Enterprise AI Coding Agents: IP Indemnity, Data Residency and 500-Seat Costs Compared

Only two of the four major AI coding agent vendors stand behind generated code with uncapped IP indemnity in their standard enterprise contracts, and the cheapest 500-seat deployment with both indemnity and SSO costs $120,000 per year, according to a MarkTechPost contract review published September 26, 2026.

The review read the current contract language for GitHub Copilot, AWS Kiro, Cursor, Devin and Windsurf, checking every term against the vendors' own pages on September 26, 2026. It targets a specific reader: the procurement lead, the general counsel and the security reviewer. Those buyers ask four questions before any rollout. Who pays if generated code triggers an IP claim? Where do prompts live? What can admins log? What do 500 seats actually cost?

The stakes are straightforward. Enterprises are deploying AI coding agents at scale, and the gap between vendors on legal exposure, data handling and true pricing is wide enough to shift six-figure budgets. The review is reporting, not legal advice; final terms should go past counsel.

Five brands are really four contracts

Cognition acquired Windsurf and, on June 2, 2026, renamed the Windsurf editor Devin Desktop. Devin and Windsurf now sit on one pricing table and one set of terms. That leaves four contracting parties for five product names: GitHub Copilot, AWS Kiro, Cursor and Cognition.

The indemnity table nobody publishes

The indemnity clause is what blocks enterprise rollouts. The differences are sharp.

GitHub Copilot (Business, Enterprise) offers "uncapped IP indemnification for unmodified outputs," per the Copilot Trust Center, under the Copilot Product Specific Terms or Microsoft Product Terms. Coverage includes unmodified code from the Copilot coding agent. The Duplicate Detection filter is no longer required since April 3, 2026. Free, Pro, Pro+ and Max plans are excluded.

AWS Kiro (Pro, Pro+, Pro Max, Power) states plainly: "Kiro offers indemnity for its output." Coverage applies to copyright claims on output under the AWS Customer Agreement and Service Terms, and it is uncapped for copyright claims per the AWS Bedrock FAQ. Conditions: no infringing inputs, no disabled service filtering features, and no coverage on the free tier.

Cursor (business customers) has the broadest language on paper. Its MSA, updated August 13, 2026, defends claims "that the Service or any Suggestion infringes or misappropriates" third-party IP. Suggestions are named explicitly, and indemnity is carved out of the 12-month fee cap. Exclusions apply if modified or combined code causes the claim, if filters are disabled, or if the customer knew infringement was likely. The consumer Terms of Service run the other way: users indemnify Anysphere.

Cognition (Devin and Devin Desktop) is the outlier. Its MSA, updated June 8, 2026, defends claims that the Services infringe patents, copyrights or trade secrets — but excludes Customer Data, defined as "including Outputs." Generated code therefore sits outside the standard promise. Indemnity is capped at 2x fees paid in the prior 12 months, and an executed order form controls over the MSA. That is where buyers should negotiate. On the Platform Terms for Pro, Max and Teams tiers (updated June 30, 2026), the cap is the greater of 6 months of fees or US$100, and indemnity "applies only to Customers on the Paid Service Tiers."

Legal exposure on generated code

Microsoft's April 3, 2026 change matters more than any feature launch for Copilot, the review argues. Before that date, one admin setting could void coverage; the required-mitigations page now lists no additional mitigations for GitHub offerings. The remaining limit is the word "unmodified" — most shipped code gets edited, so counsel should assess how far edits move code outside scope.

Kiro inherits AWS's generative AI indemnity, uncapped for copyright claims on output from services listed in Section 50.10, with paid subscribers shipping with the protection included.

Where prompts are stored

GitHub Copilot: For Business and Enterprise, the plans FAQ says IDE chat and completion prompts are not retained. Prompts from other surfaces — github.com, mobile and CLI — are retained for 28 days. User engagement data is kept for 2 years. GitHub does not train on Business or Enterprise data. Enterprises on GitHub Enterprise Cloud with data residency can pin Copilot inference to the United States or the European Union, at a cost of 10% more AI credit consumption, with a narrower model list per region.

