How to actually calculate AI ROI for your business (without the vendor-approved hallucinations)

If you've been shopping for AI projects, you've seen this slide. It's in every vendor deck. "AI delivers 10-50x ROI in the first year" or "companies save 30% of their ops cost with AI in year one." Some variant, some stat, always in a band wide enough to be meaningless.

Meanwhile, you've got a real business with a real P&L, and someone is asking you to sign off on a real number. Let me give you the honest formula I use when scoping client work — and why it's almost always more conservative than the vendor math, and yet still profitable.

The five numbers that matter

Everything else is noise. The five:

1. Hours per week your team spends on tasks AI could own. Not "could theoretically be touched by AI." Own. End-to-end. Reception-booking-a-dentist-appointment is an AI-can-own task. Drafting-a-complex-M&A-memo is not. Be ruthless here — most teams overstate this by 2-3x on first pass.

2. Blended hourly rate of the team doing that work. Fully loaded (salary + super + overhead). If your reception staff costs $40/hr in salary but you spend another $15/hr on desk, tech, management oversight, that's $55/hr blended.

3. Realistic automation rate. What fraction of those hours can AI actually do in production — not in demo? Be honest. For well-scoped reception work, 60-75% is realistic. For creative work, 20-40%. For anything involving judgement, start at 30% and be thrilled if you hit 50%.

4. Revenue lift %. What % lift in top-line revenue do you realistically expect because you shipped this AI? Not because it theoretically could, because you've matched it to a specific customer outcome. If the AI can't be traced to a revenue move, this number is 0. That's fine — not every AI project needs revenue lift. Some just save cost.

5. Build cost. What you're pencilling in for the initial build. Fixed-scope, not time-and-materials.

The formula

annual_hours_saved    = hours_per_week * (automation_rate / 100) * 52
annual_labour_saved   = annual_hours_saved * hourly_rate
annual_revenue_lift   = current_revenue * (revenue_lift / 100)
annual_benefit        = annual_labour_saved + annual_revenue_lift

payback_months        = (build_cost / annual_benefit) * 12
three_year_net        = (annual_benefit * 3) - build_cost
three_year_roi        = (three_year_net / build_cost) * 100

That's it. There is no magic. If you get a vendor showing you 50x ROI, they're doing exactly this formula but being wildly optimistic on the automation rate, the revenue lift, or both.

The four verdicts

Using this formula, every AI project falls into one of four bands:

GREEN: payback in under 6 months

This is what you want. Almost always happens when:

• You have a specific process that runs multiple times per week
• The process is mostly rules + patterns (not deep judgement)
• AI can reach 60%+ automation rate credibly
• The build cost is well-sized (you're not over-engineering)

Examples that commonly land here: reception automation, tier-1 customer support, content generation pipelines, compliance reviews, recall outreach.

CYAN: payback in 6-12 months

Reasonable. Standard green-light from most finance committees. Usually:

• Decent labour savings but no meaningful revenue lift
• Or some revenue lift but the automation rate is only 40-50%
• The project is strategic (needs to exist) but not urgent

Examples: internal knowledge base AI, ops dashboards, analyst assistants, marketing copy pipelines.

YELLOW: payback in 12-24 months

This is the "longer play" band. Worth doing only if:

• It's strategic enough to justify the wait (locks in a competitive position)
• Or it's a stepping stone to a much bigger project later
• Or the team is severely undersized and the build also buys you hiring time

Most non-strategic projects in this band shouldn't be done. The money's better spent on a narrower scope that lands in green or cyan.

RED: payback beyond 24 months

Reconsider scope. Seriously. This verdict almost always means one of:

• The automation rate is fantasy
• The revenue lift is wishful
• The build is scoped too large
• Or the underlying economics don't support AI — stick with humans

A red verdict is not a "try harder" signal. It's a "do something different" signal.

Where the vendor decks cheat

If you compare the vendor numbers to your own, 90% of the gap is in three places:

Automation rate. Vendors will tell you their AI hits 85-95% automation. In live production across a full year, with all the edge cases that 10-minute demos don't surface, you'll see 55-70% for most processes. Don't use vendor automation rates in your math.

Revenue lift. Vendors attribute all revenue growth during the AI rollout to the AI. That's not how causality works. The real lift number is the counterfactual — what would revenue have done without the AI, all else equal. For most AI projects that's honestly 0-3%. Not 20%.

