Agentic AI TCO + ROI Builder

The real cost of an AI agent is much more than the model bill. This tool adds up every cost - usage, safety, human review, compliance, monitoring, keeping it current - and tells you whether the payoff is worth it. It starts with sensible example numbers so you get an answer right away; open any section to put in your own. Anything still using an example is clearly marked.

Not sure how this works? Read the plain-English guide →

What this calculator does

Builds a board-grade agentic TCO and ROI from raw drivers. Each cost layer (tokens, retrieval, guardrails, eval, human review, compliance, observability, integration, data-freshness, model-maintenance) is computed by its own live calculator engine behind a single seamless form, then rolled up and weighted by cancellation risk into a GO / CONDITIONAL / NO-GO verdict. Layers left on defaults are flagged so a board sees what is real and what is a placeholder.

Why use it
  • One seamless form instead of bouncing between ten calculators
  • Every layer number comes from its own validated engine, so adapter-priced layers stay live-priced
  • Progressive disclosure: start instantly on defaults, open only the layers that matter
  • Flags incomplete data, so the number is honest about what you have and have not entered

Build your TCO + ROI

Put in your numbers, see your total monthly cost, and find out whether the payoff is worth it.

Scope it the way you would any project. Go top to bottom - every box already has a sensible starting number, so you get an answer right away and only refine what matters.

  1. What is the win? The value you expect once it works.
  2. Getting it built. What it takes to stand it up.
  3. Running it. What you pay day to day.
  4. Keeping it healthy. Quality, fresh data, upkeep.
  5. Handling the risks. Safety, review, compliance, downtime.
  6. How ready are you? A checklist that lowers the chance it stalls.
1. What is the win? what you expect this to be worth
Not sure? Try people helped x hours saved x their hourly cost, or extra revenue per month. A rough number is fine to start.
How long before it reaches full value. 3 to 6 months is typical.
The period to judge it over. 36 months (3 years) is common.

2. Getting it built what it takes to stand it up

Integrations and connectors guideauto
Systems you connect to: CRM, email, databases, and so on.
Simple = a clean API; complex = legacy or messy systems.
Engineer-days to build each one. 5-15 is common.
Advanced drivers (4)
Fully-loaded cost per engineer per day. ~$800-1,200.
Yearly upkeep as a share of build cost. ~15-25%.
One-time cost to scope and plan, if any.
Years to spread the build cost over. 3 is common.

3. Running it, day to day what you pay to keep it running

AI usage (the model bill) guideauto
How many calls the AI handles a day. Active users x calls each, or pull it from your logs.
Roughly the words you send in, x1.3. Short prompt ~300, long with context ~3,000+.
Roughly the words it sends back, x1.3. A sentence ~50, a paragraph ~300, a long answer ~1,500.
Advanced drivers (2)
Cheap for routine work, balanced for most, premium for the hardest reasoning.
30 if it runs every day; ~22 for business days only.
Search and retrieval guideauto
How many documents the AI searches over. Count the files in your knowledge base.
Searches a day. Usually close to your number of requests.
Advanced drivers (4)
Words per document x1.3. A one-pager ~600, a long doc ~3,000+.
Cheap is fine for most search; premium only for subtle matching.
How often you re-index everything from scratch. Once or twice a year is common.
Your monthly vector-database bill, if you pay for one. Leave 0 if unsure.
Storage guideauto
Total data you store in TB: models, training data, vector DB, logs, backups.
Storage price per TB per month. Cloud is ~$20-25.
Network (data out) guideauto
Data sent out per month in TB: to users, across regions, CDN.
Data-out price per TB. Cloud egress is ~$80-90.

