business case, largely because it sits with Benefits rather than HR and the two data sets rarely get combined for this purpose.

The full methodology for building this cost stack, including how to work with incomplete claims data, is in the companion piece linked above. Build that model first. It is the denominator every impact figure below needs to be measured against.

Once you have your own cost stack, you need impact evidence to apply against it. Here is what Unmind can verify, each tied to a specific, scoped study rather than presented as a universal truth.

The combined-tools result

The published position from Unmind is a 4.6x return for every dollar invested when self-help content and Therapy & Coaching are used together, rather than either on its own. This is the figure to anchor a combined-platform business case to, and it is directly relevant if your model assumes employees will move between self-guided support and one-on-one care rather than staying in one lane.

Two customer studies worth anchoring to

Neither needs its subject named for the numbers to hold up. One employer ran a six-month study giving its highest-stress employees access to the Unmind platform and found a 16% reduction in stress and a 10% reduction in overall work impairment, translating into a projected 3.49x ROI and an estimated $1M financial impact from the improved productivity. This is the closer analog if your own workforce has an identifiable high-stress segment, because it isolates that cohort rather than averaging across an entire population. Another employer tracked mental health scores over five years rather than a single snapshot, and found gains of 26% in the UK, 21% in the US, and 18% in Australia. The value of that second data point for a Finance conversation is less the percentage itself and more what it demonstrates: that outcomes compound over a multi-year commitment rather than resetting each renewal cycle, which matters if Finance is asking about payback period rather than a single-year return.

Impact data only converts into savings if people actually use what you buy. This is the variable most legacy EAP contracts hide, because their commercial model does not depend on engagement.

The category context: average EAP engagement runs under 2%, and legacy EAPs typically sit at 2 to 4% utilization, functioning closer to a compliance line than a used benefit. If your current contract is in that range, your realistic engagement assumption for a new model should start from where you are today rather than from a vendor's best-case figure.

Nova is the AI mental health agent Unmind built to raise that baseline. It gives employees quick, confidential support and connects them to human care, including Therapy & Coaching, when that is what they need. Nova does not provide diagnosis, treatment or crisis intervention, and it is not a substitute for the clinicians who deliver therapy and coaching at Unmind. What it does is lower the barrier to a first conversation, which is where most engagement in this category is lost. Nova users show 49% monthly retention compared with 39% for employees who do not use it, meaning people who start with Nova are more likely to keep engaging with the platform over time, which is the leading indicator your lagging cost-stack metrics depend on.

For your model, this means engagement is not a fixed input. It is the variable that determines which end of your range you land on. A higher, sustained engagement rate moves you toward the upper bound of your impact assumptions. A rate closer to the legacy EAP baseline moves you toward the lower bound.

1. Baseline your cost stack. Use the categories above (turnover, absence, presenteeism, mental-health-coded claims) and your own data. Use the companion piece for the detailed methodology.
2. Set a conservative and an aggressive engagement assumption. Anchor the low end to the legacy EAP baseline of 2 to 4%. Anchor the high end to what a comparable organization achieved in a study with a similar cohort, such as the first example above, rather than a best-case industry claim.
3. Apply the closest impact analog rather than the biggest number. If your workforce has an identifiable high-stress segment, use the first example's cohort-specific figures. If your case is about sustained investment over multiple years, use the second example's trajectory. If your case is about the value of combining self-guided tools with clinical care, use the 4.6x figure.
4. Present the range rather than a point estimate. Finance will trust a bounded range with stated assumptions far more than a single confident number with no visible working. Show your low case, your high case, and the specific inputs that move you between them.

A model built before rollout is an estimate. Once you are live, Insights & Assessments gives you real-time dashboards on engagement, stress signals and wellbeing trends across teams, benchmarked against wider data, so you can true up your projected range against what is actually happening rather than defending a number you calculated a year earlier. That is also the data set you bring back to Finance at renewal, replacing a projection with a result.

If you are building this case now, start with your own cost stack using the companion breakdown of turnover and absenteeism costs, review the ROI position Unmind publishes for how the combined-tools multiple is built, and look at Nova for how engagement gets built into the model from day one rather than measured after the fact.