Someone pitches you an AI project. They say it'll save $500K annually. You ask when. They say "within 18 months."
You want to believe them. You also know this might be fantasy.
Here's what realistic AI ROI actually looks like.
The Math That Works
A typical AI project costs $250K to build and implement. It takes 12 months. After deployment, it costs $50K annually to maintain.
For ROI to break even in year one, you need $250K in benefits. For year two, you need to account for ongoing maintenance ($50K), so you need $300K in benefits to be better off than year one.
Year 1 ROI = (Benefits - Development Cost) / Development Cost Year 2 ROI = (Benefits - Maintenance Cost) / Development Cost
For a $250K project:
- Year 1 breakeven: $250K in benefits
- Year 2 breakeven: $300K in benefits
Most AI projects achieve their first real ROI in year 2, not year 1. That's realistic.
What Kinds of Benefits Actually Happen
Labor reduction: This is the most common and reliable.
A system that automates data entry, classification, or simple decisions can reduce manual work by 30-60%. If you have 5 people doing data entry, each costing $60K annually (salary + overhead), and AI reduces this by 40%, you save $120K per year.
But here's the real math: you probably don't fire anyone. You redeploy them to higher-value work. If that higher-value work generates 30% more revenue per person, the ROI is higher. But if it just fills time, the benefit is only the reduced time on data entry.
Most companies realize 50-70% of theoretical labor reduction, because of redeployment and the reality that people aren't 100% fungible.
Process acceleration: You move faster, so you sell faster or operate more efficiently.
Example: A sales team's deal review process currently takes 1 week (4 person-days). AI analysis reduces this to 2 days. With a pipeline of 100 active deals, that saves 200 person-days annually. If each person-day of sales time is worth $2K in opportunity value, that's $400K in acceleration benefit.
But again, the real number is probably 50-70% of the theoretical maximum.
Revenue impact: The hardest to quantify, but potentially the biggest.
Example: AI recommendations increase customer purchase volume by 8%. Revenue impact is huge. But attribution is hard. Was it the AI, or improved marketing, or a seasonal effect? You need baseline data and careful measurement.
Most companies claiming revenue uplift from AI are overstating it by 2-3x.
Quality improvement: Fewer errors, better compliance, lower risk.
A lending company's AI system catches fraud 95% of the time vs 80% previously. That's 15% fewer fraudulent loans. If average fraud loss is $50K per loan and you process 500 loans annually, that's 75 fewer fraudulent loans × $50K = $3.75M benefit.
But this is also easy to overstate. You need historical data on actual fraud losses to know if the savings are real.
Timeline: When Benefits Actually Arrive
Here's a more realistic timeline:
Months 0-12: Build and implement
- Benefits: $0
- Cost: $250K (plus ongoing salaries for your internal team)
Months 12-18: Adoption and optimization
- Benefits: 20-30% of expected steady-state benefits (you're still optimizing, users are still learning)
- Cost: $25K (maintenance, training, troubleshooting)
- You're not positive ROI yet
Months 18-24: Steady state
- Benefits: 80-100% of expected steady-state benefits (system is working, team understands it)
- Cost: $50K (ongoing maintenance)
- Now you might break even or show modest positive ROI
Year 3+: Established benefit
- Benefits: 100% of steady-state (or higher if you've optimized further)
- Cost: $50K
- Strong positive ROI
This is the realistic timeline. Project completion (month 12) is not benefit arrival. Benefit arrival is month 18-24.
How to Measure ROI Correctly
Define what you're measuring before you start.
Not "improve productivity." Specifically: "reduce time spent on X by Y hours per week, worth $Z annually."
Example good definitions:
- "Customer support team spends 8 hours per day categorizing tickets. AI will reduce this to 2 hours per day. Benefit: 6 hours × 5 people × 250 days × $75/hour = $562K annually."
- "Sales process has 6 stages. Each stage review takes 4 hours. AI will reduce review time by 50%. Benefit: 2 hours × 100 deals per month × 12 months × $150/hour = $360K annually."
These are concrete and measurable.
Measure actual usage, not potential usage.
If you predict that 100 employees will use the system and save 2 hours weekly, but only 40 employees actually use it and save 1 hour weekly, the actual benefit is 40% of the estimate.
Measure against a baseline.
Before you deploy the system, measure: how much time do people spend on this now? How many errors occur? How long does this process take?
Then, after deployment, measure the same things. The difference is your benefit.
Without a baseline, you're guessing.
