AI agent ROI is net benefit divided by total cost, over a 12-month window. Net benefit is hours saved multiplied by loaded hourly cost, plus revenue gained, plus errors avoided. Total cost is the build fee plus twelve months of running costs. Track the inputs before you build. Industry research from Forrester and others suggests well-scoped agentic AI projects commonly return somewhere in the 200% to 400% range in year one — a useful band to pressure-test your own projections against.
The problemWhy most AI ROI numbers are nonsense
Industry research from the Writer/MindStudio camp has floated figures suggesting that a large share of AI investments return nothing measurable at all. The stat gets repeated because it's half-true and half-lazy. The reason so many AI projects "fail to return" isn't that the technology doesn't work. It's that nobody wrote down the baseline before they started, so there's nothing to compare the post-launch numbers to.
If you want a defensible ROI number — the kind you can put in front of a board or a partner without flinching — you have to do three unsexy things:
- Measure the baseline before you build anything.
- Pick a 12-month window and stick to it.
- Track the inputs monthly, not just at the end.
That's the entire difference between a real ROI and a vibe. This article gives you the formula, the five inputs you need to measure, a worked example, and the three mistakes that wreck the number.
The formulaOne line of maths
That's it. Everyone uses some version of this. The magic — and the argument — is what you put inside Net Benefit and Total Cost. Get the inputs right and the formula does its job. Get the inputs wrong and you get fiction.
The inputsFive numbers you actually need
Hours saved × loaded hourly cost
How many human hours the workflow removes per month, multiplied by the fully-loaded hourly cost of the person currently doing that work. "Loaded" means salary, plus taxes, plus benefits, plus overhead — usually 1.3x to 1.5x the base hourly rate. This is the time-saved component and it's usually the biggest number.
Revenue gained
New revenue the workflow is directly responsible for. Faster proposals win more deals. Better lead qualification means reps spend more time on real opportunities. Automated onboarding reduces churn. Be honest about attribution — if the agent is one of three reasons a deal closed, only count a third. This is the revenue component.
Errors avoided
The cost of mistakes the workflow prevents: rework, refunds, compliance fines, customer churn from bad experiences. A Deloitte AI impact assessment reported very large reductions in error rates for well-designed document-processing workflows in financial services. Count the dollar value of the mistakes you stop making. This is the risk component.
Build cost
The one-time cost to design and build the workflow. At Orbital Agents, Discovery and Design starts at $1,490 per workflow; Build is scoped after Discovery. Add any integration work, data cleanup, and internal team time. Be generous — undercounting the build cost is the fastest way to inflate your ROI on paper.
Run cost (12 months)
What it costs to keep the workflow running for a year: LLM API fees, tool subscriptions, monitoring, maintenance and any human-in-the-loop review time. LLM costs can drift sharply if usage grows, so model a best case and a worst case. The difference between them is often where the ROI argument happens.
Worked exampleA lead qualification agent for a $5M services firm
Illustrative only. The numbers below are a teaching example, not a case study. Use the same row structure with your own baseline inputs. Your result will be different — that's the whole point of measuring it.
Let's make the framework concrete. Imagine a professional services firm doing $5M in revenue, with three salespeople and an inbound lead volume of roughly 200 leads per month. Today, a junior sales admin spends about 25 hours a week manually qualifying, enriching, and routing those leads. Their loaded hourly cost is $45.
The firm builds a single AI workflow that handles the first-pass qualification: enrichment, ICP scoring, drafting a first-touch reply, and routing qualified leads to the right rep with context. Here's what the 12-month ROI model might look like:
| Line | Calculation | Value |
|---|---|---|
| Hours saved per month | 25 hrs/wk × 4.3 weeks × 60% automated | 65 hrs |
| Time-saved benefit (annual) | 65 hrs × $45 × 12 months | $35,100 |
| Revenue gained (annual) | Faster response → 4 extra closed deals × avg $8k | $32,000 |
| Errors avoided (annual) | Fewer dropped leads, fewer mis-routes | $5,000 |
| Net benefit | Sum of the three above | $72,100 |
| Build cost (one-time) | Discovery & Design + Build | $12,000 |
| Run cost (12 months) | LLM + tools + monitoring | $7,200 |
| Total cost | Build + 12 months run | $19,200 |
| 12-month ROI | ~275% | |
($72,100 − $19,200) ÷ $19,200 × 100 ≈ 275%. Roughly a 2.8x return in year one, which sits comfortably inside the 200%–400% band Forrester and others have reported for well-scoped agentic AI projects. The build pays back inside the first three months. Year two looks better because the build cost is front-loaded — only the run cost repeats.
