Every AI agent pitch leads with a seat price. Nineteen dollars a user. Forty dollars a user. It looks like a phone plan, and that is exactly the problem — because the seat price is the sticker, not the bill. The three things that actually determine whether an AI agent is a safe purchase or an expensive mistake almost never appear on the pricing page: who is liable if the agent's output infringes someone's copyright, where your data physically lives, and what the usage-based bill really comes to once the tool is in daily use.
This started as a developer-tooling debate — which AI coding assistant to standardize on across a large engineering org — but the three questions are not about code. They are the questions every Fort Wayne business should run before signing any AI agent or AI Employee contract, whether it writes software, answers phones, or reads client documents. Below is the checklist, built on the clearest public analyses available: a 500-seat comparison from beri.net and an enterprise copilot pricing breakdown from costbench, plus primary vendor terms. (A MarkTechPost roundup framed the same three-way comparison; it was not directly accessible for this piece, so every figure here is anchored to a source we could verify, and all prices should be confirmed against the vendor's current sheet before you sign.)
Key Takeaways
- Seat price is the sticker, not the bill. Realized all-in cost runs far higher than the per-seat number once usage-based billing kicks in — plan for the usage line, not the seat line.
- IP indemnity is not binary. Some vendors cover both training data and output; at least one caps liability at 12 months of fees, which is closer to a refund than true indemnity.
- “We don't train on your data” has fine print. It is often true for business/enterprise tiers and false for personal tiers — and employees on personal subscriptions can leak proprietary data into training.
- Data residency is the least-documented term and the most important for regulated work. If you handle HIPAA or confidential client files, where the data lives is a contract question, not a feature.
- Run the 5-question procurement checklist before you sign any AI agent contract, and weight indemnity and residency heavily if you are a regulated Allen County buyer.
- A managed gateway changes the math by putting residency, cost controls, and oversight under your control instead of the vendor's.
The uncomfortable truth underneath all three questions: the productivity case for these tools is real but noisier than the marketing suggests. beri.net's roundup cites a METR study in which experienced developers were actually 19% slower using AI tools while believing they were 20% faster, and GitClear's analysis of hundreds of millions of code changes found duplicated code blocks up 81% and copy-paste up 41%. Neither finding means the tools are bad — it means you are buying something whose value you have to verify, which makes getting the contract terms right even more important. You are not just buying speed; you are taking on liability, data exposure, and a variable bill.
Question 1: Does the vendor indemnify you for the agent's output?
When an AI agent produces something — code, copy, a contract summary — and that output turns out to infringe someone's copyright, who pays to defend you? That is IP indemnity, and the answer varies far more than the uniform “yes” on comparison tables suggests.
The strongest coverage indemnifies both the training data and the generated output. According to beri.net's analysis, Google's Gemini Code Assist is structured this way and is the broadest of the major options, and Microsoft's Customer Copyright Commitment covers GitHub Copilot's business and enterprise use — as of April 2026, Microsoft stated there are “no additional required mitigations” required for coverage. Anthropic takes a similar posture for paid API customers: its service-specific terms describe a defense obligation for authorized use of the services and the outputs generated from them, with customers retaining their inputs and owning their outputs.
Then there is the fine print that turns a “yes” into a maybe. beri.net flags that Cursor's indemnity caps total liability at the fees paid in the prior 12 months — which, at 500 seats, is roughly $240,000. The analysis calls that “a refund” rather than a true indemnity, and the distinction matters enormously: a capped indemnity protects the vendor's exposure, not yours. For several other tools, the comparison simply marks “indemnified” with no terms disclosed, which is not something you can rely on in a contract.
What to verify before signing: Does the indemnity cover output and training data? Is liability capped, and at what number? Is it automatic, or conditional on you using specific filters or settings? Get the actual contract language — the marketing checkmark is not the agreement. This is the same discipline we apply to the broader AI governance gap in software cost oversight: a control you cannot point to in writing is a control you do not have.
Question 2: Where does your data live, and can you keep it in-region?
Data residency is, by beri.net's own admission, the least-documented term across the entire category — the roundup could confirm little beyond a note that one vendor's data-residency controls are restricted to its enterprise tier. For most AI agent products, where your prompts, documents, and outputs are processed and stored is buried in a data-processing addendum, not advertised as a feature. For a lot of businesses that is a shrug. For a regulated one, it is the whole ballgame.
Tangled up with residency is a related question: does the vendor train its models on your data? Here the pattern is clearer, and it is a trap for the unwary. beri.net reports that business and enterprise tiers from the major vendors generally do not train on customer data by default — but personal tiers often do. GitHub, for example, uses Free, Pro, and Pro+ interaction data for training by default (business and enterprise are excluded), and Anthropic's service-specific terms carve out a “Development Partner Mode” that is an explicit opt-in exception where submitted data may be used, including to train models. The real exposure is not the enterprise contract you negotiate carefully; it is the engineer or staffer expensing a personal subscription and running company work through it. beri.net's recommendation is blunt and worth copying: audit expense reports for personal AI subscriptions.