Measure real business outcomes, not activity metrics.
Bad metrics: "system processed 500 documents" or "made 1,000 predictions."
Good metrics: "reduced average ticket resolution time from 5 days to 3 days" or "reduced data entry errors from 3% to <1%."
Activity metrics are easy to measure but tell you nothing about real value. Outcome metrics are harder to measure but tell the truth.
Red Flags: When ROI Claims Are Unrealistic
"We'll achieve ROI in year one."
Possible, but rare. Only true if benefits are immediate (like automating a process that people currently do manually, and the system starts working perfectly on day one).
More likely: year two.
"We expect 60% labor reduction."
Possible, but people aren't robots. Even if the system saves 60% of potential time, your team will find other work. Realistic benefit: 30-40%.
"No ongoing maintenance cost; the model will run itself."
This is wrong. Models drift. Data changes. You need monitoring, periodic retraining, and troubleshooting. Budget $50K-$100K annually for maintenance on any real production system.
"The model is so good, we can completely automate this decision."
Probably not. If the model is 95% accurate, you still have 5% errors. For critical decisions (loan approvals, fraud decisions, personnel actions), those 5% errors are unacceptable. You need human review.
Realistic scenarios: the model recommends, a human decides. The model catches most easy cases, a human handles edge cases.
"We're not sure exactly how we'll measure ROI, but the benefit will be obvious."
Red flag. If you can't define how you'll measure benefit before starting, you probably won't see it after finishing.
The Truth About AI ROI
AI is good at:
- Automating repetitive tasks (labor cost reduction)
- Processing large volumes of information (time savings)
- Spotting patterns humans miss (quality improvement)
AI is bad at:
- Being better than humans at subjective decisions (more art than science)
- Replacing high-skill work without humans staying in the loop
- Making unique strategic insights (pattern matching, not strategy)
If your project falls into "good at," ROI is probably 2-3 year positive.
If it falls into "bad at," ROI might never materialize.
Real World Examples
Example 1: Document Classification
Cost: $150K to build Benefit: Reduces manual categorization from 8 hours/day to 2 hours/day for 5-person team Calculation: 6 hours × 5 people × 250 days × $50/hour = $375K annually Year 1 (after month 12 deployment): 20% adoption, 3 months of usage = $93K benefit Net year 1: -$57K (below break even) Year 2: Full adoption, 100% benefit = $375K Net year 2: +$125K (positive ROI) Year 3: $375K benefit - $30K maintenance = $345K net benefit
Example 2: Sales Lead Scoring
Cost: $250K to build Benefit: More accurate lead scoring lets salespeople focus on high-probability opportunities Claims: "20% increase in close rate" Problem: Measuring this is hard. Lots of variables. True incremental benefit: probably 5-8%, not 20%. Realistic benefit: 8% close rate increase on $10M pipeline = $800K additional revenue (at 10% margin = $80K contribution) Year 1: 30% benefit realization, 6 months of usage = $12K benefit Net year 1: -$238K Year 2: 80% benefit realization = $64K benefit Net year 2: -$186K Year 3: 100% benefit realization = $80K benefit Cumulative ROI at year 3: Barely positive
This project barely makes financial sense, but might make strategic sense (better data for sales, more professional process, etc.).
Make Sure You're Ready First
Understanding realistic ROI is part of the picture. Also check your AI readiness and think carefully about choosing your first use case.
The Decision Framework
Before committing to an AI project, decide what you need to see:
- What's the realistic benefit? (50-70% of theoretical maximum)
- When will we see it? (Usually year 2 or later)
- How certain are we? (Confident, hopeful, or speculative?)
- What's the downside if we're wrong? (Lost $250K, but learn something)
- Is there a strategic reason beyond ROI? (Build capability, understand AI, etc.)
If the answer to #3 is "speculative" and #5 is "no," wait. Pick a project with clearer ROI.
If the answer to #5 is "yes, we need to learn AI capabilities," then ROI is secondary and the project makes sense.
But don't pretend speculative ROI is guaranteed. Be honest about what you expect to happen.
Most AI projects with well-defined benefits, clear baselines, and realistic timelines do achieve positive ROI. But it takes longer than executives want it to, and the benefit is smaller than they hoped.
That's not a failure. That's reality.
Plan for it.
Model Your Specific ROI
Every AI project is different, and so is the ROI. If you're evaluating a specific AI opportunity, we can help you run realistic numbers and understand what success actually looks like for your situation. Book a discovery call to work through your business case.
