What Does an AI Tool for a Sustainability Team Cost?

The model is the cheapest line in an AI tool's budget. Anthropic's Claude Sonnet 5.5, OpenAI's gpt-6.1-sol and Google's Gemini 4 Argon all list $2 per million input tokens as of October 29, 2026, and Anthropic's documentation puts a typical PDF page at 1,500 to 3,000 text tokens. At those rates, reading thousands of pages a year costs hundreds of dollars. The build, and the people who check its work, cost far more.

A custom AI tool that reviews documents, extracts data or drafts reports for a sustainability team costs $25,000 to $70,000 to build at the published prices of CF Labs, Council Fire's software studio. Running it costs less in model fees than in staff time: a few hundred dollars of tokens a year, against thousands in review.

The staff line is where budgets go wrong. Environmental scientists and specialists earned a median of $39.53 an hour in May 2025, and wages make up 70.0% of private employers' compensation costs, so an hour of a specialist's review costs roughly $56. Software also keeps costing after launch: U.S. federal agencies report spending about 80% of their IT budgets operating and maintaining systems they already have.

Read our figures as one studio's published pricing, not a market survey. Council Fire has been a certified B Corp since 2010, and through CF Labs we build AI tools of this kind. In Aptation, our resilience planning platform, AI drafts scores and funding matches and staff confirm every one; Trovance drafts fixes for a brand's visibility in AI search, and nothing publishes until a person approves it. Building them has made us more careful about recommending them.

In the CF Labs Radar we published today, we rate AI that answers only from your own documents as ready for a trial, and we put unreviewed AI drafts in plans, grants and disclosures on hold. This guide prices the first and budgets for the review that prevents the second.

What does it cost to build an AI tool for a sustainability team?

At CF Labs' published prices, a custom AI assistant or automation costs $25,000 to $70,000 to build before multipliers, and the fixed-price first steps cost $4,500 to $12,500. Scope moves that price far more than the choice of model does. The project estimator runs the arithmetic from seven questions; here it is by hand.

Three fixed-price ways to start

• Data & AI Readiness Review: $6,500, two weeks. An inventory of your data sources and reporting workflows, interviews with the people who produce and use the numbers, a shortlist of AI uses each checked for risk against the Montréal Declaration for Responsible AI, and a 90-day roadmap with costs.
• Tool Sprint: $12,500, two to three weeks. A working prototype on your own documents, hosted for you, with data loaded from one source, plus a walkthrough, documentation and a plan for the production build. For AI work, your reviewer approves every output during the sprint.
• Grant-Ready Tool Scope: $4,500, one to two weeks. The scope, line-item budget, technical approach and timeline a grant application needs when the funding covers a digital tool.

All three prices sit below the $15,000 federal micro-purchase threshold, a line public buyers spending federal money watch closely; your own procurement code may set a lower one. Our page on AI for sustainability teams describes what these builds do, from document review and data extraction to internal assistants and drafting with review.

The estimator's range, and what moves it

For a custom AI assistant or automation, the estimator starts at $25,000 to $70,000 over 6 to 12 weeks. Four answers multiply that range:

• Data sources: one ×1, two or three ×1.15, four or more ×1.35.
• Users: your team ×1, your team and partners ×1.1, the public ×1.2.
• Integrations: none ×1, a few such as spreadsheets or a CRM ×1.1, many or complex ones such as GIS, ERP or APIs ×1.3.
• Timing: within six weeks ×1.1, this quarter or flexible ×1.

Add-ons are priced in dollars on top: sign-in and roles at $5,000 to $10,000, maps and GIS at $6,000 to $15,000, data that updates on its own at $4,000 to $10,000, and AI features at $8,000 to $20,000. The AI add-on is for putting AI inside another kind of build, and the estimator does not charge it twice when the build is already an AI tool. Adding AI features to a calculator or scorecard, for example, moves its $12,500 to $25,000 range to $21,000 to $45,000.

With every multiplier at its top value, the base range grows about 2.3 times before any add-ons. Totals round to the nearest $1,000, or to the nearest $500 below $20,000. The result is a planning range, not a quote: CF Labs confirms scope and a fixed price on a 20-minute call.

For AI builds, the estimator does not recommend starting with a build at all. It points to the Readiness Review, to find the AI uses worth building and the data they need before committing to one, or to the Grant-Ready Tool Scope when a grant is paying.