For data residency specifically, the dependable answer is architectural. If you need prompts and documents to stay in a region — or off the public cloud entirely — you either buy a tier that contractually guarantees it or you route the agent through infrastructure you control. That second path is what a Secure AI Gateway is for: it sits between your people and the model vendors so residency, logging, and access are enforced on your side rather than trusted to each vendor's defaults. We walked through the regulated version of this in our guide to hybrid AI for confidential data off-cloud in Fort Wayne and the observability angle in keeping AI observability in your own cloud.
What to verify before signing: Where is data processed and stored, and is in-region or off-cloud an option? Does the vendor train on your data on this specific tier? Can administrators block employees from using unmanaged personal accounts for company work? Is there a signed data-processing addendum, not just a trust-center web page?
Question 3: What will the real bill be, not the sticker?
This is where the seat price does the most damage, because the industry moved the goalposts in 2026. On June 1, 2026, GitHub moved all Copilot plans to usage-based billing. Premium request counts were replaced by GitHub AI Credits consumed by token usage — input, output, and cached tokens — at each model's published API rates, as the GitHub billing documentation spells out. Base seat prices stayed familiar (Copilot Business at $19/user/month with $19 of credits, Enterprise at $39 with $39), but once a team uses the premium models and agent runs that make these tools worth buying, the credits are gone and the meter is running.
How big is the gap between sticker and reality? beri.net cites a DX estimate that the true all-in cost — seat plus tokens — lands at roughly $200 to $600 per engineer per month. Against a $19–$39 seat price, that is the real story: usage, not the seat, is the bill. The roundup also notes individual developer bills that reportedly jumped from $29 to $750, and from $50 to $3,000, per month after the June change. costbench, surveying the broader category, puts the median seat price around $54 per month but explicitly flags the hidden costs that do not appear in that number: premium-model and agent usage metered beyond the base seat, seats paid for by occasional users, and integration and rollout effort.
The table below pulls together what the public analyses report. Treat every figure as dated and directional — prices across these aggregators are inconsistent, and usage-based billing means the “cost” column is a floor, not a forecast. Confirm each against the current vendor sheet before you buy.
| Option | Seat price (reported) | Billing model | IP indemnity (reported) | Trains on business data |
|---|---|---|---|---|
| GitHub Copilot Business | ~$19/user/mo incl. $19 credits | Usage metered at API rates beyond credits | Customer Copyright Commitment | No (Business/Enterprise) |
| GitHub Copilot Enterprise | ~$39/user/mo (+ platform fees) | Usage metered at API rates | Customer Copyright Commitment | No |
| Cursor Teams | ~$40/user/mo | Overages billed in arrears | Yes, but liability capped at 12 mo. of fees | No, unless customer agrees |
| Claude Enterprise | ~$20/seat + usage | Usage at API rates | Yes (terms vary; verify) | No by default |
| Gemini Code Assist | ~$19/user/mo (Enterprise ~$45) | Hard daily quota, no overage | Covers training data and output | No |
Figures compiled from beri.net (as of Aug 10, 2026) and costbench; confirm against current vendor pricing before purchase.
The planning lesson is the one we make every time cost hides behind a per-unit sticker, most directly in our breakdown of the document AI cost-per-page decision: buy on realized all-in cost, not the advertised rate. A predictable tool with a hard quota and no overage can be cheaper to run a business on than a cheaper-looking tool whose meter never stops.
The 5-question AI agent procurement checklist
Run this before signing any AI agent or AI Employee contract. It takes an afternoon and routinely saves a year of regret.
- Indemnity: Does the vendor indemnify you for the agent's output and its training data, and is that liability capped? Get the clause, not the checkmark.
- Training: Does the vendor train on your data on the specific tier you are buying — and can you stop employees from using unmanaged personal accounts for company work?
- Residency: Where is your data processed and stored, can you keep it in-region or off-cloud, and is there a signed data-processing addendum?
- Real cost: What is the realized all-in cost per user per month at your actual usage — not the seat price — and can you cap spend with admin budgets?
- Control: Can you enforce logging, access limits, and spend caps on your side, or are you trusting each vendor's defaults?
If a vendor cannot answer all five in writing, that is your answer. The checklist is also a useful internal document: it forces the person who wants the tool and the person who owns the risk to sit at the same table before money is committed, which is exactly the oversight most AI purchases skip.
For Allen County's regulated buyers: weight residency and indemnity first
If you run a legal practice, a healthcare clinic, or a manufacturer handling confidential designs in Fort Wayne or anywhere in Allen County, the five questions are not equally weighted. Residency and indemnity move to the front, because your obligations do not bend to a vendor's default settings.
A law firm bound by client confidentiality cannot route privileged documents through a tool whose data-processing terms it has not read. A clinic subject to HIPAA cannot treat “the vendor says they don't train on our data” as compliance — it needs a Business Associate Agreement, a documented residency answer, and the ability to prove where protected health information went. A manufacturer with proprietary CAD files or customer contracts has trade-secret exposure the moment that data crosses into an unmanaged model. For every one of these buyers, the worst-case scenario is not an oversized bill; it is a confidentiality breach, a regulatory finding, or an IP claim with a capped indemnity that does not cover the damage.