4. Keeping it healthy quality, fresh data, upkeep

Quality checks guideauto
Test cases you check quality against. 50-500 is typical.
How often you run the test set. Weekly ~4, daily ~30.
Advanced drivers (4)
Words per test case sent in, x1.3.
Words per test answer, x1.3.
Size of the grading instructions, in tokens. ~500 is common.
Size of each grade the judge writes back. ~100.
Monitoring guideauto
Total requests a month.
Steps recorded per request. Simple ~3-5, agentic ~20-50.
How your tracing tool bills: per span, per trace, or self-hosted.
Advanced drivers (5)
Share of traffic you trace. 100% at low volume, less at scale.
Your tool's price per million spans, if it bills that way (~$2).
Your tool's price per 1,000 traces, if it bills that way (~$2.50).
Extra cost for keeping data longer. 1 = standard, 2 = long retention.
Monthly server cost if you self-host tracing.
Keeping data fresh guideauto
Total size of your knowledge base in tokens (~words x1.3).
How often you update the data. Daily ~30, weekly ~4.
Share of data that changes each refresh. Often 1-5%.
Advanced drivers (4)
Price to re-embed per million tokens (~$0.10). A guess is fine.
Compute cost to run one refresh job (~$50-100).
How often you re-embed everything, e.g. after a model change. 6-12 months.
Monthly cost of tools that watch for data drift, if any.
Model upkeep guideauto
How often a vendor retires your model and forces a move. 1-2 a year lately.
Engineer-days to move and re-test per migration. ~10-20.
Days a month spent tuning prompts and fixing drift. 1-3.
Advanced drivers (3)
Fully-loaded cost per engineer per day. ~$800-1,200.
How often you re-tune or fine-tune. 0-2 a year.
Cost of one re-tune: data, compute and time. $2k-20k.

5. Handling the risks around it safety, review, compliance, downtime

Safety guardrails guideauto
Total AI calls a month. Roughly requests/day x days.
Share of calls you check for safety. 100% for sensitive uses, less if you sample.
Checks per call, e.g. input + output = 2.
Advanced drivers (3)
Words sent to the safety check x1.3. Usually similar to the prompt.
Words the safety check returns. Usually small, ~50.
Monthly fee for a guardrail tool, if you use one. Leave 0 if not.
Human review guideauto
Total AI actions a month.
Share a person reviews. Start high (10-20%) and lower as you trust it.
Minutes a person spends per review. 1-5 is common.
Advanced drivers (4)
Reviewer's fully-loaded cost per hour (pay + overhead). Often ~1.4x their wage.
What one missed mistake costs you: a refund, a fix, a lost customer.
Problems that slip through per 1,000 unreviewed actions. A guess is fine.
Share of problems your reviewers actually catch. 80-95% is typical.
Compliance and audit guideauto
How regulated this use is. High = decisions about people (hiring, credit, health).
How many separate AI systems fall under this.
People-equivalents on compliance. 0.25 = one person a quarter of their time.
Advanced drivers (6)
Fully-loaded yearly cost of one compliance person. ~$150k-200k.
Yearly cost of outside audits, if any.
Yearly cost of compliance/governance tooling, if any.
Audit data stored a month, in GB. A guess is fine.
How long you keep logs. Often 12-24 months by policy.
Storage price per GB per month. Cloud is ~$0.02-0.10.
Vendor lock-in cushion guideyou enter
What you'd spend to avoid being stuck with one vendor: a fallback provider or abstraction layer. A small monthly insurance.
Downtime (lost revenue) guideauto
Hours a month the system is unavailable. Even 1-2 adds up.
Revenue you would normally earn in an hour the system serves.

Other costs easy-to-miss extras

Data prep and other extras guideyou enter
Anything else: data prep and labeling, one-off rollbacks, small extras.
6. How ready are you? tick what is true - each one lowers the chance it stalls
Across the industry about 40% of AI agent projects get shelved. Ticking the boxes above brings this down.
Per-unit costs optional - to see cost per request, user and outcome
How many people use it in a month. Leave 0 to skip per-user cost.
Real results a month: tickets resolved, docs processed, leads handled.
Name the result you care about, e.g. ticket resolved.

