AI ROI evaluation: learn the HARD way

Companies are seriously investing in AI, but is ROI really there to back it up?

Surveys show that, while 91% of organizations actively increase AI investments, only 15% of AI adopters report measurable ROI. For most companies, investing in AI is a question of prestige with some wishful thinking on the side: turns out, nobody wants to stay behind in the long run.

At the same time, AI can deliver much more than a vague sense of being future-proof. Sombra’s own case studies indicate very clearly: if your company has the basic prerequisites to use AI as more than another trendy gadget, you can increase the saving potential up to $2M annually, reduce time spent validating information by 30%, and even shorten delivery times to 2-3 weeks. All thanks to AI integration.

Based on my experience with live deployments, I’ve developed a simple framework to help companies evaluate whether AI adoption will actually generate ROI.

Here’s the HARD framework for you to evaluate AI ROI before you commit to an investment.

Evaluate Human Impact & Cost of Failure

Start with the H of human impact. Locating where AI will impact people is your first step. What will happen to your team if AI agents mess up? Who will be in charge of the decisions?

Here are three questions to ask:

• What decisions will AI deployment change?
• Who owns these decisions?
• Who will be affected by these mistakes?

The less hypothetical the answers are, the better. What do you need AI to do when things go wrong: hand the task to a human, escalate it, or stop and ask for confirmation? Giving clear answers to each of them will help set the error budget. Besides, you’ll immediately see how much failure is acceptable for your specific case.

Evaluate your wins from this situation to know if you’re only getting marginal productivity wins or a meaningful business lever. 

Review for Audit & Regulations Before You Deploy

Now, the A: A is for Audit. If AI is there to make decisions, will your team still be able to explain, reproduce, and defend them if things go wrong? Here are some questions to figure out:

• Do we log inputs and outputs, model versions, and approvals?
• Can we replay a specific scenario or decision later to see why something happened?
• What data is sensitive, who has access to it, and what are our retention policies?

To bring ROI, AI needs to be safe. Boundaries are paramount to this. As each industry requires a different level of transparency, it’s essential to figure out all the controls before scaling.

Check The Readiness of Data

R — time to pressure-test data readiness! Most AI projects fail not because the model is not fit for the task, but because data is messy and no one owns it. Here’s what to look for:

• At least one AI-ready data product with an owner and SLAs.
• The sources of ground truth and their reliability.
• Knowledge of the company’s repetitive roadblocks.

If no specific team or person owns the data quality, you can’t measure the results of AI adoption properly. Messy, unreliable data is the main reason AI deployment fails in the production stage.

Assess Deployment & Infrastructure Risks

And finally, D stands for Deployment. The last question of this framework is fairly simple: can your company actually run this in real life, every day, not just in a demo? This is exactly the stage where ROI often disappears into thin air. The “real work” is not a sound model itself: rather, it’s the shipping, monitoring, and maintaining it so teams actually use it. 

The AI model needs to be fast and reliable enough for your company. Your team should be able to track down the specific moment when decisions start getting worse (in case that happens). And you always need an owner to fix and update it if anything breaks.

This HARD framework might highlight some hard truths about your company — yet it will also help you make better decisions about AI adoption.