This is the design problem Cloud Radix's Secure AI Gateway and managed AI Employees were built to answer. Instead of accepting each vendor's residency and logging defaults, the gateway puts those controls on your side: regulated files can stay off the public cloud, every interaction is logged for audit, and access is governed per user and per data type. The AI Employee does the work; the gateway keeps the data where your obligations require it to stay. For a regulated Northeast Indiana buyer, that is the difference between adopting AI and adopting a liability.
Buy the terms, not just the tool
The AI agent market is optimized to sell you on capability and quote you a seat price. The three questions that actually protect your business — who is liable, where your data lives, and what the real bill is — take deliberate effort to answer, and most of them are not on the pricing page. That effort is the entire job of good procurement.
Cloud Radix helps Fort Wayne and Northeast Indiana businesses run exactly this diligence — and, where the answers do not add up, deploy managed AI Employees through a Secure AI Gateway that keeps residency, cost, and oversight under your control instead of a vendor's. If you are weighing an AI agent purchase and want a second set of eyes on the indemnity, residency, and real-cost terms before you sign, our AI consulting service will run the checklist with you — get in touch and bring the vendor's contract.
Frequently Asked Questions
Q1.What is IP indemnity for an AI agent, and why does it matter?
IP indemnity is the vendor's promise to defend and cover you if the agent's output infringes someone's intellectual property. It matters because without it, a copyright claim over AI-generated code or content is your problem to litigate. The strongest coverage indemnifies both the model's training data and its output; weaker versions cap liability at the fees you have paid, which functions more like a refund than real protection.
Q2.Do AI agent vendors train their models on my data?
It depends on the tier. Business and enterprise tiers from the major vendors generally do not train on customer data by default, but personal and free tiers often do — GitHub, for instance, uses Free, Pro, and Pro+ interaction data for training by default. The common hidden risk is employees running company work through personal subscriptions, so auditing expense reports for unmanaged AI accounts is a practical safeguard.
Q3.What is data residency and when does it matter for a business?
Data residency is where your data is physically processed and stored. It matters most for regulated businesses — HIPAA-covered clinics, law firms under confidentiality duties, manufacturers with trade secrets — where keeping data in a specific region or off the public cloud is a legal obligation, not a preference. It is also the least-documented term in most AI agent contracts, so it has to be verified in a signed data-processing addendum rather than assumed.
Q4.Why is the seat price not the real cost of an AI agent?
Because most AI agents moved to usage-based billing in 2026, the per-seat price only covers a base allotment. Once a team uses the premium models and agent features that justify the purchase, usage is metered on top. Public estimates put realized all-in cost far above the seat price — on the order of several hundred dollars per user per month in heavy use — so you should budget and cap against actual usage, not the sticker.
Q5.What should a regulated Fort Wayne business prioritize when buying an AI agent?
Residency and indemnity first. A regulated buyer should confirm where data lives, secure a data-processing addendum (and a Business Associate Agreement where HIPAA applies), and get the indemnity terms in writing before weighing cost. Routing AI through a secure gateway you control, rather than relying on each vendor's defaults, lets you keep confidential files in-region or off-cloud and prove it in an audit.
Q6.Can a small business run this procurement checklist without a legal team?
Yes. The five questions — indemnity, training, residency, real cost, and control — are plain-language diligence that any owner or operations lead can ask, and requiring written answers is itself most of the protection. For contract language or regulated-industry obligations, a consulting partner can run the checklist with you and flag the terms that need a lawyer's eyes, which is far cheaper than discovering a gap after you have signed.
Sources & Further Reading
- beri.net: beri.net/article/best-ai-coding-assistants-500-engineer-organizations-2026 — The Best AI Coding Assistant at 500 Seats: Pricing, Indemnity, Training, and Security
- costbench: costbench.com/software/developer-copilots-enterprise — Enterprise AI Developer Copilots Pricing 2026
- GitHub: github.blog/news-insights/company-news/github-copilot-is-moving-to-usage-based-billing — GitHub Copilot is moving to usage-based billing
- GitHub Docs: docs.github.com/en/billing/concepts/product-billing/github-copilot-premium-requests — GitHub Copilot premium requests
- Anthropic: anthropic.com/legal/service-specific-terms — Service-Specific Terms
- MarkTechPost: marktechpost.com/2026/09/26/ai-coding-agents-for-enterprise-ip-indemnity-data-residency-and-500-seat-cost-compared — AI Coding Agents for Enterprise: IP Indemnity, Data Residency, and $500-Seat Cost Compared
Weighing an AI Agent Purchase?
Bring the vendor's contract. We will run the indemnity, residency, and real-cost checklist with you — and, where the terms do not add up, deploy a managed AI Employee through a Secure AI Gateway that keeps control on your side.
