Liability allocation
When it fails, who pays?
Every failure comes with a bill, and laws, network rules and contracts decide who gets it. Liability shapes pricing, risk appetite and which features a company is willing to build.
What it is
Liability allocation is the set of rules that decides who absorbs the cost when something goes wrong: the customer, the platform, a network or an insurer. It is set by law, by network rules and by contracts, and it often shifts depending on what each party did or didn't do.
Follow the liability and you'll understand why a company says yes to some features and no to others.
Where you'll see it
| Industry | Who pays when it fails |
|---|---|
| Payments | Fraud liability shifts to the card issuer when a payment was properly authenticated |
| Telco | Fines for non-compliant messages; who pays for fraudulent traffic |
| Healthcare | Who absorbs a denied claim or a billing error |
| Payroll | Penalties for late or wrong tax filings |
| SaaS | Service credits when uptime targets are missed |
Questions to ask in week one
- For each major failure, who pays today?
- What would shift that cost to someone else?
- Where do our contracts cap liability, or pass it through?
- Which failures are insured, and which come straight out of margin?
- What has the company paid for before, and what changed afterwards?
The trap
Treating liability as a legal footnote. Liability sets the decision culture: the more a failure costs, the slower and more careful every team around it becomes. Know why before you try to speed them up.
In the Field Guides
AI agents: orchestration platforms
Courts have said a bot isn't responsible for itself; the open questions are the vendor's share and an agent's rights on other people's websites.
| Failure | Who absorbs it | Mechanism |
|---|---|---|
| Agent misstates a policy to a customer | The deployer | Misrepresentation (Moffatt v. Air Canada, 2024) |
| Vendor's AI screens out job applicants | Possibly the vendor too, as the employer's agent | Mobley v. Workday (collective certified 2025) |
| Agent shops where the site forbids agents | Unsettled | Amazon v. Perplexity |
| Injected or unauthorized action | The customer | The clouds' responsibility splits |
| Runaway token bill | The customer on consumption; the vendor on outcome pricing | Contract |
| Copyright claim on an output | The lab or cloud, with exclusions | IP indemnities (Microsoft, Google, OpenAI, Anthropic) |
| Damage from an agent's action | No indemnity found that covers it | Contract silence |
| Defective AI software sold in the EU | The producer, without proof of fault | Product Liability Directive, from 2026-12-09 |
Air Canada argued its chatbot was a separate legal entity. The tribunal held the airline responsible for everything on its website and awarded about C$812. Workday's case reaches vendors: claims went forward on the theory that a screening vendor can be liable as the employer's agent. Perplexity's case turned on whose permission an agent carries. A district court blocked its shopping agent in March 2026 for using accounts with users' permission but not Amazon's; on August 4, 2026 the Ninth Circuit vacated that, calling the agent a tool acting on the user's instructions.
Contracts push the act to the customer. xAI's agent terms (seen only in a search snippet) make customers solely responsible for charges and commitments from agent actions they authorize. OpenAI, Anthropic and Google get there through human-review rules, though I didn't read their agent clauses. Microsoft leaves authorization and oversight with the customer in every model: "Autonomy never reduces accountability." Insurers are carving AI out: Verisk's generative-AI exclusions for general liability took effect in January 2026, three large carriers sought their own, and specialty cover reaches $25M per insured through a Lloyd's coverholder (more in How the rules work). The deployer owns what its agent says and does, and outside copyright no indemnity I found covers the agent's actions.