Verdict

0% your data · 15 of 15 sections still use example numbers. Fill them in for an accurate total.
Where this project stands
GO The payoff beats the full cost, even after allowing for risk. Build it.
RISKY The payoff works and it pays back, but the margin is thin. De-risk before you commit budget to scale. you are here
NO-GO Even before risk, the payoff does not cover the full cost. The case is not there yet.
Return after risk is 38% with payback in month 4, but return after risk below the 100% GO line. Close these before committing budget to scale.
What it takes to reach GO
  1. Lift return after risk to 100%. It is 38% now; the GO line is 100% (every $1 returns about $2 after risk). Raise the monthly value, cut your top cost layer, or lower the cancellation risk to close the gap.
  2. Tick the remaining readiness checks. We measured the 'before', so we can prove the improvement; One person clearly owns this project. Each one lowers the cancellation risk, which lifts the return.
What this means Over 36 months, after allowing for the 31% chance the project is cancelled, every $1 spent returns about $1.38 (a 38% return), and the project pays for itself in month 4. The model bill is only 0% of the cost; the other 100% is the work around it.
Return after risk
38%
before risk 99%
Pays for itself in
4 mo
from launch
Total cost / month
$213,700
$2,564,396 / year
Risk it gets cancelled
31%
from 40%, 4/6 checks done

Cost per unit (your total cost spread across your volume)

Per request
$7.1233
30,000 / mo
Per 1k tokens
$2.8493
 
Per user / month
$10.68
20,000 users

The report at a glance

Cost by area
$213,700 per month
Handling the risks around it94%
Keeping it healthy3%
Other costs1%
Getting it built1%
Running it, day to day0%
Pays for itself in month 4
month 1 month 36
after riskbefore riskpays back

Where the money goes (each cost worked out separately)

Human review default
$180,00084%
Compliance and audit default
$16,5368%
Model upkeep default
$4,6672%
Data prep and other extras default
$2,5001%
Vendor lock-in cushion default
$2,0001%
Downtime (lost revenue) default
$2,0001%
Integrations and connectors default
$1,9171%
Keeping data fresh default
$1,7051%
Safety guardrails default
$1,2481%
Network (data out) default
$4500%
Storage default
$3750%
AI usage (the model bill) default
$2230%
Quality checks default
$460%
Monitoring default
$200%
Search and retrieval default
$140%

How to bring the cost down (based on your numbers)

  1. Start with your biggest cost
    Human review is $180,000/mo (84% of the total). Lower the share of actions sent to a person, or cut minutes per review - usually the single biggest lever.
  2. Most of the cost is not the model
    Only 0% of your spend is the model itself, so shopping for a cheaper token price barely moves the total. Put your effort into the large non-model costs above.
  3. Close the readiness gaps
    Not yet ticked: We measured the 'before', so we can prove the improvement; One person clearly owns this project. Each one lowers the chance of cancellation.
  4. Lift the return
    To clear the GO line, raise the monthly value by pointing the AI at higher-value work, or cut your top cost above. Closing the readiness gaps also shrinks the cancellation discount.
  5. Fill in your real numbers
    15 sections still use example numbers, so the total and per-unit costs are rough until you enter your own.
Quality safety-net is covered
An AI gets worse over time unless you both catch problems (quality checks) and fix them (refreshing data and updating the model). Both are budgeted here, so quality holds up.
Board one-liner
RISKY: 38% return after risk over 36 mo, payback month 4, at a 31% cancel-risk discount. Hidden costs = 100% of total cost.
Verdict RISKY RW ROI 38% TCO/mo $213,700 Data 0%

Each cost is worked out by its own dedicated calculator, and the ones tied to live pricing update on their own. Yearly costs are shown per month. The "risk it gets cancelled" number is the chance the project is dropped before it pays off - the readiness checklist lowers it. Anything still on an example number is a placeholder, not your figure - fill those in before you rely on the total. This is a budgeting aid, not financial advice.

