What this calculator answers
Every founder pitching an agentic AI investment eventually lands on the same question: what does payback actually look like? Most ROI calculators are vendor-published, which means the assumptions are often fixed, optimistic, or hidden behind a lead form. This calculator keeps the model exposed. It answers four practical questions: how much labour the agent can remove, how much model usage costs at the token level, how many months it takes to recover the build cost, and whether the investment still looks useful across a 36-month discounted cash-flow view.
The useful output is not a single magic number. It is the relationship between the assumptions. A support agent with 100,000 monthly tickets can look excellent at 60 percent automation and weak at 30 percent automation. A workflow with small prompts can survive retries. A workflow with long context windows and critique loops may spend more on model calls than it saves in labour. That is why the sensitivity table is part of the tool rather than a decorative extra.
How to use it
- Start with monthly task volume and human cost per task from your ticketing, operations, or payroll data.
- Set automation, review, and retry rates conservatively enough that a skeptical operator would accept them.
- Choose the closest model tier and enter realistic input and output token counts from a sample prompt trace.
- Add build, maintenance, ramp, and discount assumptions, then inspect payback, NPV, and sensitivity results.
- Copy the scenario URL and ask your team to challenge the assumptions before using the number in a deck.
What goes into the math
Monthly LLM cost equals input token cost plus output token cost, multiplied by monthly task volume and retry rate. Monthly labour saved equals task volume multiplied by automation rate, adjusted down by the human-in-the-loop rate, and then multiplied by human cost per task. Monthly net savings subtracts model cost and maintenance. Payback divides build cost by net monthly savings. NPV discounts each month of net savings with monthly compounding from the annual discount rate.
The model pricing file cites official provider pricing pages from Anthropic, OpenAI, and Google AI. You should still verify pricing before committing budget, because provider price tables and model names change faster than most annual planning cycles.
Where AI agent ROI usually breaks down
ROI usually breaks when the labour-saving story ignores the production system around the model. Handle-time savings get overstated because the old human workflow had hidden judgment, exceptions, and quality control. Retry rate gets undercounted because prototypes run on clean examples while production receives incomplete, angry, duplicated, or policy-heavy inputs. Prompt iteration, eval design, observability, governance review, and incident response are treated like optional polish even though they are what make an agent safe enough to leave the demo environment.
The best agent ROI patterns are high-volume, repetitive, well-instrumented workflows where the agent can operate inside a narrow policy box: support triage, internal knowledge retrieval, structured extraction, routing, reconciliation, and first-pass QA. The weakest patterns are low-volume bespoke decisions, creative work where review takes longer than doing the work, and regulated tasks where a human must repeat the entire reasoning path anyway.
Assumptions and methodology
This calculator assumes AI agent value comes from avoided labour, not vague productivity uplift. It treats automation rate as the share of tasks completed without full human handling, human-in-the-loop rate as the share of automated tasks still requiring review, and retry rate as additional model attempts caused by failures or validation loops. It does not include revenue uplift, churn impact, brand risk, compliance cost, or opportunity cost unless you add those into the human cost or maintenance assumptions yourself.
Need a defensible implementation plan?
A calculator can challenge the economics. The architecture still needs a scoped workflow, eval plan, observability design, governance model, and rollout path.
Book an architecture reviewFrequently asked questions
What inputs do I need before using this calculator?
You need monthly task volume, average human cost per task, expected automation rate, review rate, retry rate, token assumptions, build cost, maintenance cost, and discount rate. Rough numbers are fine for a first pass, but board-ready ROI should use ticket exports, payroll-loaded cost, and measured token traces from a pilot.
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 monthly maintenance. If monthly net savings is zero or negative, payback is shown as never because the deployment costs more to run than it saves under the current assumptions.
Why do you ask about retry rate?
Retry rate captures failed calls, validation retries, tool errors, and agent loops that require another model pass. It matters because LLM cost scales with every attempt, not just every user-visible task. Under-counting retries is one of the fastest ways to make an agent ROI model look better than production reality.
What's a realistic automation rate for support tickets?
A realistic support-ticket automation rate depends on ticket quality, policy complexity, and escalation rules. High-volume routine categories can land around 40-70 percent after a controlled rollout. Bespoke, regulated, or emotionally sensitive tickets usually need lower automation assumptions and a higher human-in-the-loop rate.
How accurate are the LLM cost estimates?
The calculator uses provider pricing stored in the project data file and multiplies token assumptions by monthly task volume and retry rate. It is accurate for the assumptions entered, but token counts vary by prompt, context length, tool traces, and model family. Treat the result as a scenario model, not an invoice forecast.
Should I include build cost or just run cost?
Include both. Run cost tells you whether the agent is operationally efficient, but payback and NPV need the one-time build cost too. Build cost should include architecture, integration, evals, observability, security review, prompt iteration, and rollout work, not only the first prototype.
What's a sensible discount rate for NPV?
A 10-15 percent annual discount rate is a practical default for early-stage software investments because execution risk is material. Lower rates can fit mature companies with predictable financing costs. The calculator compounds the annual rate monthly across 36 periods so the NPV behaves like a standard investment model.
Can I share my scenario with a teammate?
Yes. The share button encodes your inputs into the URL hash and copies the full link. The hash is read in the browser when the page loads, so the same assumptions reproduce without storing scenario data on a server or requiring an account.