A worked example: supplier questionnaires

Take a hypothetical corporate sustainability team that receives about 400 supplier questionnaire responses a year, each around 20 pages of PDFs and spreadsheets, the raw material behind its CDP disclosure. It wants an assistant that pulls 40 values from each response into a review queue, cites the page behind every value and holds each record until a named analyst approves it. In the estimator, that is two or three data sources, used by the team, with a few integrations, sign-in and roles so reviewers and approvers get different permissions, and delivery this quarter.

The estimator puts the timeline at 7 to 14 weeks. The range is wide because nobody has scoped the work yet; a Tool Sprint on one of those sources, or the Readiness Review, is how it narrows.

Who this is for: sustainability, ESG and program leads who need a defensible number for a budget request before anyone has written a scope. Our one-workflow build brief helps you write that scope.

What does an AI tool cost to run each year?

In our worked example, running the tool costs about $9,800 a year, and roughly 92% of that is staff time. Price five lines: model usage, hosting, monitoring, maintenance and human review. The example keeps the supplier scenario and labels every assumption, so you can swap in your own.

Model and API usage

Providers charge per million tokens, and output costs more than input. These are the list prices on each provider's own page on October 29, 2026:

Our example reads 400 responses of about 20 pages, or 8,000 pages a year. We plan on 8,000 input tokens a page: Anthropic counts 1,500 to 3,000 text tokens for a typical page and, because each page also goes in as an image, up to 4,784 image tokens on its newest models. Add 2,000 output tokens per response for the extracted values and their page references, then double everything for testing, re-runs and corrections.

That comes to 128 million input tokens and 1.6 million output tokens a year:

• About $272 a year on any of the three $2 models, and about $136 at batch rates.
• About $14 on Claude Haiku 5.5 or gpt-6-luna at their short-prompt rates. Claude Haiku 5.5 costs five times as much for prompts over 100,000 tokens, so split long documents.
• About $1,360 on top-tier models priced at $10 / $50, such as Claude Fable 5.1 and gpt-6-astra.

Two cautions. These are planning figures, and each provider counts tokens its own way, so count the tokens in a sample of your own files before you sign. Prices also move: Google's page schedules Gemini 3.8 Flash to double, from $0.75 to $1.50 per million input tokens, on January 1, 2027.

Hosting

A small internal tool needs somewhere to run and a database. For scale, Vercel lists its Pro plan at $20 a month with $20 of usage included, and Supabase lists its Pro plan from $25 a month, together $540 a year before extra usage. The free tiers are not meant for this work: Vercel's Hobby plan is for personal, non-commercial use, and Supabase pauses free projects after a week of inactivity. Ask any builder who holds the hosting accounts; CF Labs either hosts what it builds or hands it over with documentation and training.

Monitoring and evaluation

Someone has to keep checking that the tool still gets it right after launch. In our example, a reviewer re-checks 20 approved records a month against their source pages at 15 minutes each, or 60 hours a year: about $3,400 at the rough loaded rate of $56 an hour. For state and local governments, wages were 61.2% of compensation costs in June 2026, which lifts the same hour to about $65.

Maintenance and model retirements

AI tools add a maintenance cost of their own: the model underneath gets retired. Anthropic gives at least 60 days' notice and is retiring Claude Sonnet 4.5, whose model ID is dated September 29, 2025, on November 30, 2026. Each change of model means re-running your acceptance tests and sometimes rewriting prompts, so budget reviewer time for it at least once a year, along with the usual fixes and updates.

Human review

Review is the largest running cost, and the one that makes the tool safe to use. In our example an analyst spends 15 minutes checking each response's values against the cited pages: 100 hours a year, or about $5,600.

Model fees are under 3% of that total. So put your negotiating effort into review time and volume: a model that costs twice as much but halves the review saves money.

Time saved counts as a saving only if something changes, such as a contractor you stop paying or a deadline you now meet. Our custom software budget worksheet keeps staff capacity separate from cash for that reason.

Who this is for: budget owners who must defend a recurring line once the build is paid for, and the finance teams asked to approve it.

What do off-the-shelf AI assistants cost?

Standard seats that include an AI assistant are published at $7 to $25 a user a month, less for eligible nonprofits. For a small team that reads and drafts occasionally, a seat is the right first purchase, at a fraction of the cost of any build. We list only prices the vendor publishes on its own page:

• Claude Team lists standard seats at $20 a month billed annually or $25 monthly for teams of 2 to 150, with premium seats at $100 or $125, and states no model training on your content by default. Claude Enterprise lists $20 a seat a month, billed annually, plus usage at API rates, with a 20-seat minimum.
• Claude for Nonprofits lists Team at $8 a user a month for registered 501(c)(3) organizations and international equivalents, and $3 for nonprofits in low- and middle-income countries.
• Google Workspace lists Business Standard, which includes the Gemini assistant in Gmail, Docs and Meet, at $14 a user a month, and Business Starter, with Gemini in Gmail, at $7.