📖 Data sources & methodology 158 text models · 9 embeddings · 37 vision · 55 audio · 8 vector DBs across 10 vendor pages · last verified 2026-07-21

Methodology

  • All prices are USD per 1 million tokens, current as of 2026-07-21.
  • Vendor-published values have no mark. Inferred/extrapolated values are marked with * and listed below.
  • Batch API discounts are 50% off standard rates across providers that offer Batch mode.
  • Prompt caching discounts vary by provider (typically 80-90% off cached input tokens).
  • Regional data-residency surcharges (Anthropic 1.1x, OpenAI 1.1x, Google regional tiers) are NOT included in base rates.
  • Long-context pricing tiers apply when input exceeds model threshold.
  • Embedding prices are input-only (no output tokens generated).

Primary sources

Last-verified date is the most recent successful daily snapshot (aicost_pricing_snapshots) or, when no snapshot exists yet, the latest successful crawler run (aicost_crawler_runs). 10 of 10 vendors are currently verified. Aggregator services (TokenCost, AI Pricing Guru, etc.) are not listed.

Anthropic
2026-07-21
https://www.anthropic.com/pricing
Daily snapshot since Sep 2023 · 624 days captured
Anthropic Docs
2026-07-21
https://platform.claude.com/docs/en/about-claude/pricing
Daily snapshot since Sep 2023 · 624 days captured
OpenAI
2026-07-21
https://openai.com/api/pricing/
Daily snapshot since Sep 2023 · 625 days captured
Google AI
2026-07-21
https://ai.google.dev/gemini-api/docs/pricing
Daily snapshot since Dec 2023 · 600 days captured
Google Vertex
2026-07-21
https://cloud.google.com/vertex-ai/generative-ai/pricing
Daily snapshot since Dec 2023 · 600 days captured
DeepSeek
2026-07-21
https://api-docs.deepseek.com/quick_start/pricing
Daily snapshot since May 2024 · 539 days captured
xAI
2026-07-21
https://x.ai/api
Daily snapshot since Nov 2024 · 457 days captured
Mistral
2026-07-21
https://mistral.ai/pricing
Daily snapshot since Dec 2023 · 598 days captured
Cohere
2026-07-21
https://cohere.com/pricing
Daily snapshot since Sep 2023 · 624 days captured
Voyage AI
2026-07-21
https://docs.voyageai.com/docs/pricing

Inferred values (marked with * in calculator tables)

Derived from industry conventions, not directly published by the vendor. Typical conventions: cached input = 10% of base (90% off), Batch API = 50% of base (50% off).

