How to Measure ROI From an AI Feature Before You Build It
The four-line spreadsheet we use with founders to decide whether an AI feature is worth shipping — and how to prove it after launch.
Every AI feature pitch has the same slide: "This will save 10 hours per user per week." Nobody believes it, including the person presenting. The reason is simple — the math behind that number is usually invented on the way to the meeting.
At Augere Labs we've killed almost as many AI features as we've shipped. The ones we shipped had a spreadsheet. The ones we killed didn't. Here's how the spreadsheet works.
Why AI ROI is different
Regular feature ROI is a two-line calculation: users × conversion × price. AI ROI has three extra variables that trip up teams:
- Inference cost per action. Not zero. Not a one-time build cost. A recurring per-user tax.
- Acceptance rate. The output only counts if the user uses it. A 30% acceptance rate turns a 10-minute saving into a 3-minute one.
- Regression cost. When the model gets it wrong, the user has to notice, correct, and sometimes apologize to their own customer. Each of those has a cost.
The features that "obviously" have great ROI usually fall apart when you multiply those three together.
The four-line spreadsheet
Before we scope an AI feature we ask the founder to fill in exactly four rows:
- Value per accepted output. In dollars. What's this worth if it works?
- Expected acceptance rate. Between 0 and 1. Be pessimistic.
- Cost per attempt. Model + retrieval + review time.
- Volume per month. How often does this fire?
Monthly value = (Value × Acceptance × Volume) − (Cost × Volume). If that's not at least 5× the build cost annualized, we don't build it.
What to put on each line, honestly
Value per accepted output
Two ways to price it: time saved × loaded hourly rate, or revenue enabled. Time is easier; revenue is more defensible. If you're using time, be strict — 4 minutes saved is 4 minutes, not "an hour of context switching." Loaded rate for a US knowledge worker is around $60–$120/hour in 2026. For a support agent, $25–$40. For a sales rep, use their fully-loaded cost including commission.
Acceptance rate
Assume 40% for a well-designed inline feature, 15% for a first-time-visible feature, and 5% for anything that requires the user to seek it out (like a chat sidebar on a page they don't visit).
Founders always want to write 70%. They're always wrong. Real production numbers, from live projects: draft email suggestions in a CRM landed at 46%, auto-categorization in a ticketing tool at 62%, a "suggested reply" widget at 11%. Wide range. Design decides the number more than the model does.
Cost per attempt
This is the line teams get wrong most often. It's not just the model bill. It's:
- Input tokens × input rate (system prompt + retrieval context + user input)
- Output tokens × output rate
- Embedding cost if you're doing retrieval
- Reranker cost if you're using one
- Human review cost if there's a QA loop
Realistic 2026 numbers for a mid-complexity RAG call: $0.008–$0.04 per attempt on cheaper models, $0.03–$0.15 on Claude Sonnet or GPT-4o. Multiply by 10,000 daily calls and you have a real line item.
Volume
Only count the volume you'll actually hit in the first 90 days. Founders like to project the second-year steady state; that's a forecast, not a budget. If you can't get to positive ROI on 90-day volume, the feature is a bet, not a build.
An example that failed
A founder we were about to work with wanted an AI feature that would summarize competitor pricing pages daily and email the sales team. The pitch was "saves each rep 30 minutes a day." Sixteen reps × 30 minutes × $50/hour × 22 days = $8,800/mo saved.
We asked four questions:
- Would every rep read the summary daily? (No — probably 3–4 of them.)
- Would they trust it enough to skip their own check? (No — they'd still spot-check the source.)
- How much would 16 competitor pages × 22 days × $0.06 per summary cost? ($21/mo.)
- How much would the build cost? ($7K.)
Real value became 4 reps × 10 minutes × $50 × 22 = $733/mo. Payback: 10 months. We suggested a much smaller v1: pull the pages, diff them, flag changes only. Build cost: $1,200. Value: same. Payback: 6 weeks. That's the version we shipped.
The metrics that convince a board
After launch, three numbers matter — and none of them are "AI usage."
- Time-to-completion delta. Median seconds from workflow start to finish, before vs after. This is the ROI number.
- Downstream conversion. Did the workflow the AI supported get completed at a higher rate?