AWS Kiro: Enterprise content is never used for service improvement, per Kiro's data protection page. Data is stored in the region where the Kiro profile is configured; inference stays inside the US or Europe geography, except for models tagged experimental. Admins can encrypt data with customer-managed KMS keys.

Cursor: With Privacy Mode on, Cursor holds zero data retention agreements with all model providers and does not train on customer code. File contents are cached temporarily, encrypted with client-generated keys. Cursor does not publicly offer a customer-selectable processing region, and every request still passes through Cursor's backend.

Cognition: The MSA bars training on Customer Data without written consent. The Customer Dedicated Deployment runs Devin's devbox in a single-tenant VPC linked by AWS PrivateLink — but the agent's reasoning layer still runs in Cognition's cloud. On self-serve paid tiers, Cognition may train on Customer Data until the customer opts out; on Teams, only an admin can opt out.

Audit logging and admin controls

GitHub Copilot: The enterprise audit log records Copilot events under action:copilot. Agent activity appears under actor:Copilot for 180 days. GitHub is explicit that prompts sent locally are not in the audit log. Admins set model, preview-feature and spending policies.

AWS Kiro: SSO runs through IAM Identity Center, Okta or Entra ID. Admins can enable prompt logging and daily user activity reports, both landing in an S3 bucket inside the customer's AWS account — prompt records stay under customer control.

Cursor: Teams includes SAML or OIDC SSO, team-wide Privacy Mode and usage analytics. Audit logs, SCIM, repository and model access controls, and an AI code tracking API require Enterprise.

Cognition: Teams offers an admin dashboard with analytics. SAML or OIDC SSO, centralized admin controls and VPC deployment are Enterprise only, and the Enterprise API exposes audit log endpoints.

True per-seat cost at 500 seats

List prices are monthly, in USD, before usage overages and taxes. The "true" figure is the cheapest tier delivering indemnity plus SSO.

• GitHub Copilot Business: $19 per seat — $9,500 per month, $114,000 per year, with 1,900 AI credits per seat pooled. SAML SSO and data residency require GitHub Enterprise Cloud, bringing the real combined cost to $40 per seat ($240,000 per year) if you don't already pay for it. Copilot Enterprise plus Enterprise Cloud runs $60 per seat ($360,000 per year) with 3,900 AI credits per seat.
• AWS Kiro Pro: $20 per seat — $10,000 per month, $120,000 per year, with 1,000 credits per user. Add-on credits cost $0.04 each. Self-serve tiers cover teams up to 500.
• AWS Kiro Pro+: $40 per seat — $240,000 per year, 2,000 credits per user, same indemnity as Pro.
• Cursor Teams: $40 per seat — $240,000 per year — but lacks audit logs and SCIM; Enterprise is custom priced.
• Cognition Teams (Devin, Devin Desktop): $40 plus an $80 team fee, with daily and weekly quotas. Teams is capped at 200 users, so 500 seats means Enterprise at custom pricing.

Two meters sit on top of every seat. GitHub moved Copilot to usage-based AI credits on June 1, 2026, where 1 credit equals $0.01. Kiro bills extra credits at $0.04, and enterprise overages are off by default. Agent-heavy teams should model usage, not just seats.

What to put in the RFP

The review closes with concrete procurement steps: ask each vendor to confirm in writing whether edited or agent-written code stays inside indemnity scope; for Cognition, redline the Customer Data exclusion so Outputs are covered and lift the 2x supercap; for Copilot, confirm whether you buy under GitHub terms or Microsoft Product Terms; require region commitments in the order form, not just a settings page; and price 12 months of agent usage at your real model mix, not the seat line alone.

The bottom line for buyers: Copilot offers uncapped indemnity for unmodified output with no filter requirement since April 3, 2026; Kiro is the cheapest 500-seat path with indemnity, SSO and customer-owned prompt logs; Cursor's MSA names Suggestions and removes indemnity from the fee cap; and Cognition's standard terms exclude Outputs from indemnity, covering Devin and Windsurf alike. At 500 seats, Cursor and Cognition both require custom Enterprise pricing — which means procurement teams that skip the order-form negotiation may find the contract, not the model, is what they actually bought.