Entity and identity
What is the unit of record, and how do we know it is the same one?
Every industry keeps records about one core thing, and has a rule for deciding that two records describe the same one. Get the rule wrong and every bill, report and audit downstream is quietly wrong too.
What it is
The entity is the noun the business can't function without: the account holder, the patient, the subscriber, the employee, the legal matter. Identity is the rule that decides the record in front of you today is the same one you saw last week, even after a name change, a merger, a new phone or a typo.
The hard part is almost never the entity. It's the identity rule, and who is allowed to override it.
Where you'll see it
| Industry | The entity | How identity gets decided |
|---|---|---|
| Payments | Account holder, merchant, card | Identity checks on people and businesses; a token stands in for the raw card number |
| Telco | Phone number, sender, brand | The number is an address, not a person; ownership is proven with documents and registrations |
| Healthcare | Patient | Matching on name, birth date and identifiers; a duplicate record is a safety risk |
| Payroll | Employee | Tax ID plus employment record; one person can hold two jobs at the same company |
| Legal | Client, matter | Conflict checks decide whether a "new" client is really an old adversary |
Questions to ask in week one
- What is the one record everything else hangs off?
- What makes two records the same one, and who decides when it's ambiguous?
- Can one real-world thing end up with two records, or one record cover two things? What breaks when it does?
- Which identifiers do we issue, and which are issued by someone else?
- What happens to the record when the real-world thing goes away?
The trap
Software PMs think of an ID as a primary key you mint. In regulated industries the identifier that matters is often issued by someone else (a tax authority, a numbering plan, a card network). You don't control the identity, you mirror it, and your data model has to survive the day the issuer changes its mind.
In the Field Guides
AI agents: orchestration platforms
Every agent action involves at least three identities, and most logs record only one.
| Entity | Who issues the identity | Where it breaks |
|---|---|---|
| End user | The company's identity provider | Retrieval must run as them, even on a run resumed hours later |
| Agent | The platform or cloud | A shared service account erases it from downstream logs |
| Tool or MCP server | Its OAuth client ID; a registry namespace | Names are unique per server only, so one tool can shadow another |
| Run, agent version, tenant | The platform | Without a version, a run can't be reproduced |
| Legal role | The law: EU "deployer" or "provider" | Repurposing an agent for hiring can make you the provider |
The key standard is old. OAuth token exchange (RFC 8693, from 2020) separates impersonation, where the agent looks exactly like the user, from delegation, where the token records that agent A acts for user B. Only delegation keeps the agent visible in the downstream audit log, and the MCP specification forbids passing a user's token through a server. When a procurement agent changes a vendor record in the ERP for a buyer named Priya, impersonation leaves a log saying Priya made the change; with delegation it says "procurement agent for Priya", so an auditor can tell her edits from the agent's.
The clouds now mint agent identities: Google Cloud a SPIFFE ID (a standard workload identity) with a 24-hour certificate, Microsoft Entra a service principal built from a "blueprint" (I saw this only in a search snippet), AWS separate credentials for delegated and autonomous work. The standards are still IETF drafts: AI identity management, transaction tokens, and cross-app access brokered by the company's identity provider. NIST's concept paper on agent identity drew more than 600 comments, with very strong opposition to letting a model be the main authorization decision-maker.
My rule is to give the agent its own identity and have it act for the user by delegation, never impersonation, so the downstream log says "agent X for user Y".
AI coding agents: writing, reviewing and shipping code
An agent's change has three possible actors: the developer who asked, the agent's own app identity, and whoever set up the automation that started it. GitHub records its agent as the commit author, with the requester as co-author and a link to the session log; an automation's pull request is credited to whoever created the automation. Locally, agents mostly run as the developer, with the developer's tokens, so the audit log can't tell "the developer typed it" from "the agent did it with the developer's token". Payment-card rules now ask for the agent to have its own credentials (see How the rules work).
The law adds an author. Under US copyright only a human can be one: the US Copyright Office said in January 2025 that prompts alone don't make you an author, and the Supreme Court declined to hear Thaler's appeal on March 2, 2026. The Linux kernel lets agents add an "Assisted-by:" tag and forbids them from signing off a change, so a human stays responsible.