AI coding agents: writing, reviewing and shipping code
| Failure | Who absorbs it | Mechanism |
|---|---|---|
| Copyright claim on an unmodified suggestion, on a paid plan with the filter on | The vendor | IP indemnity, usually outside the liability cap |
| Copyright claim on modified or combined code, or a patent claim | The customer | Indemnity exclusions |
| Agent deletes data or breaks production | The customer | Caps of about 12 months of fees; lost data and consequential damages excluded |
| Secret leaked from an agent's commit | The customer | No vendor term found |
| Runaway usage bill | The customer on usage pricing | Overage on by default unless an admin turns it off |
| Model withdrawn mid-contract | The app vendor, then its customers | Change-of-control clauses; no indemnity found |
Vendor terms push the operational risk down. Cursor's individual terms make the user "solely responsible" for suggestions it runs automatically and cap its liability at the greater of six months' fees or $100; its enterprise agreement, Anthropic's commercial terms and AWS's customer agreement cap at 12 months of fees, with IP indemnities outside the cap. GitHub's indemnity covers unmodified suggestions with the public-code filter on, and almost no shipped code is unmodified, so it covers the snippet more than the product.
When an agent did real damage, the platforms paid in goodwill: Replit promised a refund and added a default split between development and production databases. Insurers are moving the other way, with generative-AI exclusions entering general liability forms in 2026 (more in AI agents: orchestration platforms).
AI evals and observability: how teams measure AI quality
| Failure | Who absorbs it first | How |
|---|---|---|
| The agent states a wrong policy | The company running it | Air Canada paid C$812.02 after a tribunal held it responsible for its chatbot |
| The provider's model or serving stack degrades | The builder, toward its own customers | No quality SLA; lab caps of about 12 months of fees; accuracy disclaimed |
| The agent takes a harmful action | The customer, then the agent vendor | Contract and goodwill (Replit) |
| A quality claim turns out to be false | The vendor making the claim | Texas v. Pieces; SEC "AI washing" penalties |
The deployer owns what its AI says and does. Providers disclaim accuracy, cap liability at roughly a year of fees and write SLAs for uptime (Bedrock's credits apply when a region is down), so when a lab's change causes the regression, the builder still faces its customers. Inside the company, the PM who owns the quality bar owns the gate decision, so I'd record each sign-off with the report behind it.
AI governance and model risk
| Failure | Who absorbs it | Mechanism |
|---|---|---|
| Biased screening or scoring of people | The deployer by default; increasingly the vendor that scores | Anti-discrimination law; California's "agent" rule; Connecticut's no-defense rule; Workday and SafeRent |
| Automated denial without real human review | Patients first, then the insurer in litigation | Contract and good-faith claims; state laws requiring physician review |
| Chatbot misstatement | The company running it | Air Canada |
| Misleading claims about an AI product | The vendor | FTC and state attorneys general (Pieces, accessiBe, DoNotPay) |
| Vendor model changed or retired | The deployer | Notice periods, no compensation |
| Discrimination claim tied to a vendor's tool | Split by contract | Vendor caps of about 12 months of fees; whether discrimination is carved out isn't public |
| AI loss outside insurance cover | The insured company | Generative-AI exclusions in liability forms since January 2026 |
The law reaches whoever makes or shapes the decision about a person: the deployer by default, and more and more the application vendor whose tool scores, ranks or screens. Model labs sit behind usage policies and caps of about a year of fees, and I found no case where a lab was held liable for a deployer's decision about a person. So the evidence splits: the deployer keeps the inventory, oversight design, notices, outcome monitoring and records; the vendor keeps testing, documentation and intended-use limits; contracts decide who pays between them. Vendors' certificates don't move this: AWS says customers aren't "automatically certified by association" with its ISO 42001 certificate.
Agent identity and delegated access: how AI agents act for people
| Failure | Who absorbs it first | How |
|---|---|---|
| A hijacked agent misuses a user's access | The user or their employer | The agent makers' terms make the user responsible for the agent's actions |
| An integration vendor's token store is stolen | That vendor's customers, in forensics and rotation | Lawsuits aimed at the platform; the vendor's cap limits the rest |
| An agent with write access destroys data | The customer | Terms exclude lost data; refunds are goodwill |
| An agent buys the wrong thing | The consumer or the merchant | Probably counts as authorized; no network rule yet |
| A spoofed agent scrapes a site | The site owner | Bot-defence costs |
In every case I found, the money stayed where the contract put it, on the user. Vendors paid in engineering and reputation: Replit separated development and production databases, and GitHub said it would shorten token lifetimes for package publishing.
