How to Calculate AI Agent ROI: Templates for Support, Sales & Knowledge
Enterprise Agent ROI should use auditable costs and benefits, not demo vibes. This template covers: (1) what you spend in a year; (2) which metrics monetize benefit for support, sales, and internal KB; (3) how to report payback months to finance. Figures are illustrative assumptions, not GeonAI performance promises. Pricing structure: /blog/custom-ai-agent-pricing-factors. Email [email protected], subject “Enterprise Agent inquiry”. /agents are capability references only (not a public trial of all 363+ presets).
One formula—align with finance first
- Annual net benefit = annual monetizable benefit − annual run cost
- ROI (%) = annual net benefit ÷ annual total cost × 100 (total cost = amortized build + run; amortization often 12 or 24 months—write it down)
- Payback months = one-time build ÷ monthly net benefit (annual net ÷ 12)
- Only count benefits with a measurable baseline; “brand lift” stays out unless you have a separate monetization model
Funding sheets need three columns: pessimistic / base / optimistic—and pessimistic should still be investable. Red flags: /blog/ai-agent-project-red-flags.
Cost checklist (miss one item and ROI looks fake)
| Cost line | Typical content | Estimation tip |
|---|---|---|
| Build (one-time) | Discovery, integration, KB cleanup, guardrails, UAT | By milestone; include compliance tests |
| Inference (annual) | API or private GPU | Daily sessions × avg tokens × price; +20–40% peak buffer |
| Human review (annual) | Sampling + high-risk review hours | Often the largest run cost for external CS/finance |
| Knowledge ops (annual) | Doc updates, version sync, QA samples | Person-days by change frequency |
| Platform/channels (annual) | IM, tickets, gateway, monitoring, log storage | Do not forget audit retention |
| Change / phase-2 | New intents, systems, model swaps | Reserve 10–20% or separate change orders |
Template A: Support (deflection + handle time)
Use when you have volume, seat cost, and AHT. The goal is rarely “replace every agent”—it is auto-resolve share × variable cost per contact, minus rework from wrong answers.
- Baseline: monthly contacts Q; human-resolve share; fully loaded seat cost C_seat per FTE-month
- Variable cost per contact ≈ (AHT minutes / 60) × (C_seat ÷ productive hours/month)
- Auto-resolve rate d (only cited, low-risk, no-handoff intents—replace with pilot truth)
- Annual savings ≈ Q × 12 × d × variable cost × (1 − rework rate)
- Subtract: inference + review + knowledge ops
| Assumption (example) | Value | Notes |
|---|---|---|
| Monthly contacts Q | 8,000 | IM + ticket entry |
| Variable cost / contact | ¥12 / ~$1.70 | Backed out from AHT + seat cost |
| Auto-resolve d (base) | 35% | Process/policy only; no refund writes |
| Rework / escalate loss | 8% | Wrong promises → second contact |
| Gross annual savings | ~¥386k | 8000×12×0.35×12×(1−0.08) |
| Annual run (infer+review+ops) | ¥120k | Example |
| Build amortized (12 mo) | ¥150k/yr | Assume ¥150k one-time build |
| Net / ROI | ~¥116k / ~43% | Vs ¥270k annual total cost |
Re-run with d = 20% (pessimistic) and 45% (optimistic). If pessimistic net is negative, shrink intent scope before funding. Support architecture: /blog/custom-customer-service-agent.
Template B: Sales / pre-sales (speed and conversion)
Sales teams often count “more leads” as cash. Finance usually accepts hours saved on follow-up × loaded rate, or attributable pipeline lift with a control. Conversion lift without a control stays an observation—not ROI numerator.
