Every AI pitch deck includes a ROI slide. Most of them are wrong — not because the numbers are fabricated, but because they measure the wrong things. Here is how to build an honest AI business case.
Why Most AI ROI Calculations Are Misleading
Walk into any AI vendor presentation and you'll see the same math: "Our tool saves 3 hours per employee per week. You have 500 employees. That's 1,500 hours per week. At $50/hour, that's $3.9 million in annual value. The tool costs $500K. Done."
This calculation is wrong for three reasons:
1. It confuses time freed with money saved. The 1,500 hours you recapture from AI automation don't disappear from payroll. Employees keep getting paid. The value is only realized if those recaptured hours are redirected to higher-value work — which requires active management, not passive hope.
2. It ignores implementation costs. Enterprise AI deployments involve integration work, change management, training, data preparation, and ongoing maintenance. These costs routinely run 2-4x the software license cost in year one. Most ROI calculations omit them.
3. It ignores quality effects. AI tools often change quality as well as speed. Sometimes quality improves (AI catches errors humans miss). Sometimes it degrades (AI outputs require more review than original work). The net effect on quality is almost never accounted for.
The Three-Factor Framework
McKinsey's AI practice uses a framework that breaks value creation into three distinct categories, each requiring different measurement approaches:
Factor 1: Time Freed
What to measure: Hours per employee per week on tasks before AI, versus after AI, for the specific tasks being automated. Crucially: what do employees actually do with freed time? Use time tracking or manager observation, not surveys (people systematically overestimate their time savings).
Honest math: If 50% of freed hours translate to higher-value work and 50% are absorbed by new administrative work, your time-freed value is half what the simple calculation suggests.
Factor 2: Error Reduction
What to measure: Error rate before AI versus after, cost per error (including downstream correction costs and reputational effects), and false positive/negative rate of the AI itself.
Honest math: AI error reduction often offsets AI errors. A tool that reduces manual errors by 40% but introduces a 15% AI error rate may be a net positive — or not, depending on relative error costs. Calculate both.
Factor 3: Decision Quality
What to measure: This is the hardest to quantify. Proxy metrics include: time to decision, consistency of decisions across similar cases, outcomes of AI-assisted decisions vs. non-AI-assisted decisions (requires A/B testing or historical comparison).
Honest math: Decision quality improvements often don't show up in 6-month ROI calculations. They compound over years and are best measured in outcome data — customer retention, sales win rates, default rates, patient outcomes — not in process metrics.
Realistic Timelines
Based on enterprise deployment data, expect:
- Month 1-6: Net negative. Implementation costs, productivity dip during transition, and change resistance.
- Month 6-12: Break-even territory. Tools are integrated, users are trained, workflows are adapted.
- Month 12-24: Positive ROI on measurable metrics. Quality effects beginning to show.
- Month 24-36: Full ROI realization including compound quality effects.
Any business case projecting payback in under 12 months for a significant enterprise deployment should be scrutinized carefully.
What Good Measurement Looks Like
The companies measuring AI ROI well share three practices:
-
Define "success" before deployment, not after. Pick 2-3 measurable outcomes that matter to the business and commit to measuring them rigorously.
-
Run a controlled pilot first. Deploy to a subset of users, keep a control group, and measure real outcomes before scaling. This builds credibility for the business case and surfaces problems early.
-
Track value realization, not just deployment. Many organizations track the rollout of AI tools. Few track whether the expected value actually materialized six months later. Build value realization reviews into the deployment plan.