How To Tell If An AI Agent Will Actually Save Your Business Money (Before You Buy One)

Before you buy any AI agent, tool, or agency package, run this five-question test. It takes 15 minutes. If the numbers pass, buy. If they do not, walk. You do not need to be technical, and you do not need to trust the vendor's ROI claim. The math below works on any AI tool priced between $19 a month and $50,000 a project.

Every vendor promises savings. Almost nobody publishes the actual math a buyer should run before they sign. This is that math.

#The five-question diagnostic

Get pen and paper (or a spreadsheet). Answer each in order. Skipping one invalidates the result.

#Q1: What specific, recurring task will the AI handle?

Name one task. Not a category ("customer support"), a specific task ("draft the first-response email to inbound support tickets tagged 'billing'"). If you cannot name one, do not buy yet. The most common AI-purchase mistake is buying a "platform" before you have one specific job for it to do.

#Q2: How many hours a week does that task take a human today?

Real hours, not vibes. Ask the person doing the task. Track it for a week if you have to.

If the answer is under 3 hours a week, stop. The cost of setup, calibration, and verification will exceed the savings. AI shines when there is enough volume to amortise the overhead. Below 3 hours a week, keep it human.

#Q3: What is the fully-loaded hourly cost of that human?

Not the salary rate. The loaded rate. Formula: base rate x 1.4 to 1.7. That multiplier covers benefits, payroll taxes, management overhead, tools, and lost productivity when they context-switch. A $22/hr contractor is really $30 to $37 an hour of business cost.

If you use a VA at $18/hr and forget to load the number, you will over-count the savings.

#Q4: What is the AI's total monthly cost?

Three lines, always:

• SaaS subscription (the sticker price)
• Setup, amortised over 12 months (if setup takes 8 hours of your time at your own loaded rate, that is real. Divide by 12.)
• Realistic weekly maintenance (5 to 30 minutes a week for the first quarter. Load that too.)

Vendors usually quote only the sticker. That is the smallest of the three lines.

#Q5: What is the verification cost?

The one everyone skips. Someone has to look at what the AI produces during the first 3 months. Not everything. Just a sample. Target 15 to 20 percent sample rate on the outputs for weeks 1 to 8, dropping to 10 percent afterwards. That is real reviewer time at a real loaded rate.

Miss this line and you build your ROI on a fantasy. This is also the number-one root cause of the week-2 collapse we described here. If you skip the verification budget, you skip the verification, and by week 3 the agent has drifted into wrong-but-plausible answers and nobody noticed.

#The math

Put it all together. Monthly saving equals:

(hours saved per week x 4.33 x loaded rate) - (SaaS + amortised setup + maintenance + verification)

Positive number = buy. Zero or negative = walk. Simple.

#Worked example: meeting notes with Fathom (pass)

Say you run 12 meetings a week and post-meeting admin takes 40 minutes total (about 3.3 minutes per meeting, honestly counted). Your loaded rate is $30/hr.

• Savings side. Fathom automates ~80% of that admin, saving ~32 min/week. That is 0.53 hr/wk x 4.33 x $30 = $69/month.
• Cost side.
  • Fathom Pro: $19/month
  • Setup (1 hour at $30 loaded): $2.50/month amortised
  • Maintenance (15 min/week at $30 loaded): $32.50/month
  • Verification (skim the summary of every meeting first month, 30 seconds each = 6 min/week at $30 loaded): $13/month
• Net saving: $69 - $67 = +$2/month.

That is a bare pass. Not a win, not a loss. For a founder running only 12 meetings a week, Fathom pays for itself but does not print money. The real win comes at 20+ meetings a week, where savings scale linearly and costs stay flat.

Lesson: even a "cheap" AI tool needs a real volume floor. Below the floor, the math fails.

#Anti-example: a $2,000/month AI sales agent (fail)

Same test. Vendor claims "10 hours a week saved." Your actual sales rep, loaded, costs $35/hr. Actual investigation shows the AI saves maybe 4 hours a week in reality (some tasks it does not handle, some it needs escalation on).

• Savings side. 4 hr/wk x 4.33 x $35 = $606/month.
• Cost side.
  • SaaS subscription: $2,000/month
  • Setup (10 hours of your team at $50 loaded): $42/month amortised
  • Maintenance (2 hours/week at $50 loaded): $433/month
  • Verification (25% sample rate for first 8 weeks at $50 loaded): $260/month
• Net saving: $606 - $2,735 = -$2,129/month.

Walk. And when the vendor says "but it will save 10 hours by month six," ask them to sign that number into the contract. They will not.

#How vendors inflate the numerator

Watch for these three tricks in every quote:

1. "Save 10 hours a week." Usually true for a fictional persona, not you. Ask which specific tasks the 10 hours came from and add them up.
2. Multiplying by an inflated rate. Vendor uses "$75/hr fully-burdened US employee rate" when your actual worker is a $18/hr offshore VA. Insist on your real rate.
3. Ignoring calibration weeks. The first 4 weeks are usually net negative because of setup and reviewer time. Ask for the 12-month total, not the "steady state" savings.

#How buyers understate the denominator

The other side of the same coin:

1. Zero setup cost assumption. There is always a setup cost. If the vendor says "5 minutes to install," that means 5 minutes for the technical part. Add 4 to 12 hours for actually integrating it into your workflow.
2. Zero maintenance assumption. Nothing runs itself for a year without a human touch. 5 to 30 minutes a week for the first quarter is realistic on any AI tool.
3. Skipping verification entirely. The biggest miss. Even a "just works" AI tool needs a human eye on 15 to 20 percent of outputs during the first quarter. That is real cost.

#When to buy without running the test

You do not need to run this test if all three are true:

• The task takes 5+ hours a week
• The tool is under $50/month total (SaaS + setup)
• You have someone who can watch it for 15 minutes a day for the first month

At that price and that volume, the numerator almost always beats the denominator. Just try it. Fathom, Reclaim, Superhuman all fit this profile for most founders. Anything above $200/month, run the full test.

#Bottom line

The five-question test is boring math. That is the point. Every AI agent purchase we have seen fail was because the buyer never ran it. Every one that succeeded had a positive answer before signing.

If you want us to run the test on any specific tool you are considering (including anything we might sell you), come talk to us at /get-started. We would rather tell you the numbers do not work now and earn the business later than sell you a losing deal and lose the relationship in month three.