B2B payments: cross-border payouts
Contracts move most losses to the customer; laws and sanctions move some back.
| Failure | Who absorbs it | Mechanism |
|---|---|---|
| Payout arrives short | The worker or supplier, unless the customer grosses up | The charge option (SHA, BEN) |
| Wrong but valid account (a typo) | The customer, by contract; on SPEI, NEFT and Pix, the sender | Credit on the account number |
| Wrong IBAN in the euro area | The payer if VoP warned them; the provider if it skipped VoP | Instant Payments Regulation |
| FX move between quote and conversion | The provider while the quote is valid; the customer after, and on returns | Quote terms |
| FX move between payroll invoice and pay day | The EOR if the invoice is fixed; the client if it's reconciled | Contract |
| Local partner fails or is designated | The provider; customers if not safeguarded | Safeguarding, contract |
| Sanctions breach | The US person processing, under strict liability | OFAC; contracts move cost, not liability |
| Contractor reclassified | The client; a contractor-of-record provider only up to its cap | Local labour law |
| Permanent establishment from a remote worker | The client, not the EOR | Tax treaties |
| No W-8BEN on file | The US payer: presumed US person, 1099, backup withholding | Treas. Reg. 1.6041-4 |
An architecture firm wires a $5,000 invoice to a Philippine supplier on SHA. An intermediary and the supplier's bank each take a fee, the supplier sees a short payment, and the firm ends up sending the difference.
Law puts employer duties on the EOR, contracts push them back to the client, and courts look through both: Brazilian courts can declare the client the employer under "primacy of reality" (substance over form), Philippine law makes the principal jointly liable, and in Germany unlicensed labour leasing makes the client the employer. A certified US PEO (professional employer organization) is the one legal shield: the IRS treats it as the only employer for wages it remits. And sanctions ignore contracts: in 2021 OFAC fined a public payout company about $1.4M for 2,220 apparent violations after weak screening.
B2B payments: order to cash
As a rule of thumb, on payments loss follows whoever started the payment: the payer usually eats a bad push, the seller a bad pull. On credit, the seller eats the loss unless it moved it beforehand, to an insurer, a factor, a guarantor or a secured position.
| Failure | Who absorbs it | Mechanism |
|---|---|---|
| Payer tricked into paying a fraudster (BEC) | The payer, if its bank followed the agreed security procedure | UCC 4A-202; Studco v. 1st Advantage (4th Cir., 2025) |
| The seller's invoice in that case | Still legally unpaid, unless contract says otherwise | In practice a negotiation |
| Unauthorized ACH debit from a business account | The seller | The originating bank warrants the authorization; the seller indemnifies it |
| Card-absent fraud | The seller, unless authenticated (3-D Secure) or covered by Visa's compelling-evidence rules | Network rules |
| Returned check, Canadian business PAD claim, surcharge violation | The seller | Provisional credit reversed; debited back; fines through the acquirer |
| Customer can't pay | The seller, unless insured or sold without recourse | Insurers cover up to 90% (EDC) or 90% to 95% (EXIM); the seller keeps the rest, losses above the limit and the waiting period |
| Invalid deduction | The seller, unless it proves otherwise in time | The payer already holds the cash |
| Preference claw-back | The seller returns the cash unless a defense applies | 11 U.S.C. 547 |
| Payer pays the seller after notice that the invoice was sold | The payer still owes the factor | UCC 9-406 |
Say a customer gets an email that looks like it's from your AR team, with "new bank details", and wires $48,000 there. If its bank followed the agreed security procedure, the customer bears the loss, and your invoice is still open, so your seller is left with a customer who believes it already paid. Nacha's 2026 fraud rules leave UCC 4A unchanged; the buyer's controls are in procure to pay.