Five people on Claude Team standard seats cost $1,200 a year billed annually, and the same five get Gemini inside an $840-a-year Business Standard suite. If your organization already licenses another assistant, such as ChatGPT Business or Microsoft 365 Copilot, run the same arithmetic with the price on the vendor's own page. For nonprofits and foundations with lean teams, an $8 seat makes the case for building even harder to win.

What a seat does not give you is a workflow: a queue where each extracted value waits for a named reviewer, a record of who approved what, and a feed into your reporting system. When you need those, you are back to building, or to a platform that already does it. Our guide to building or buying sustainability software covers that choice.

Who this is for: teams of up to a few dozen people whose AI use is reading, summarizing and drafting inside documents they already have.

What governance work should you budget time for?

Governance costs hours from your reviewer, your data owner and whoever accepts the risk. Plan for four kinds of work: choosing risk checks, settling data terms, naming reviewers and testing accuracy.

NIST's AI Risk Management Framework

The NIST AI RMF is intended for voluntary use, was released on January 26, 2023, and is being revised as part of the White House AI Action Plan. Its Generative AI Profile, published in July 2024, names confabulation, confidently stated but false content, among the risks unique to or worsened by generative AI. Use it to choose the checks that fit one tool, and do not describe the result as a NIST certification.

OMB's AI memos, for public bodies

Three OMB memos govern federal AI use today, and none of the later memos on OMB's list replaces them:

• M-25-21 (April 3, 2025) requires seven minimum practices for high-impact AI, meaning AI whose output is a principal basis for decisions with a legal, material, binding or significant effect on rights or safety. The practices include pre-deployment testing and an impact assessment covering direct costs and expected savings, checked by an independent reviewer. Agencies had 365 days to document them and must safely stop using high-impact AI that does not comply.
• M-25-22 (April 3, 2025) covers buying AI, for solicitations issued 180 days or more after it. It tells agencies to make contracts permanently bar vendors from using non-public agency data and outputs to train publicly or commercially available AI without the agency's consent, and to scope licensing to avoid vendor lock-in.
• M-26-04 (December 11, 2025) requires contract terms on two "Unbiased AI Principles," truth-seeking and ideological neutrality, for any large language model an agency procures after that date.

These memos bind federal agencies, not cities, counties or nonprofits. They still make a free template: M-25-21's practices are a sound checklist for any public body's AI use, and M-25-22's data clause is worth copying into any AI contract. An assistant that extracts project data for staff to approve probably falls outside M-25-21's definition of high-impact AI. Resilience teams should note what falls inside: the memo presumes high impact when AI is the principal basis for decisions on public infrastructure that would pose a significant risk to safety if it failed.

Data privacy

Know where your documents go before anyone uploads one. Claude's Team plan states no model training on your content by default, while Google's API pricing page says content sent on its free tier is used to improve Google's products and paid-tier content is not, which rules the free tier out for supplier contracts or resident data. Public bodies also have records and retention rules to square with a vendor's terms; our guide to smart community data privacy lists the questions to ask.

Human review

Name the reviewer before anyone writes code. CF Labs builds to the Montréal Declaration, and under its responsibility principle, as we apply it, a person on your team accepts or rejects every output and the record shows who did. That reviewer's hours belong in the budget, as they do in the worked example above.

Accuracy testing

Grounding a model in your own documents lowers its error rate, and only testing tells you by how much. Legal research vendors had promoted retrieval-augmented generation as eliminating or avoiding hallucinations, yet in a 2025 peer-reviewed study their tools each hallucinated more than 17% of the time. Google's AI playbook for sustainability reporting advises starting with a prototype and testing its outputs against human-verified data. Build a test set from real files with known answers, including a superseded version and a record with a missing year, and agree on the pass mark before the demo; our pilot acceptance checklist sets out the tests.

Who this is for: resilience and sustainability staff in public agencies, and any team whose AI output reaches a board, a funder or the public. Council Fire works with municipalities and government agencies on the programs these tools serve.

Where does AI not pay off for a sustainability team?

AI does not pay off where the hours are few, the answer must be exact or nobody will check the output. Price the build in hours first. At about $56 an hour, the $25,000 low end of CF Labs' AI range equals about 440 hours of a specialist's loaded time, and the worked example's $37,000 to $99,000 equals about 650 to 1,750 hours.