A plan tier is an identity too. The same tool trains on your code or not depending on the account: GitHub Copilot's Free, Pro and Pro+ plans have used interaction data for training by default since April 24, 2026 unless the user opts out, while Business and Enterprise are excluded. A developer signing in with a personal account on company code is the classic leak.
AI evals and observability: how teams measure AI quality
The unit that gets evaluated is a release bundle: the dated model version, sampling settings, prompt, tool definitions, retrieval settings and guardrail versions, treated as one thing. Changing any part makes a new bundle. A judge has an identity too: model version, judge prompt and the example answers in it. "GPT-x judge with prompt v3" and "GPT-x" are different instruments, and a score means little unless the report names both.
"The model" is several things: the named model, a dated snapshot, an alias that points wherever the provider chooses, the serving platform and the product wrapped around it. Anthropic's 2025 bugs hit the same model at different rates on its own API, Amazon Bedrock and Google's Vertex AI, so pinning a version wouldn't have helped. Its 2026 problems sat in the Claude Code product layer, and the company said the API and inference layer weren't affected, so teams that built on the API were spared.
Below the bundle sit the eval case, the versioned dataset, the run, the trial (one attempt) and the trace; OpenTelemetry gives traces standard IDs, and nothing gives a dataset row or a grader one. Vendor names shift too: Galileo became Splunk Agent Observability within weeks of its sale.
AI governance and model risk
The governed unit depends on who's asking. US bank guidance governs a "model", which the April 2026 text narrowed to complex quantitative methods built on statistical, economic or financial theory, leaving out spreadsheet arithmetic, deterministic rules and generative AI. Canada's OSFI E-23 widens "model" to include AI and machine learning methods and judgment-based tools. The UK's PRA covers vendor models "regardless of technology". The EU AI Act, NIST and ISO 42001 talk about an "AI system", and most company registers track a use case. I'd key the register on the use case (purpose, users, data, decisions affected) and link models, vendors, prompts and datasets to it, because one model serves many use cases and one use case can swap models.
Roles carry identity too. Banks name a model owner, developer, validator and user. The EU separates the provider, who places a system on the market under its own name, from the deployer, who uses it professionally, and those roles attach per system and can flip. US law keys on the decision-maker's legal role (creditor, employer, insurer, health care provider) and now pulls vendors in as the employer's "agent" (Workday, California's rules), as a consumer reporting agency (a theory in a 2026 suit against Eightfold) or under housing law (SafeRent). For each use case I'd write down which entity decides and which vendor a court could call its agent.
Agent identity and delegated access: how AI agents act for people
Three identities sit behind every agent request: the software (which agent), the person it acts for, and the operator (the company that runs it). They're easy to collapse into one record and expensive to separate later.
The vendors model the agent differently. Microsoft's Entra creates each agent identity from a blueprint (a template that holds credentials and mints tokens for many agents), adds an optional user account for systems that need a mailbox, and assigns a human sponsor who answers for it. Google ties each agent's SPIFFE ID to the resource that hosts it, so redeploying creates a new identity and old grants don't carry over. On the open web the identity is the operator's signing key, and the individual agent is invisible.
In the incidents, the identity that mattered was often the integration: Drift's Salesforce app, Klue's integration, Context.ai's Google OAuth client. Salesforce and Google contained those cases by blocking the app.
Nobody has a reliable count of agent identities: vendor ratios of machine identities to people range from 17 to 109 per person, which mostly shows their definitions differ.
B2B payments: cross-border payouts
Every payout touches at least six identities, and most rails check only one of them.
| Entity | Who issues the identity | Where it breaks |
|---|---|---|
| Payer | Provider's KYB, for its partner bank | Owners and sanctions need refreshing |
| Payee | Their tax authority: RFC or CURP (Mexico), CPF or CNPJ (Brazil), PAN (India), TIN (Philippines) | SPEI requests now carry RFC or CURP |
| Payee's account | CLABE, Pix key, account + IFSC, sort code, IBAN, card, phone | One payee, several identifiers over time |
| Payee's bank | A BIC, screened against US and local orders | A bank under a US order can't be paid |
| Legal employer | The EOR's local entity, or a partner's | The client never sees the partner |
| Tax status | W-8BEN (foreign) or W-9 (US) | The W-8BEN expires |
A design agency pays a freelancer in Mexico every month, and one day an email arrives asking to switch to a new CLABE. SPEI will credit the new account without looking at the name, so the check has to live in your own change-of-details flow.