B2B payments: procure to pay
In US business payments the sender usually carries the loss on a bad push, and almost everything AP sends is one.
| Failure | Who absorbs it | Mechanism |
|---|---|---|
| Buyer pays a fraudster after a fake bank change | The buyer | UCC 4A-202: if the bank followed the agreed security procedure, the order binds the customer. 4A-207: the receiving bank may rely on the account number |
| The real supplier's invoice | Still owed by the buyer | Pay again, or negotiate |
| Altered or counterfeit check | Banks, under UCC 3 and 4 warranties; the buyer if it declined Positive Pay or missed a cut-off | Deposit agreement |
| Duplicate payment | The buyer, until recovered | Supplier credit balance; recovery-audit fees |
| Late payment | The buyer | Contract fees or interest; Prompt Payment interest for US federal buyers; lost discounts |
| Missing or wrong TIN | The buyer as payer | 24% backup withholding duty; IRS notices |
| Invalid GST/HST number | The buyer | Input tax credit at risk |
| A payment hub fails while holding funds | Buyer and supplier, depending on safeguarding | State permissible-investment rules; RPAA safeguarding |
Positive Pay loss-shifting comes from bank agreements rather than statute. Courts back the rest up: in Studco v. 1st Advantage (4th Cir., 2025) the receiving bank wasn't liable for a name and account mismatch that no employee actually knew about. Nacha's 2026 rules leave UCC 4A unchanged, and the US has no push-fraud reimbursement scheme like the UK's. Recovery is thin: in AFP's 2025 survey, a fifth of victims got nothing back. What a hub's payment "guarantee" covers is a contract question.
Since the buyer pays for a bad push, the cheap protection sits before release: verify the change, hold payments to new details, and have two people sign off.
B2B payments: real-time payments and open banking
| Failure | Who absorbs it first | How |
|---|---|---|
| A US business pays a fraudster after a fake "new bank details" email | The payer | UCC 4A; the receiving bank may rely on the account number |
| A fake request for payment gets paid | The sender's bank, on a valid warranty claim | RTP and FedNow rules; a quality complaint isn't a breach |
| A UK consumer is tricked into paying | Sending and receiving banks, half each, up to £85,000 | Reimbursement rules |
| A UK 50-person company is tricked | The company | Outside the reimbursement scope |
| An EU payer sends to a wrong IBAN and the bank skipped the name check | The payer's bank | Instant Payments Regulation |
| An EU consumer pays a fraudster posing as their own bank | The consumer until around 2028, then the bank | PSD2, then the Payment Services Regulation |
In the US, Article 4A puts a business transfer on the customer if the bank followed a commercially reasonable security procedure the two agreed, in good faith. In business email fraud the customer's own staff send the payment, so it's theirs. The beneficiary bank can rely on the account number even when the name doesn't match, unless someone there actually knew: in Studco v. 1st Advantage (Fourth Circuit, 2025), unreviewed automated alerts didn't count as knowing.
Outside the UK, payers carry most of the loss. In the European Banking Authority's and ECB's data, users bore about 85% of credit-transfer fraud losses in the European Economic Area in 2024, mostly from scams that tricked them into paying.
B2B payments: spend management
Strangers' fraud has a legal cap and a dispute process; your own people's misuse has neither.