- Hours saved (stable): monthly leads × minutes saved on draft/summary × loaded hourly rate × 12
- Attributed pipeline (careful): only if CRM tags Agent touch and has a control; incremental closed × margin, then apply a ~50% haircut
- Do not count directly: site UV, brand buzz, unattributed revenue YoY
| Assumption (example) | Value | Notes |
|---|---|---|
| Monthly leads worked | 600 | SDR queue |
| Minutes saved / lead | 8 | First reply draft + field summary |
| Loaded hourly cost | ¥180 | Fixed loaded cost, not commission |
| Annual hours saved | ~¥104k | 600×12×(8/60)×180 |
| Attributed margin (conservative) | ¥60k | Only after a control; example |
| Annual run + amortization | ¥140k | Example |
| Annual net | ~¥24k | Hours + lift − cost; tight |
Sales ROI often breaks even on hours first, then validates lift. Playbooks: /blog/sales-assistant-agent-playbook, /blog/presales-consultation-agent-design.
Template C: Internal knowledge / RAG (time-to-find)
The cleanest internal benefit is minutes/week finding docs ↓ × headcount × hourly rate. Rework from wrong/outdated policy is secondary and needs an incident log—or leave it out of the numerator.
| Assumption (example) | Value | Notes |
|---|---|---|
| Covered employees | 200 | Permissioned internal Agent users |
| Baseline find-time | 40 min/person/week | Survey or ticket tags |
| After go-live | 18 min/person/week | Same sample, 4-week mean |
| Saving | 22 min/person/week | |
| Blended hourly cost | ¥120 | Role-weighted |
| Gross annual savings | ~¥458k | 200×22/60×52×120 |
| Annual run + amortization | ¥180k | Includes KB cleanup + ACL ops |
| Net / ROI | ~¥278k / ~154% | Vs annual total cost |
KB ROI is sensitive to citation rate and stale-doc hits: fast wrong answers erase savings. Guardrails: /blog/enterprise-rag-knowledge-base-agent, /blog/deepseek-rag-knowledge-base.
Which template to run first
| If you already have… | Start with | Why |
|---|---|---|
| Volume, AHT, seat cost | A Support | Few variables; finance-friendly |
| CRM handle time, lead tags | B Sales (hours first) | Conversion lift needs a control later |
| Survey/telemetry on find-time | C Knowledge | Usually less external-compliance friction |
| No baseline at all | Measure 4–6 weeks first | Otherwise the ROI sheet is an essay |
90-day measurement plan (make the sheet real)
- Weeks 0–2: freeze intent scope and success definition; align cost accounts
- Weeks 3–6: read-only/draft only; daily log d, handoff rate, missing citations, wrong-promise samples
- Weeks 7–10: recalc base case with real d and inference invoices; cut money-losing intents
- Weeks 11–12: one-pager for finance—assumptions, formula, three scenarios, payback, next investment gate
How to use this with GeonAI
Send your filled base case (with data sources) to [email protected]. We align scope via the 4-step playbook so quotes are not “full seat replacement” while delivery is “FAQ bot.” Include known cost lines and target payback months when you ask for a proposal.
Frequently asked questions
Should the numerator be revenue or cost savings?
Attributed incremental revenue can count; most support/KB projects are safer on cost savings. Do not stuff unattributed revenue YoY into the numerator.
Over how many years should we amortize build?
Match finance policy—often 12 or 24 months. Keep amortization consistent across AI projects or cross-comparisons break.
Does private deployment kill ROI?
It raises build and ops. Treat residency as a constraint, then compare API/VPC/private TCO—not “pick public because ROI looks best.” See the private-deploy guide.
We have no history for auto-resolve rate d.
Label intents for 2–4 weeks and use a pessimistic d; replace with pilot truth. Never paste a vendor demo resolve rate into the sheet.
Can a 5% conversion lift go straight into ROI?
Only with a control or clear attribution window—and haircut it. Otherwise keep it observational and use follow-up hours saved in the numerator.
Will GeonAI contractually guarantee ROI?
No guaranteed outcomes detached from your baseline and scope. We deliver acceptance metrics and milestones; the ROI sheet stays owned by business and finance.