Suppose the team in our example spends two hours on each response by hand today; replace that with your own figure. The tool would turn 800 hours of work into 160 hours of review and monitoring, and the 640 hours saved are worth about $36,000 a year at the loaded rate. The build would pay back in roughly one to three years, if the saved hours go to work that matters.

Five places where it rarely pays:

• Standard-bound arithmetic. Emissions math belongs in formulas you can audit, under a method such as the GHG Protocol. Google's reporting playbook notes that a spreadsheet formula or simple script is sometimes faster and more reliable, and keeps AI for ambiguous tasks that rules cannot handle.
• Once-a-year tasks. A report section that takes a week, about 40 hours, would need more than a decade to repay a $25,000 build even if the tool did all of it.
• Applications a funder expects you to write. NIH stopped considering applications substantially developed by AI from its September 25, 2025 receipt date. Before automating grant reporting, audit which questions you can stop asking.
• Output nobody will check. Our Radar puts unreviewed AI drafts in plans, grants and disclosures on hold, and a public claim needs a person to check it against its evidence before it ships; our guide to substantiating environmental claims covers what counts.
• Data nobody owns. If no one can say which file is current, fix ownership first; the ESG Data Management Guide covers owners, sources and controls.

The footprint is real but small per query. Google reported in August 2025 that a median Gemini Apps text prompt used 0.24 watt-hours of energy, 0.03 grams of CO2e and 0.26 milliliters of water, after a 33-fold drop in energy per prompt over 12 months. A 20-page document is far larger than a median chat prompt, so ask your provider for usage-based figures. CF Labs tracks compute footprint the way it tracks cost, and our piece on balancing AI benefits with energy use covers the wider tradeoffs.

Who this is for: whoever writes the business case, and the finance colleague who will test it.

Which option fits your team?

Start at the cheapest row that does the job, and move down only when that row's over-buying risk no longer applies.

Who this is for: anyone choosing a first AI purchase for a sustainability, ESG, resilience or program team.

Related Resources

Budgeting and buying software

• How to write an RFP for AI tools in a sustainability program: accuracy tests, human review and the rules public buyers cite.
• What drives the cost of custom sustainability software?: a budget worksheet for every recurring line.
• Before you commission sustainability software: write a one-workflow build brief: define the task before anyone prices it.
• Build or buy sustainability software? When each one wins: where a platform beats a custom tool.
• Build or buy climate resilience planning tools?: the same decision under public procurement rules.
• How to choose sustainability software: buy, configure or build?: the decision matrix to start from.
• CF Labs Project Estimator: a planning range and timeline from seven questions.

Using AI with care

• AI sustainability reporting: a pilot acceptance checklist: the tests a pilot should pass before rollout.
• How to substantiate environmental claims AI can cite: keep every claim tied to evidence.
• Balancing AI benefits with energy use: the footprint side of the ledger.
• Water Benchmarking for AI Data Centers: how data centers' water use is measured.
• CF Labs Radar: our quarterly ratings of climate data, tools and AI methods.

The data your tool depends on

• ESG Data Management Guide: owners, sources and controls behind any tool.
• ESG Data Collection: Step-by-Step Guide: gathering the records an assistant will read.
• CDP Disclosure Guide: Climate, Water, and Forests: where supplier data often ends up.
• Smart Community Data Privacy: Key Questions: what public bodies should ask before sharing data.
• Sustainability data and technology guides: the library's data and tooling guides.

Start with one task and the documents behind it

Bring the task that eats your team's week and the documents behind it, whether supplier responses, hazard mitigation plans or grantee reports. If you can already name the task, the data and the reviewer, a Tool Sprint will put a working prototype on your own documents in two to three weeks for $12,500. If you cannot, start with the Data & AI Readiness Review: $6,500 and two weeks for an inventory of your data and workflows, a shortlist of AI uses checked for risk against the Montréal Declaration and a 90-day roadmap with costs. To talk it through first, use our contact page and choose Software & AI (CF Labs).

About the Author

George Chmael II is Council Fire's Founder & CEO. He has led the firm since its founding and through its B Corp certification in 2010. Council Fire's sustainable business practice helps companies, NGOs and foundations build sustainability programs that hold up, and CF Labs builds the software and AI that put them to work.

Written by

George Chmael II

Founder & CEO

Corporate lawyer turned social entrepreneur, guiding organizations toward three bottom lines for 25+ years.

Full bio →All 85 articles by George →

From the library

• Topic hubSustainability Data & Technology→
• GuideESG Data Management Guide→
• CF LabsThe tools we build→

Around The Fire, in your inbox.

New articles on climate resilience, sustainable business and the end of extraction, sent when they publish.

Email addressSubscribe