- Retention delta on the cohort that used it. Users who accepted at least one AI output N times, retention vs users who didn't.
Anything that isn't tied to one of those three is a vanity metric. Track fewer things; defend them harder.
Cost per useful output is the metric to protect
Once a feature is live, the metric to watch weekly is cost per accepted output — total inference spend divided by outputs the user kept. It's the honest denominator, and it usually gets worse before it gets better as users find creative ways to hammer the feature.
A pattern we see: month one is great, month three is a disaster because one power user with a script has generated 40% of your calls. Add per-tenant caps from day one. Not because the user is bad — because unbounded is unbounded.
What to ignore
- Total AI usage. A user who accepts nothing is a cost, not a win.
- Session length increases. Might mean engagement. Might mean the AI is slow.
- Model benchmarks. Ours-is-better-on-MMLU has no bearing on your feature's ROI.
- Vendor case studies. Read them for patterns, not for numbers.
A common misconception
"AI features pay for themselves." Some do. Most don't automatically. The ones that pay for themselves have three things in common: a specific workflow (not "AI copilot"), a real user paying real money for the surrounding product, and a UI that lets the user accept or reject cleanly. Take any of those three away and the math stops working.
When ROI is the wrong question
Occasionally the correct answer is "build it anyway." Two cases:
- Competitive parity. If your top three competitors ship the feature, not shipping is its own cost — just don't call it ROI-positive when it isn't.
- Learning value. The first AI feature a team ships teaches them how to ship the next one. Budget it as a training expense, not a revenue play.
Being honest about which bucket a project is in beats pretending it's revenue when it's really learning.
How this changes our scoping calls
We now spend the first 20 minutes of any new-project call filling in the four-line spreadsheet with the founder. Roughly 30% of the time, we talk them into a smaller v1. Roughly 10% of the time, we talk them out of building at all. The remaining 60% we scope, quote, and ship. It's the highest-ROI thing we do in a sales call.
Frequently asked questions
What's a good ROI target for an AI feature?
Annualized value should be at least 5× the build cost within 12 months. Below 3× and you're better off spending the money elsewhere; above 10× and you should probably raise your ambition and ship a bigger version.
How do I estimate acceptance rate before shipping?
Run a 30-example Wizard-of-Oz: hand-produce the AI output for real user data, put it in front of five users, and count how many they accept. It's the closest thing to a real number you can get in a week.
How long does it take for an AI feature to break even?
For a well-scoped feature at Augere Labs prices ($3K–$15K build), typical payback is 2–8 months. Anything with a longer payback usually deserves a smaller v1.
Should I include the cost of my team building it in ROI?
Yes. Include salary or agency invoice as the build cost, not just cash outflow. Otherwise the math flatters internally-built features that took a quarter of engineering time.
What if my AI feature is for internal ops, not customers?
Same spreadsheet, different value line. Instead of revenue, count hours saved × loaded cost. Internal AI features often hit ROI faster than customer-facing ones because acceptance rate is close to 100%.
Want us to run the numbers with you?
We do this exercise on every first call. If you have an AI feature in the pipeline and you're not sure if it's worth building, that's a good use of 30 minutes. Book one and bring the workflow — we'll fill in the spreadsheet live.
Or read our AI product spec guide next if the ROI checks out and you're ready to scope the build.
FAQ
Frequently asked questions
What's a good ROI target for an AI feature?+
At least 5× the build cost annualized within 12 months. Below 3× and the capital is better spent elsewhere; above 10× and you should probably ship a more ambitious version.
How do I estimate the acceptance rate for a new AI feature?+
Run a Wizard-of-Oz test with 30 real inputs. Hand-produce the AI output, put it in front of five real users, and measure how many they accept. That's the closest real-world number you can get in under a week.
How long does an AI feature take to break even?+
For a scoped feature costing $3K–$15K to build, payback typically lands between 2 and 8 months. Longer paybacks usually indicate the v1 scope is too large.
Should I count the model bill in ROI calculations?+
Yes, every attempt — not just successful ones. A 30% acceptance rate means 70% of your cost has no revenue attached, and that's the honest denominator.
Building something similar?
Let's talk in 30 minutes.