| Failure | Who absorbs it | Mechanism |
|---|---|---|
| Stolen card, third-party fraud (credit or charge) | The merchant if the dispute is won; else the company, or an individually liable employee up to $50 | Reg Z; at 10+ cards the contract decides |
| Same, on a prepaid or business debit card | The company or the program manager | No Reg E; Visa's zero liability excludes Corporate and Purchasing |
| Employee misuse | Company, employee or both, by liability model | Not "unauthorized use": the employee had authority |
| Misuse by an employee who has left | Network misuse insurance, if the card was canceled in time | See below |
| Business expense on an individual card, California | The employer | Labor Code 2802 |
| Stand-in approval | The issuer at network level, passed down by contract | Visa Rules |
| Customer can't pay its charge balance | Whoever holds the receivable, often the fintech and its funders | Participation agreements |
| Unsubstantiated charge never repaid | The employee, as wages; the employer owes payroll tax | Treas. Reg. 1.62-2 |
| Late reimbursement | The employer | California interest and fees; New York misdemeanor |
| Purchase made by an AI agent | The cardholder, as if it had made it | Visa rule on agentic payment providers |
The waivers. Visa's US program covers misuse by employees who've left: up to $100,000 per cardholder at five or more cards ($5,000 at one to four), for charges billed up to 75 days before the termination notice or made up to 14 days after, if the card was canceled within two business days, with one paid claim per person. Mastercard's covers $25,000 per cardholder at two to four cards and $100,000 at five or more, but not one-card companies, lost or stolen cards, or owners and shareholders above 5%. The Visa text I found dates from 2004 to 2007, though bank pages repeat its figures. In Canada, Visa requires misuse insurance of at least CAD 100,000 per cardholder.
So the law limits what a stranger's fraud can cost a company, misuse by its own staff has no cap, and the waiver covers only leavers cut off within two days.
B2B payments: stablecoins and blockchain payments
| Failure | Who absorbs it first | How |
|---|---|---|
| Wrong address or wrong chain | The payer | Issuer terms: final, no responsibility |
| Payment to a fraudster | The payer | Same; a freeze only if the issuer acts fast |
| An innocent business's wallet is frozen | The holder | Issuer terms and court orders; no appeal |
| Depeg | Whoever holds the coin at the time | Market price; direct redemption only for approved customers |
| Issuer insolvency | Holders, first in line for the reserves | GENIUS Act priority, after a delay |
| Custodian or ramp insolvency | The business customer, possibly as an unsecured creditor | The terms decide who owns the coins |
The contract decides most of it, then statute, then general law. Issuer terms disclaim almost everything: Tether's say holdings "are not segregated assets held in your name or for your benefit", and Circle's say a holder without a Mint account has one remedy, to stop holding USDC. Statute adds the GENIUS Act's priority on reserves for every holder, business or consumer. General law adds UCC Article 12, which lets a good-faith buyer who takes control of a digital asset keep it free of others' claims; 24 states and Washington DC had adopted it by July 2024, and New York signed it in December 2025.
Businesses get no Reg E, whose error rights cover consumers only; the CFPB withdrew its 2025 attempt to stretch them to stablecoins. Who pays for a freeze or a fraud is mostly decided by who holds the coin when it happens.
Card issuing: launching a card program
The network makes the bank liable, the bank pushes the liability to the program by contract, and when the program can't pay, the bank or the cardholder ends up holding it.
| Failure | Who pays first | How |
|---|---|---|
| Anything on the BIN, including an agent's losses | The bank, to the network | Visa's rules make the BIN holder liable for all of it |
| Fraud and operating losses | The program manager | Indemnities, reserves, collateral |
| Credit losses on a partner program | The program manager, through participations or guarantees; then the bank | Contract, then the bank's balance sheet |
| Unauthorized consumer transactions | Mostly the issuer side | Reg E caps the consumer at $50 or $500, depending on how fast they report; network zero-liability rules go further |
| Stand-in approvals when the processor is down | The bank, then the program by contract | Visa holds the issuer responsible for stand-in decisions |
| Ledger shortfall | End users, if nobody else can pay | Synapse: users first, then $46.2 million from the CFPB's fund |
| Sponsor bank order | The bank (penalties); the program (frozen growth, remediation work) | Consent order, then contract |
Coastal Community Bank showed how the backstop works in 2026. One credit partner's guarantee fell short, and Coastal booked $68.8 million in the April-June quarter, mostly a write-down of the partner's credit-enhancement asset, on a consumer portfolio of about $500 million. The contract put the loss on the partner; the partner couldn't pay, so the bank took it.
On debit fraud, the Fed's 2023 data has issuers bearing about 28% of losses, and for a program that share arrives as a contract line. Every relaxed setting (instant approval, instant loads, skipped checks) is also a decision about who absorbs the next loss.
