AI Agent ROI Calculator

About this tool

What this calculator answers

Every founder pitching an agentic AI investment lands on the same question: what does payback actually look like? Most ROI calculators online are vendor-published. They start with the conclusion, hide the formulas, and treat model cost as a rounding error. That is not how a serious buyer evaluates automation. A support agent, ops copilot, underwriting assistant, or internal knowledge worker has to survive real task volume, retry behaviour, human review, maintenance, and model-price drift.

This calculator produces four numbers that matter in a board deck or architecture review: monthly labour saved, monthly LLM cost, payback period, and 12 / 24 / 36-month NPV. Monthly labour saved tells you what work the agent plausibly replaces. Monthly LLM cost tells you the token bill before infrastructure and vendor overhead. Payback shows how many months it takes to recover the build cost. NPV discounts future savings so a dollar in month 30 is not treated like a dollar today.

How to use it

Start with the current-state inputs. Monthly task volume should come from your ticketing system, CRM, back-office queue, or event logs. Human cost per task can be calculated from loaded labour cost divided by tasks handled, or estimated from vendor BPO pricing if the work is outsourced. Handle time is not in the core savings formula, but it forces a reality check: if the claimed automation saves fewer minutes than the review process adds, the ROI is fragile.

Then set agent assumptions. Automation rate is the share of tasks the agent completes without human completion. Human-in-the-loop rate is the share of automated tasks that still need review or approval. Retry rate captures failed tool calls, invalid structured output, model refusals, clarification turns, and orchestration loops. Token assumptions should come from a small pilot trace, not a prompt pasted into a playground once.

Finally add build cost, maintenance, discount rate, ramp-up, and decay. Ramp-up matters because teams rarely reach full automation in month one. Decay is usually zero for stable workflows, but it should rise when policies, product catalogues, documents, or user intents change quickly. A strong ROI case still works after these frictions are included.

What goes into the math

Monthly LLM cost equals input tokens times input price plus output tokens times output price, multiplied by monthly task volume and one plus retry rate. Monthly labour saved equals monthly task volume times automation rate times one minus human-in-the-loop rate, multiplied by human cost per task. Monthly net savings subtracts LLM cost and maintenance. Payback is build cost divided by monthly net savings. NPV uses monthly compounding from the annual discount rate across 36 periods.

Assumptions and methodology: the calculator estimates cash impact from task automation, not total enterprise value. It assumes the human cost per task already includes wages, management overhead, software seats, and vendor margin where relevant. It treats LLM usage as variable cost, maintenance as monthly fixed cost, build cost as a month-zero outflow, and ramp-up as a linear path to full automation. It does not assume revenue uplift unless the user manually converts that uplift into cost-per-task value.

Where AI agent ROI usually breaks down

ROI breaks when the saved work is overstated. A vendor demo may handle a perfect ticket in 20 seconds, but production queues include duplicates, missing fields, angry customers, policy exceptions, and edge cases that require judgement. If human review takes nearly as long as doing the work from scratch, automation rate becomes a vanity number.

Retry rate is the second leak. Agents retry tool calls, repair JSON, fetch another document, ask clarification questions, or run a critique pass. Those loops are often invisible in a slide deck but very visible on the invoice. Prompt iteration, eval design, observability, governance review, and escalation tooling also need budget. A serious deployment needs traces, red-team cases, rollback paths, and someone accountable for drift.

Model-price drift across 36 months is another risk. OpenAI, Anthropic, and Google publish token prices on their official pricing pages, and those prices change as models and service tiers evolve. This tool stores pricing in JSON with a last-verified date so the assumption is visible. Treat the sensitivity table as the antidote to false precision: if the case only works when every assumption is perfect, it is not a durable ROI case.

When the ROI is real and when it is not

Agentic ROI is usually strongest in high-volume, routine, structured work: L1 support, internal knowledge retrieval, document triage, compliance pre-checks, lead enrichment, and operations queues with clear escalation criteria. These workflows have enough volume to amortise build cost and enough repetition for evals to catch regressions.

The ROI is weaker for creative work, low-volume bespoke work, relationship-heavy work, and tasks where the human review burden is larger than the original task. If the agent mostly creates drafts that a senior person rewrites, you may still have product value, but it is not labour-deflection ROI. Use a different business case for quality, speed, or revenue.

How this calculator differs from vendor calculators

This calculator is vendor-neutral. Every assumption is editable. The formulas are shown in plain English with values plugged in. Pricing is kept in a repo data file instead of being buried in a component. The goal is not to prove that an AI agent should be built. The goal is to reveal the assumptions that make the investment worth building or worth rejecting.

Sources used for pricing and methodology include the official OpenAI API pricing page, Anthropic pricing documentation, Google Gemini API pricing, and industry research such as McKinsey's State of AI reporting on the gap between AI adoption and measurable business value.

Frequently asked questions

What inputs do I need before using this calculator?

You need monthly task volume, approximate human cost per task, expected automation rate, expected human review rate, token assumptions, model choice, build cost, and monthly maintenance cost. If you do not know a value, start with the defaults and replace them with support desk, operations, or finance data once available.

How is payback period calculated?

Payback period is build cost divided by monthly net savings. Monthly net savings equals labour saved minus LLM cost minus maintenance. If monthly net savings is negative or zero, payback is shown as never because the system does not recover its build cost under the current assumptions.

Why do you ask about retry rate?

Retry rate matters because agent workflows often call the model again after failed tool calls, invalid JSON, policy checks, or user clarifications. The calculator multiplies token cost by one plus retry rate. A 10% retry rate means the same task volume creates roughly 10% more model spend.

What's a realistic automation rate for support tickets?

A realistic support automation rate depends on ticket quality and escalation policy. For clean, repetitive L1 tickets, 40% to 70% can be plausible after a pilot. For messy, account-specific, compliance-heavy, or emotionally sensitive work, a lower rate is safer until you have measured deflection and review time.

How accurate are the LLM cost estimates?

LLM cost estimates are only as accurate as the token assumptions and current provider prices. This page stores provider pricing in a JSON data file with a last-verified date, then applies the formula input tokens plus output tokens times task volume and retry multiplier. Real bills can differ when tools, caching, or long-context premiums apply.

Should I include build cost or just run cost?

Include build cost if you are making an investment decision, preparing a board deck, or comparing vendors. Run cost alone is useful for operations planning, but ROI and payback require the one-time build cost because the first question is usually how long it takes to recover the upfront spend.

What's a sensible discount rate for NPV?

A sensible discount rate for startup planning is often 10% to 20% annually, with 12% as a reasonable default. Use a higher rate when the project is risky, the cash runway is tight, or the automation depends on uncertain adoption. Use your finance team's hurdle rate if one exists.

Can I share my scenario with a teammate?

Yes. The share button encodes the calculator inputs into the URL hash. Anyone opening that link sees the same task volume, model, token assumptions, rates, build cost, and discount rate. The encoded scenario is local to the URL and does not require a login or server-side storage.