Name checks exist only on some rails: the euro area's Verification of Payee (mandatory since October 9, 2025), the UK's Confirmation of Payee, Pix showing the recipient's name and masked tax ID, India's free lookup (since April 1, 2025, though banks still credit "only on the account number"), and card networks' Account Name Inquiry. SPEI has none and credits on the CLABE alone.
The standard is moving. FATF's revised Recommendation 16 (June 2025) asks for names, an account or reference, and town and country for both parties above USD or EUR 1,000, plus name-and-account checks at the receiving bank, in national law by the end of 2030. The W-8BEN expires on "the last day of the third succeeding calendar year". And in June 2025 the US Justice Department announced actions against North Korean IT workers who used fake identities to land remote jobs.
B2B payments: order to cash
The rail doesn't know your customer, and no single "customer" record does either. An ACH payment names the payer by routing and account number only; it says nothing about the legal entity or the invoice. Upstream, "customer" splits into the legal entity that owes, sold-to, ship-to, bill-to and payer; credit limits sit at the parent while orders arrive per ship-to. A hospital group's shared-service center might pay for its twelve hospitals from one bank account, so the bank shows you one payer while your credit team manages twelve customers.
| Entity | Who issues the identity | Where it breaks |
|---|---|---|
| Seller | Its legal entities, each with its own remit-to account | Remit-to details must stay current; a factoring deal changes them |
| Payer | The seller's ERP customer master | A parent's shared-service center pays for all its subsidiaries at once |
| Credit account | The seller's credit team: limit, terms, risk class, guarantor | Limit set at the parent, orders per ship-to |
| Payer bank account | The payer's bank | A new "remit-from" account is the main reason cash goes unmatched |
| Order, shipment, proof of delivery | The seller's order system; the carrier | The evidence that wins or loses a deduction |
| Invoice, credit memo, deduction case | The seller's ERP and AR team; the payer adds its PO number | Deductions resolve against memos, not the invoice |
| Remit-to details | The seller | The attack surface for business email compromise (BEC) |
Business checks on sellers, payers and customers
An AR fintech runs full KYB (know your business) on its sellers: legal entity, owners of 25% or more, a control person, sanctions. Payers get lighter checks: bank-account validation, fraud signals, sanctions. A seller granting terms has no legal KYB duty, but it does carry strict-liability sanctions exposure, and it may pull a consumer report on a guarantor, not on an officer who isn't personally liable.
Name-to-account checks are voluntary in the US. The Fed announced a Payee Name Verification service in December 2025, launch unconfirmed, and under UCC 4A-207 a bank may rely on the account number when name and number disagree.
You build and keep the mapping from bank account to payer to credit account to invoice yourself, and cash application and credit control both depend on it.
B2B payments: procure to pay
AP keeps more things apart than it looks from outside:
| Entity | Who issues the identity | Where it breaks |
|---|---|---|
| Supplier (legal entity) | Tax authorities: a TIN certified on Form W-9 (US); a Business Number and GST/HST registration (Canada) | The invoice name doesn't match the tax record |
| Vendor record | The buyer's ERP vendor master | One supplier, several records: a duplicate-payment and fraud risk |
| Remit-to and bank account | The supplier, through the buyer's change process | The attack surface for business email compromise (BEC) |
| Buying entity | The buyer's own legal entities, often many | An invoice addressed to the wrong entity is rejected |
| Invoice, credit memo | The supplier | A re-sent or renumbered invoice becomes a duplicate |
Is the supplier real?
The W-9 is where a US supplier certifies its taxpayer identification number (TIN) under penalty of perjury. The IRS's free TIN Matching checks 25 name-and-number pairs instantly, or 100,000 within 24 hours; a missing or wrong TIN triggers 24% backup withholding, tax the buyer must hold back. Canada has no W-9: AP collects the Business Number and checks the GST/HST number in the CRA registry, keeping the result as evidence for tax credits.