Build cost. Vendors quote the software. Not the integration. Not the training. Not the inevitable scope creep. Not the 12-month rebuild when they change their API. Add 30-50% to any vendor-quoted build cost unless it's fixed-scope with a hard ceiling.

The one real secret

Honest AI ROI numbers are usually better than the sceptics think, worse than the vendors claim.

A well-scoped project with realistic inputs almost always lands in the green-to-cyan band. That's good. It means AI projects are genuinely worth doing. It also means you don't need 10x ROI to justify shipping — 1.5x in year one is fine, because the compounding matters more than the headline.

The ROI calculator on pulseagency.dev

I built this calculator because I got tired of clients walking in with vendor math and wondering why my estimates were lower. The calculator uses exactly the formula above. It takes your five numbers, gives you:

• Annual hours saved
• Annual labour savings in AUD
• Annual revenue lift in AUD
• 3-year net
• 3-year ROI %
• Payback months
• Verdict (GREEN / CYAN / YELLOW / RED)

And it emails you the analysis so you can show it to your board.

No pitch at the end. No "click here for a demo." If the numbers don't work, go do something else. That's a feature, not a bug.

A worked example

Let's run a concrete case. Melbourne accounting firm, 8 staff, wants an AI that handles tier-1 client questions (tax thresholds, GST timing, super caps, etc.).

Inputs:

• Hours per week: 18 (three partners each spend ~6 hrs/week on repeat questions)
• Hourly rate (blended): $180
• Automation rate: 65% (question variety is high, clients sometimes want to talk to a human)
• Current annual revenue: $1.8M
• Revenue lift: 3% (faster response times → better client retention → 3% annual retention uplift)
• Build cost: $28,000

Math:

• Annual hours saved: 18 × 0.65 × 52 = 608 hrs
• Annual labour saving: 608 × $180 = $109,440
• Annual revenue lift: $1.8M × 0.03 = $54,000
• Annual benefit: $163,440
• Payback: $28,000 / $163,440 × 12 = 2.1 months
• 3-year net: ($163,440 × 3) − $28,000 = $462,320
• 3-year ROI: $462,320 / $28,000 × 100 = 1,651%

GREEN verdict. Payback in 2.1 months. That's a real project.

Now let's stress-test. What if the automation rate is only 40%?

• Annual hours saved: 18 × 0.40 × 52 = 374 hrs
• Annual labour saving: 374 × $180 = $67,320
• Annual benefit (labour only, ignore revenue lift as speculative): $67,320
• Payback: $28,000 / $67,320 × 12 = 5 months

Still green. This is how you test — change one variable, see if the verdict holds. If the verdict only holds under heroic assumptions, the project is weaker than it looks.

What the ROI calculator can't do

It's a planning tool, not an oracle. What it doesn't capture:

• Team morale. A good AI deployment raises it. A bad one crashes it. Neither shows up in the math.
• Customer experience compounding. 24/7 availability, sub-second response times, multilingual support — these create value you won't see in the first year.
• Strategic moat. Getting good at AI now vs. starting 18 months later. Harder to quantify but probably bigger than the direct ROI.
• Downside risk. What if the build fails? What if the vendor changes pricing? What if regulation shifts? The calculator shows expected value; it doesn't show variance.

Use it as a sanity check, not a decision-maker. If it says GREEN, talk to a builder. If it says RED, stop.

Use it

The ROI calculator is here. Costs nothing, takes 90 seconds, emails you your analysis. If the math works, book a call and we'll turn it into a scope. If it doesn't, don't build it.

FAQ

Q: Should I include soft benefits (team morale, brand lift) in the formula? A: No. Use the calculator to get the hard-number ROI. Then discuss soft benefits separately with your board. If the hard numbers already justify the project, soft benefits are upside. If they don't, soft benefits rarely close the gap.

Q: What if my revenue lift is genuinely high — like 20%+? A: Then you've got a product decision, not a process decision, and the math skews even greener. But double-check the causal claim. "We shipped AI and revenue went up" is correlation; "this specific AI feature unlocked this specific customer behaviour that generates this specific revenue" is causation.

Q: Should I use the calculator before or after talking to a builder? A: Before. If the numbers don't work, don't pay for a scope session. If they do, show up to the call with the output printed — it anchors the conversation on reality, not vendor fantasy.

Q: What's the biggest mistake in AI ROI math? A: Using the vendor's automation rate. Close second: forgetting that labour savings rarely translate to headcount reduction in the first year (you usually redeploy, not fire). Third: assuming revenue lift compounds at the same rate every year. Don't.