Your numbers will be different. The shape of the calculation is what matters. If you can't fill in those eight rows with real baseline data before you build, you aren't ready to build yet.
Reality check. The example above sits squarely inside the 200%–400% band Forrester has reported for agentic AI projects. Single well-scoped workflows tend to outperform enterprise-wide programmes on year-one ROI because scale inflates cost faster than it inflates benefit. That's why the Orbital Agents rule is one workflow at a time.
Hard vs softWhat counts, and when
There are two kinds of AI ROI and you need both — but you need to label them clearly.
Hard ROI
Hard ROI is the cash number. Hours saved in dollars, revenue added, costs avoided, errors prevented. It's measurable, it's auditable, it goes on the P&L. This is the number you lead every business case with. If you can't defend a hard ROI number, you don't have a business case — you have a hope.
Soft ROI
Soft ROI is the strategic value. Faster decisions because your Monday report writes itself. Better customer experience because support responds in 30 seconds not 3 hours. Higher employee retention because the interesting work is no longer buried under admin. Independent research has suggested that organisations tracking both hard and soft returns see meaningfully higher overall value than those tracking only cash savings.
The rule: hard ROI leads, soft ROI supports. Put your hard number in the headline. Put your soft benefits in the narrative underneath. Never — ever — use soft benefits to rescue a weak hard number. Finance people can smell that from across the building.
Service ↗ We model the full ROI inside every Discovery & Design engagement. Starting from $1,490.The trapsThree mistakes that kill your ROI number
1. Counting the whole person instead of the hours
"The agent replaces a full-time employee, so I save $80,000 a year." No, it doesn't. It removes 60% of that person's work — and you still need them for the other 40%. Count the hours saved, not the headcount. Counting headcount is how you get an ROI number that evaporates under scrutiny in the first board meeting.
2. Forgetting the run cost grows
LLM API pricing is usage-based. If your workflow handles 10x more volume in month six than month one, your run cost goes up proportionally. Budget for the volume you'll actually hit, not the volume you have today. Usage-based pricing is the single biggest reason a good year-one ROI becomes a bad year-two ROI.
3. Measuring only at the end
The IDC has reported that AI projects typically take around 14 months to fully realise their value. If you measure ROI for the first time at the 12-month mark, you'll be staring at an incomplete picture and probably panicking. Measure every month. Track the five inputs the same way you track revenue. Adjust the workflow when the numbers drift.
Want the ROI modelled for your workflow, not a hypothetical?
Discovery & Design includes a full 12-month ROI model for your specific use case. One workflow, one number, one call.
Book a Free Scoping Call →Frequently asked
What's a good ROI for an AI agent?
Industry research from Forrester and others has reported well-scoped agentic AI projects commonly returning somewhere in the 200% to 400% range in year one. Use that band as a sanity check for your own projections, not as a target to aim at — your real number depends entirely on the inputs you measure for your specific workflow, your specific volume, and your specific baseline costs.
How do you calculate AI agent ROI?
Net benefit divided by total cost, times 100. Net benefit is hours saved multiplied by cost per hour, plus revenue gained, plus errors avoided. Total cost is the build fee plus 12 months of running costs. Measure over a full 12-month window and track the inputs monthly, not just at the end.
What's hard ROI vs soft ROI?
Hard ROI is the cash number: time saved in dollars, revenue added, costs avoided. Soft ROI is the strategic value: faster decisions, better customer experience, employee retention. Track both, but lead every business case with the hard number. Soft benefits belong in the supporting narrative, never in the headline figure.
How long until an AI agent pays for itself?
A well-scoped workflow typically pays back its build cost inside three to six months, with full 12-month ROI commonly landing in the 200% to 400% range reported by industry research. Research from IDC has reported that organisations realise value from AI projects in roughly 14 months on average — a single well-scoped workflow usually beats that comfortably if it's measured properly from day one.
Should I include intangible benefits in my ROI calculation?
Yes, but label them clearly. Intangibles like improved employee experience and faster decision quality are real, but CFOs discount them heavily. Quote your hard ROI number first, then add intangibles as a secondary narrative rather than padding the primary figure.
Get a real ROI number, not a spreadsheet guess.
30-minute free call. We'll walk your first workflow through the five inputs and give you a defensible projection.
Book a Free Scoping Call →