Is the bank account really theirs?
The bank account is the field a fraudster most wants to change, and no Nacha rule makes a buyer validate a supplier's account before an ACH credit, though Nacha's December 2025 tips recommend it, plus confirming changes through contacts already on file and two people signing off. The Fed announced a Payee Name Verification service in December 2025, with no launch date I've seen. And UCC 4A-207 (US commercial law on funds transfers) lets the receiving bank rely on the account number when name and number disagree.
Say an AP clerk gets an email from a packaging supplier they've paid for years: new bank, new account number, please update before Friday's run. The email looks right. The safe move is to call the supplier on the phone number already in the vendor master, not the one in the email, and hold the first payment to the new account until someone has confirmed it.
B2B payments: real-time payments and open banking
On US rails the payee is a routing number and an account number. There's no directory of aliases like Pix keys or UPI's payment addresses, and no required name check, so a US request for payment needs the payer's account details up front, from onboarding, an earlier payment or an account link.
Other rails tie the account to a person or company: Pix keys must match Brazil's tax-ID records (since July 2025), UPI apps show the name the bank has on file, and the UK and EU check names before payment. The Fed sells Payee Name Verification, which works on any US rail and draws on 12 months of payment history.
Who the payer is decides who pays after a scam. Consumers are covered in the UK and, in a small way from 2027, in India. In the UK so are micro-enterprises (fewer than 10 staff and €2 million or less in turnover or balance sheet) and small charities, while a 50-person company gets no cover anywhere.
Platforms add a layer. A payables platform sending RTP payments for 2,000 small businesses through its sponsor bank sends "on behalf of" them, and RTP's new rules ask the bank to know and screen each one, with minimum requirements from March 31, 2028.
B2B payments: spend management
Each card carries three identities, and each expense adds two more.
| Entity | Who issues the identity | Where it breaks |
|---|---|---|
| Legal issuer | The sponsor bank, as BIN licensee and creditor | The brand on the card isn't the issuer |
| Company (account) | The bank's KYB, done by the program manager: owners of 25% or more, a control person | Refresh is a bank duty pushed onto the fintech |
| Employee cardholder | The company's HR records | Leavers keep cards; work state and province drive tax and labor rules |
| Merchant | The acquirer assigns the category code; the name is cut to 25 characters | The code describes the merchant, not the item |
| Receipt issuer | The merchant's own name and, in Canada, its GST/HST number | The receipt name doesn't match the card descriptor |
Reg Z (the US credit card rule) counts organizations as cardholders, so on a corporate account the company is the cardholder. The employee becomes a debtor, and needs a credit check, only under individual ("the cardholder will be solely liable", in Mastercard's definitions) or joint liability. Under every model, though, the employee's details drive rules: work state, province (which HST factor), Quebec (a second tax slip), owning 10% or more of the company (no lodging per diem for related parties), and a federally regulated Canadian employer (a 30-day reimbursement rule).
Visa's merchant data rules have the acquirer pick a code for the merchant's main business and cut the name to 25 characters; marketplaces use one code (5262) for everything they sell. So the network knows roughly who was paid, never what was bought. US companies stopped reporting beneficial owners to FinCEN in March 2025, but banks still collect them, and sponsor banks push that work onto the program manager. I'd model the bank, the company, the employee and the merchant as separate identities, because each answers a different question.
B2B payments: stablecoins and blockchain payments
Six companies can touch one payment (payer, payee, orchestrator, off-ramp, issuer, custodian), and each is checked by someone. On the chain the counterparty is only an address: free to create, anonymous on its face, and one company can have hundreds. The work is mapping addresses to companies: an allowlist that ties each address to a verified owner, and a deposit address or payment request per payer so incoming coins can be matched. Regulators are pushing the same way. In the EU a firm must check who owns a self-hosted wallet above €1,000, with a signed message or a test transfer, and Brazil asks providers to identify the owners of self-hosted wallets they deal with.
