Everyone wants to know the ROI of AI. The problem is that most AI ROI calculations are somewhere between optimistic and delusional. Vendors inflate the upside, teams undercount the costs, and nobody accounts for the organisational change required to actually capture the value.
We've seen AI business cases that claim 10x returns while conveniently ignoring data preparation costs, change management effort, and the six months of iteration needed to get model accuracy above a useful threshold. That's not ROI measurement — it's storytelling.
Here's how to do it properly.
Why Most AI ROI Calculations Are Wrong
The typical AI ROI calculation looks like this: "This process costs us £500K per year in labour. AI will automate 80% of it. Therefore, AI saves us £400K." This is wrong for several reasons:
- It ignores implementation costs: Building, deploying, and maintaining AI is not free. Development, infrastructure, data preparation, testing, monitoring, and ongoing model maintenance all cost money.
- It assumes perfect adoption: Just because AI can do something doesn't mean your team will use it. Change management, training, and workflow redesign are real costs that rarely appear in the business case.
- It conflates potential with reality: An AI model that achieves 80% accuracy in a lab does not automate 80% of the workload in production. Edge cases, human review, and error correction eat into the savings.
- It ignores time to value: Most AI projects take 3-6 months to reach production. The ROI clock doesn't start until users are actually using the system.
The Four Types of AI Value
AI creates value in four distinct ways. A robust ROI framework needs to capture all of them, not just the easy-to-measure ones:
- Cost reduction: The most straightforward to measure. Automating manual tasks, reducing errors, cutting processing time. Calculate the current cost of the process (labour, tools, rework) and subtract the cost of the AI-augmented process (infrastructure, development, remaining human oversight). Be honest about what "remaining human oversight" actually looks like.
- Revenue growth: AI that enables better pricing, faster sales cycles, improved conversion rates, or new products. Harder to measure because attribution is tricky. Use controlled experiments — A/B tests or phased rollouts — to isolate the AI impact from other variables.
- Risk reduction: AI that improves compliance, catches fraud, reduces safety incidents, or strengthens security. The value here is in avoided losses, which are inherently uncertain. Use historical loss data and probability-weighted estimates rather than best-case scenarios.
- Employee productivity: AI that makes knowledge workers faster and more effective. This is the trickiest category because "saving 2 hours per week" doesn't automatically translate to cost savings unless you can redeploy that time to higher-value work. Be specific about what happens with the recovered time.
Setting Baselines That Actually Work
You cannot measure improvement without knowing where you started. Yet most organisations skip baselining or do it poorly. Here's how to set baselines that hold up:
- Measure before you build: Spend 2-4 weeks measuring current process performance before any AI development begins. Track volume, time per task, error rates, and costs. This baseline is your evidence.
- Use real data, not estimates: "We think it takes about 30 minutes per ticket" is not a baseline. Instrument the process and measure actual time spent across a representative sample.
- Account for variability: Processes have peak and trough periods. Your baseline should cover enough time to capture this variation — typically at least 4 weeks of data.
- Document your methodology: When you present ROI numbers six months later, you'll need to explain exactly how the baseline was measured. If you can't defend the baseline, the ROI figure is meaningless.
A Practical Tracking Framework
We use a simple three-tier framework to track AI ROI over time:
- Leading indicators (weekly): Usage metrics, adoption rates, model accuracy, processing volume. These tell you whether the system is working as intended and whether people are actually using it.
- Lagging indicators (monthly): Cost per transaction, time savings, error rates, throughput improvements. These measure actual business impact but need time to accumulate.
- Strategic indicators (quarterly): Revenue impact, customer satisfaction changes, competitive positioning. These capture the broader business value that AI enables over time.
Our AI ROI Calculator helps you model these numbers before you commit to a project. It forces you to input real costs and realistic assumptions, not fairy tales.
Common Pitfalls to Avoid
After helping dozens of organisations measure AI ROI, here are the mistakes we see most often:
- Cherry-picking metrics: Only reporting the metrics that look good while ignoring ones that don't. If accuracy improved but adoption is at 20%, that's not a success story.
- Ignoring maintenance costs: AI systems require ongoing monitoring, model retraining, and infrastructure costs. Year-one ROI is not the same as year-three ROI if you're not accounting for these.
- Comparing against the wrong baseline: Comparing AI performance to the worst-case manual process rather than the average. This inflates the improvement.
- Forgetting opportunity cost: The money and time spent on AI could have been spent on other improvements. Your ROI needs to be better than the next-best alternative, not just better than doing nothing.
- Measuring too early: AI systems typically improve over the first 3-6 months as models are refined and users learn the system. Measuring ROI at week two gives you a misleadingly low number.
Struggling to build a credible ROI case for AI? We help organisations measure what matters and build investment cases that stand up to scrutiny. Book a strategy call and we'll walk you through our approach.