Cloud security (CNAPP): posture, workloads and permissions
The customer carries almost all of it, and the contracts say so. AWS's customer agreement (updated August 14, 2026) says "You are responsible for properly configuring and using the Services" and caps AWS's liability at the fees paid for the service involved in the 12 months before the claim. Microsoft lists configurations, identities and data as the customer's in every column, though its page is "not intended to convey legal conclusions". Google says customers "always remain responsible for their access policies and data", and its "shared fate" program adds blueprints and an insurance channel without changing legal liability.
| Failure | Who pays first | How |
|---|---|---|
| Misconfiguration or a leaked customer key | The customer | Clean-up, settlements, fines |
| The provider's own key or flaw (Storm-0558) | Customers in clean-up; the provider in engineering and reputation | No fine; free logs; a security program |
| A platform default (MFA optional) | Customers so far; the platform is being tested in court | Snowflake's multidistrict litigation |
| A security tool as the way in (Trivy) | Every downstream customer, rotating keys | The vendor's reputation |
| A long cloud outage | Customers, mostly uninsured below waiting periods | Insurers' estimates for the October 2025 AWS outage ran from $38 million to $581 million |
Snowflake is the case to watch. It's a defendant in federal multidistrict litigation in Montana, and October 2025 rulings let negligence and consumer-protection claims proceed on several tracks; two plaintiffs' claims against it were dropped in December 2025 after their companies settled. With no liability finding and no certified class, the theory that an optional-MFA default is negligence is alive and unproven, and I found no public case where a customer recovered breach losses from a cloud provider.
Providers pay in engineering instead: AWS shipped instance metadata version 2 in November 2019, about four months after Capital One, Snowflake is making MFA mandatory, and Microsoft made logs free after Storm-0558.
Identity and trust: KYB, KYC and ongoing monitoring
The regulator fines the licensed party, which pushes the cost down by contract until it reaches someone who can't push further.
| Failure | Who pays | How |
|---|---|---|
| Customer identification or due-diligence failure | The bank, then its fintech | Penalties and consent orders on the bank; indemnities, remediation and freezes for the fintech |
| Merchant fraud at a payfac | The acquirer to the network, then the payfac | Network rules; contract |
| Wrong industry boarded (a high-risk merchant as low-risk) | The acquirer, then the payfac | Network assessments |
| Marketplace skips seller verification | The marketplace | FTC civil penalties; Temu paid $2M in the first case, in 2025 |
| Synthetic identity becomes a bust-out (maxing out credit, then vanishing) | Whoever extended the credit | Often booked as a credit loss, not fraud |
| Deepfake-opened account used as a mule (an account that moves stolen money) | US: mostly the scam victim. UK: sending and receiving payment firms, 50/50 | UK reimbursement rules since October 2024 |
| Account takeover | Consumer accounts: the institution. Business accounts: usually the business | Reg E for consumers; contract for businesses |
| Biometrics collected without consent | The platform and its identity vendor | Illinois's BIPA lets people sue; Texas's law lets the attorney general |
| Bad data from a registry or vendor | Whoever relied on it | Registries don't vouch for filings; vendor contracts disclaim accuracy |
| Sanctions breach | The US person who processed it | Strict liability; contracts move the cost, not the liability |
Illinois's biometric law is the one people underestimate: $1,000 per negligent violation and $5,000 per intentional one, with a right to sue. Since a 2024 amendment, repeated scans of one person count once, but class sizes still make it costly: Onfido settled for $28.5M. Financial institutions are exempt, and courts have extended that to some of their vendors; a marketplace or gig platform doesn't get the shield.
If the landscaping company is really a front and $80,000 of stolen card payments flow through it before you catch on, the cardholders' banks charge the payments back, the acquirer collects from you, and you can't collect from a company that's gone. Every relaxed setting (auto-approval, instant payouts, skipped selfies) is also a decision about who absorbs that kind of loss, and it belongs in the spec and the contract.