Vendor / Model Field Why it’s inferred
Anthropic — Claude Sonnet 4.6 cachedInput Derived at 10% of input rate — Anthropic publishes 90% cache-hit discount on this tier.
Anthropic — Claude Sonnet 4.5 cachedInput Derived at 10% of input rate; same 90% cache-hit convention as Sonnet 4.6.
Anthropic — Claude Sonnet 4.5 batchInput Derived at 50% of standard input — Anthropic documents uniform 50% Batch discount.
Anthropic — Claude Sonnet 4.5 batchOutput Derived at 50% of standard output — Anthropic documents uniform 50% Batch discount.
Anthropic — Claude Haiku 4.5 cachedInput Derived at 10% of input rate — Anthropic 90% cache-hit discount convention.
OpenAI — GPT-5.4 Mini cachedInput Derived at 10% of input — OpenAI documents automatic 90% discount on cache hits across GPT-5.x tier.
OpenAI — GPT-5.4 Nano cachedInput Derived at 10% of input — OpenAI 90% cache-hit convention.
OpenAI — GPT-5.4 Nano batchInput Derived at 50% of input — OpenAI Batch API uniform 50% discount.
OpenAI — GPT-5.4 Nano batchOutput Derived at 50% of output — OpenAI Batch API uniform 50% discount.
OpenAI — GPT-5.4 Pro cachedInput Derived at 10% of input — OpenAI 90% cache-hit convention.
OpenAI — GPT-5.4 Pro batchInput Derived at 50% of input — OpenAI Batch API uniform 50% discount.
OpenAI — GPT-5.4 Pro batchOutput Derived at 50% of output — OpenAI Batch API uniform 50% discount.
OpenAI — GPT-5.2 cachedInput Derived at 10% of input; no residency uplift.
OpenAI — GPT-5.2 batchInput Derived at 50% of input.
OpenAI — GPT-5.2 batchOutput Derived at 50% of output.
OpenAI — GPT-5 cachedInput Derived at 10% of input.
OpenAI — GPT-5 batchInput Derived at 50% of input.
OpenAI — GPT-5 batchOutput Derived at 50% of output.
OpenAI — GPT-5.5 Pro cachedInput Derived at 10% of input — OpenAI does not publish a cached rate for *-pro models; using the family convention.
OpenAI — GPT-5.5 Pro batchInput Derived at 50% of input.
OpenAI — GPT-5.5 Pro batchOutput Derived at 50% of output.
OpenAI — GPT-5.2 Pro cachedInput Derived at 10% of input — pro-tier convention.
OpenAI — GPT-5.2 Pro batchInput Derived at 50% of input.
OpenAI — GPT-5.2 Pro batchOutput Derived at 50% of output.
OpenAI — GPT-5.1 batchInput Derived at 50% of input.
OpenAI — GPT-5.1 batchOutput Derived at 50% of output.
OpenAI — GPT-5 Pro batchInput Derived at 50% of input.
OpenAI — GPT-5 Pro batchOutput Derived at 50% of output.
OpenAI — GPT-5 Nano cachedInput Derived at 10% of input.
OpenAI — GPT-5 Nano batchInput Derived at 50% of input.
OpenAI — GPT-5 Nano batchOutput Derived at 50% of output.
Google — Gemini 3 Flash cachedInput Derived at 10% of input — Google caching discount convention ~90%.
Google — Gemini 3.1 Flash-Lite cachedInput Derived at 10% of input — Google caching convention.
Google — Gemini 3.1 Flash-Lite batchInput Derived at 50% of input — Google Batch API uniform 50% discount.
Google — Gemini 3.1 Flash-Lite batchOutput Derived at 50% of output — Google Batch API uniform 50% discount.
Google — Gemini 2.5 Pro cachedInput Derived at 10% of input.
Google — Gemini 2.5 Flash cachedInput Derived at 10% of input.
Google — Gemini 2.5 Flash-Lite cachedInput Derived at 10% of input — Google caching convention.
Google — Gemini 2.5 Flash-Lite batchInput Derived at 50% of input — Google Batch API uniform 50% discount.
Google — Gemini 2.5 Flash-Lite batchOutput Derived at 50% of output — Google Batch API uniform 50% discount.
Google — Gemini 2.0 Flash cachedInput Derived at 25% of input per Google 2.0 family caching rates.
Google — Gemini 2.0 Flash batchInput Derived at 50% of input — Google Batch API uniform 50% discount.
Google — Gemini 2.0 Flash batchOutput Derived at 50% of output — Google Batch API uniform 50% discount.
Google — Gemini 2.0 Flash-Lite cachedInput Derived at 10% of input — Google caching convention.
Google — Gemini 2.0 Flash-Lite batchInput Derived at 50% of input — Google Batch API uniform 50% discount.
Google — Gemini 2.0 Flash-Lite batchOutput Derived at 50% of output — Google Batch API uniform 50% discount.
xAI — Grok 4 (legacy) cachedInput Extrapolated at 25% of base.

Pricing is cross-verified against the LiteLLM community registry when available. Daily snapshots are kept in aicost_pricing_snapshots; every change is logged to aicost_price_changelog with old & new values for full audit trail. Read the full methodology →