Before you commission an AI agent, put a number on the work it would take on. Not a vendor slide—a worksheet your finance lead can challenge. This post walks through a practical ROI model Phi Digital uses with prospects: hours × fully loaded cost × automation share, minus build, usage, and optional retainer.
Use the interactive AI Agent ROI Calculator for wage-based scenarios. Sources on that page were checked 2026-10-25; this article does not invent BLS wages—it points you there.
Why ROI math fails for AI projects
Three patterns kill credibility:
- Counting 100% of a role as automated when the agent handles a slice of repetitive tasks
- Ignoring model/API usage billed to your own provider accounts
- Comparing only year-one cash without separating one-time build from steady-state savings
Agents usually augment a team. Treat “automation share” as an explicit assumption, not a promise.
The core formula
Annual labor addressed (estimate)
hours_per_week × 52 × fully_loaded_hourly_cost × people × automation_share
First-year net (estimate)
labor_addressed − build − first_year_usage − first_year_retainer
Steady-state net (estimate)
labor_addressed − annual_usage − annual_retainer
Payback (months, rough)
build ÷ (monthly_labor_addressed − monthly_usage − monthly_retainer)
Label every output as an estimate. Scope changes the build price; volume changes usage.
Step 1 — Measure the work
Pick one workflow (refunds, lead research, invoice extraction). For two weeks, log:
- Hours spent per person per week
- How many people do it
- Percentage that is repetitive and rule-bound (your automation share)
If you cannot measure hours, you are not ready to buy a custom agent. Start with the free AI audit or an AI Assessment.
Step 2 — Fully loaded cost
Hourly wages understate cost. Fully loaded cost includes benefits and overhead. The ROI calculator uses U.S. BLS occupational medians and a benefits uplift from Employer Costs for Employee Compensation (details and SOC codes live on the calculator sources—checked 2026-10-25). Prefer that page over copying wage numbers into a slide deck from memory.
If you operate outside the U.S., use your own payroll fully loaded rates. The structure of the formula stays the same.
Step 3 — Automation share (be conservative)
Example thinking, not a claim: an agent might handle 25–70% of a repetitive queue while humans keep exceptions, relationship work, and approvals. Phi’s own product work shows agents can ship large volumes of code under human gates—but client ROI still hinges on your share assumption. See running products with an AI agent fleet for measured engineering throughput, which is a different metric than labor replacement.
Step 4 — Subtract Phi costs (published starting prices)
From AI agent services (USD starting prices, not guarantees):
| Cost | Starting point |
|---|---|
| AI Assessment | $2,500 (~1 week), fully credited against a build if you proceed |
| Single Agent build | from $5,000 |
| Agent Suite | from $12,000 |
| Managed retainer | from $2,000/month |
| Website Agent | $990 setup + $299/month (first payment $1,289) |
Builds are typically billed 50% upfront and 50% on delivery. You own the code built for you; Phi licenses pre-existing components for use in your system. Model/API usage is billed to your accounts—not included as “free” in the build.
For small-business site Q&A, compare Website Agent economics separately: A Website AI Chat Agent for Small Businesses. Do not force a $5k+ custom agent into a FAQ job.
Step 5 — Model usage and retainer
Usage. Estimate monthly tokens/API calls from pilot volume, then annualize. Wrong by 2× is still better than $0.
Retainer. Optional. Include from $2,000/month if you want monitoring, model upgrades, and ongoing improvements—toggle this on the ROI calculator.
Assessment credit. If you buy the assessment and then a build, the $2,500 is credited—so do not double-count it as a permanent cost in the go-forward case.
Hypothetical worked example
Hypothetical only—not a customer result.
Assume:
- 2 people × 15 hours/week on repetitive ticket triage
- Fully loaded cost: $50/hour (your number; replace with payroll or calculator occupation)
- Automation share: 40%
- Build: Single Agent starting at $5,000
- Usage assumption: $500/month
- No retainer in year one
Labor addressed:
15 × 52 × $50 × 2 × 0.40 = $31,200 / year
First-year costs:
$5,000 + ($500 × 12) = $11,000
First-year net estimate:
$31,200 − $11,000 = $20,200
Monthly surplus after build month (rough):
($31,200 − $6,000) / 12 ≈ $2,100 → payback on $5,000 on the order of a few months under these assumptions.
Change share to 20% or usage to $1,500/month and the story changes. That sensitivity analysis is the point.
For wage-table scenarios (customer service, data entry, bookkeeping, SDR-adjacent roles, admin, paralegal), use the calculator’s replacement-cost panel rather than inventing medians here.
What not to put in the spreadsheet
- Invented testimonials or “industry average ROI”
- 100% headcount elimination without a transition plan
- Recruiting, training, quality, and speed gains as precise dollars unless you have a measurement plan
- Guaranteed payback dates from a vendor
Phi publishes process timing (Assessment ~1w → Architecture 1w → Build & Test 2–4w → Deploy & Monitor 1w) so you can place costs on a calendar—not so you can claim a fixed ROI.
Choosing the right product after the math
| Outcome of the worksheet | Sensible next step |
|---|---|
| Site FAQ value only | Website Agent |
| Unclear workflows / systems | Free audit or $2,500 assessment |
| Clear single workflow, positive net | Single Agent from $5,000 |
| Several coordinated workflows | Agent Suite from $12,000 |
| Need ops after launch | Retainer from $2,000/month |
Chatbot vs agent fit: AI Agent vs Chatbot. Assessment contents: What an AI Agent Assessment Includes.
FAQ
Should I include the assessment fee in ROI?
Include it in the discovery path. If you proceed to a build, it is credited—model the “proceed” case without counting it twice.
Can I use the calculator for Singapore payroll?
Use your local fully loaded rates in the hours model. The occupation table on the calculator is U.S. BLS-based (sources checked 2026-10-25). Singapore companies considering the assessment may pay S$3,500 PayNow to Phi Intelligence via hi@phi.llc—that is a payment path for the assessment, not a different ROI formula.
Is payback the only metric that matters?
No. Error rate, cycle time, and coverage of after-hours work matter. Start with cash ROI so the project survives scrutiny, then add quality metrics in the build plan.
Next step
Run your numbers on the ROI calculator, then contact with the workflow, share assumption, and systems list—or buy the assessment online from services if you want a fixed-scope quote with the fee credited toward the build.
Share the worksheet with both ops and finance
Ops owns hours and share assumptions. Finance owns fully loaded rates and payback thresholds. If those two never meet, AI projects become either underfunded science or overpromised magic. Bring both to the assessment kickoff—or at least attach the sheet to the contact note.