Login and access for apps: authentication and authorization
Who pays after a takeover is settled in the contract, by the regulator, in court and by the insurer.
| Failure | Who absorbs it first | How |
|---|---|---|
| Reused passwords, no MFA, on your app | You, then your users | Fines, class actions, notification costs |
| Your identity vendor is breached | You, toward your customers | The vendor's cap: a year of fees in the terms I read |
| A provider's own signing key or bug | The provider in engineering and reputation; customers in clean-up | No fine in Storm-0558 |
| A help desk resets the wrong person | The company that owns the help desk, then maybe its outsourcer | Contract, if proven; Clorox's suit is unresolved |
| A permission check fails | The app | I found no regulator case where that alone was the headline cause |
I read two vendors' terms directly. Okta's master subscription agreement caps liability at the fees paid in the 12 months before the incident, excludes lost data and consequential damages, has no separate cap for security incidents, and makes the customer responsible for activity under its users' logins. Clerk's standard terms (updated October 1, 2026) have the same 12-month cap, cumulative across claims, with no data-breach carve-out. I didn't confirm other vendors' current terms, and negotiated contracts can differ.
"Whoever skipped MFA pays" is only the opening rule. The cap limits what can move to a vendor to a year of fees, against losses that can run to millions. Insurance pays only if your MFA answers were true (in 2022 an insurer had a policy rescinded after the customer overstated its MFA), and 23andMe's class settlement came from insurance and then a bankruptcy estate.
Models and inference: what AI runs on
| Failure | Who absorbs it | Mechanism |
|---|---|---|
| Outage | The builder, minus a small credit where an SLA exists | Bedrock: 99.9% per region, credits of 10-100% of that region's charges as future credit; lab standard tiers: best effort, no credit |
| Silent quality drop | The builder | No SLA found that counts wrong answers as downtime |
| Retirement and migration | The builder | Notice periods; no compensation |
| Copyright claim on an output | The vendor, if the indemnity's conditions are met | Paid plan, covered model, output unmodified, filters on; usually outside the liability cap |
| Copyright claim over training data | The lab | Bartz, the New York Times case; reaches builders only through prices |
| Harm from an answer | The deployer by default | Air Canada's chatbot case; consumer cases against model makers are pending or settled |
| Unused compute | The lab, then its investors | Take-or-pay contracts: about 80% of Anthropic's commitments, as reported |
| GPU cloud failure | Lenders, then customers mid-contract | Debt against contracts |
Builders carry price and capacity risk, labs carry volume risk, GPU clouds carry financing and obsolescence risk, and household electricity customers carry part of the grid cost. Liability caps in the labs' and clouds' terms are typically 12 months of fees.
Observability: metrics, logs and traces
The customer carries most of the risk in observability, and the contracts say so.
| Failure | Who pays first | How |
|---|---|---|
| Bill shock from a spike | The customer | Usage is billable by contract; credits are a goodwill gesture |
| Vendor outage | The customer, in lost visibility; the vendor, in credits and lost usage revenue | SLA service credits, the "sole and exclusive remedy" |
| Missed alert leads to an outage | The customer and its users | Nothing in the vendor's contract covers it |
| Personal data logged | The customer as controller | GDPR fines land on the controller, not the logging tool |
| Privileged host agent crashes machines | Customers and insurers, then the vendor if negligence is proven | Liability caps, tested in court |
| AI agent's wrong suggestion | The customer, through the human who approved it | Vendors' terms don't accept liability for remediations |
The SLAs are short. Grafana Cloud promises 99.5% of requests succeed each month and credits 10% to 100% of the affected service's monthly fee as availability drops; it also covers alert evaluation, so a missed alert is contractually defined there. New Relic targets 99.8%, with termination as the remedy if it misses badly two months running, and I found no public Datadog SLA to compare.
Datadog's own outage on March 8, 2023 shows the vendor side. A security update auto-applied across its nodes knocked out networking, and the company said it lost about $5 million of revenue, because under usage pricing it couldn't bill for data it couldn't ingest. Customers were blind for most of a day and got credits at best.
CrowdStrike's July 2024 outage is the extreme case for any privileged host agent, which observability agents are becoming as they take on security work. One estimate put direct losses to Fortune 500 companies at $5.4 billion, and CrowdStrike says its liability cap limits its exposure to "single-digit millions". A court let Delta's negligence claims proceed in May 2025.
Phone numbers, messaging and voice
The default answer to "who pays?" is the sender, then whoever has a contract with the sender.
| Failure | Who pays | Mechanism |
|---|---|---|
| Message blocked or filtered | Sender | Blocked messages can still be billed (stated explicitly for unverified toll-free) |
| Non-compliant content | Carrier fines the provider, which passes it down by contract | T-Mobile's Sev-0 fines are reported at $500 to $2,000 per violation since 2024 |
| SMS pumping | Sender, for every message | Fraudsters split termination revenue with a complicit operator; customers who switch off a provider's geographic protections carry the loss explicitly |
| TCPA violation | Sender | Class actions; the reassigned-number safe harbor covers only a caller who proves it checked |
| Miscategorized WhatsApp template | Business | Charged at the category Meta applied at send time |
| RCS replies and fallback | Platform as RCS partner | Owes carriers for billable events, including US replies and STOP taps; fallback SMS is its own cost |
| False attestation or RMD filing | Signing provider | Token revocation, RMD removal, a $10,000 base forfeiture for false RMD information |
| 911 failure | Shared | The NET 911 Act gives VoIP providers liability protection no less than local phone companies get; notice duties stay with the provider; terms push address accuracy onto customers |
| Port-out fraud on mobiles | Wireless carrier | Authentication, notification and lock duties |
Take a fintech whose sign-up page texts a code to any number typed in. Overnight a bot enters 100,000 numbers in a few small countries, each text costs several cents, and the carriers there split that revenue with the fraudster. The fintech gets a bill of several thousand dollars for sign-ups that never existed, and it'll ask whether you should have stopped it.
Every setting that relaxes a protection (geographic permissions, rate limits, consent checks) is also a decision about who absorbs the loss, so write the answer into the spec and the contract.
Voice AI: real-time voice agents
The deployer owns what its agent says. When Air Canada's chatbot invented a bereavement-fare rule in 2024, the tribunal rejected the airline's argument that the bot was a separate legal entity and made it pay about C$812, and the same principle applies to a voice agent that quotes a refund on a recorded call.
Platform contracts push the rest down to the customer:
| Clause | Retell (June 2026) | Vapi (September 2026) |
|---|---|---|
| Who gets consent | The customer only | The customer |
| AI disclosure | Required at the start of outbound calls | Not explicit |
| Customer indemnity | Covers regulatory fines, the TCPA included; uncapped | Covers TCPA claims |
| Vendor's cap | 12 months of fees | The greater of $100 or 12 months of fees |
| Accuracy | Not warranted; the customer assumes the risk | All warranties disclaimed |
Plaintiffs are testing whether that holds. A December 2025 suit claims OpenAI and Twilio are liable as makers of AI robocalls that another company sent through them, two California rulings in 2025 let wiretap claims proceed against AI vendors able to use call audio for their own models, and nine biometric suits filed in May 2026 accuse ElevenLabs and big tech companies of building voiceprints from public recordings to train models. An indemnity doesn't stop a plaintiff from suing the vendor directly; it only moves the bill, and only if the customer can pay.
Carriers carry their own share through attestation: Lingo Telecom paid $1 million for signing spoofed AI calls. Fraud losses depend on the payment rail and the country. In the US the victim of a scam that tricks them into sending money usually bears it. Pindrop reportedly offers a deepfake warranty of up to $1 million a claim, the only explicit transfer of this risk from a security vendor I found.
Field Guides are learning notes, not legal or compliance advice. Rules and fees change; check the cited primary sources before you act on anything here.