Picture a supplier that emails to say its USDC address has changed. Nothing on the chain checks that the new address belongs to the supplier, so the check lives in the payer's change-of-details flow, as it does for bank details: a call to a known contact and a test payment.
Identity also decides which law applies. Under Treasury's August 2026 proposal, a coin is "issued in the US" when the issuer or the recipient is in the US, so the address map also says who counts as American. On bank tokenized-deposit networks every participant is the bank's own client, which makes identity stronger than on a public chain.
Card issuing: launching a card program
A card has at least three legal identities behind it, and embedded issuing adds a fourth. The network member is the bank, the issuer of record. The program manager is the bank's registered agent; Visa calls it a Third Party Agent and needs a separate registration for each bank relationship. Then there's the cardholder, and on a platform, the business customer whose workers or users get the cards.
With the platform's logo on the front, the bank is still the issuer, named in the cardholder agreement and liable to Visa. Adding another sponsor bank means another agent registration, another contract, another set of audits and probably another design approval.
Prepaid adds a role. FinCEN's prepaid access rule has each program name one "provider of prepaid access" by contract, which registers as a money services business. A bank can't be that provider, but where a bank has "primary oversight and control" of the program, nobody has to register, and most banking-as-a-service contracts are written that way.
The card itself has two identities: the card number, in an account range inside the bank's BIN, and the network tokens that wallets and merchants store in its place. Visa assigns responsibility for a token to the bank behind the BIN, so tokens follow the bank.
A check I'd run early on: is the same legal entity named as issuer on the card, in the cardholder agreement and in any FinCEN registration? It usually is; when it isn't, I'd find out why before launch.
Cloud security (CNAPP): posture, workloads and permissions
The unit is the cloud resource: an AWS ARN, an Azure resource ID, a Google resource name. Vendors bill on it, and each counts it differently:
| Billing entity | Who uses it | Rough list price (Oct 2026) |
|---|---|---|
| Workload (a running VM, container or serverless function) | Wiz | $240-380 a year, less at volume |
| Resource unit (1 VM = 12 functions = 18 images = 125 identities) | AWS Security Hub | $3.75 a month |
| Billable resource (servers, storage accounts, databases) | Microsoft Defender CSPM | About $5 a month |
| Host | Datadog, Sysdig | $10-25 a month at Datadog |
| Developer | Wiz Code, Snyk | About $585 a licence a year for Wiz Code |
The same estate can come out very differently under each, so comparing quotes starts with converting them to one unit.
Cloud identity is the other entity: people, roles, service accounts, workload identities and integration roles. Machine identities far outnumber people, though estimates run from 10 to 1 (Microsoft, 2023) to "up to 40,000" to 1 (Sysdig, 2025), which mostly shows the definitions differ. Microsoft's 2023 research found identities use about 1% of the permissions they're granted. Non-human identities are where cloud breaches run: old access keys, OAuth tokens, CI tokens and, in Microsoft's Storm-0558 case, a signing key. The scanner is one too: a CNAPP holds read rights, and often snapshot and encryption-key rights, in every account you connect.
Customers and businesses that a platform verifies are a different kind of identity, covered in Identity and trust; AI agents' identities are in AI agents: orchestration platforms.
Identity and trust: KYB, KYC and ongoing monitoring
Products tend to blur the legal entity (the landscaper's LLC), the account on your platform (its two locations, each with its own payouts) and the person (an owner who also owns a second company you onboarded last year).
Keep them as separate records. One entity can have many accounts. One person can own several entities, sign for others and be a consumer customer too. When a fraud team asks "what else is this person connected to?", that link is the answer.
Identifiers are local, and none covers every small business:
| Identifier | Who issues it | Where it falls short |
|---|---|---|
| State file number | Each US Secretary of State | One per state; a company registered in four states has four |
| EIN | IRS | Not public; a sole proprietor may use their SSN instead |
| Corporation number and Business Number | Corporations Canada or a province; CRA | Federal and provincial registers are separate |
| Company number (CRN) | Companies House | UK only |
| EU register number (EUID) | National registers | Ownership data sits in a separate register |
| LEI | GLEIF, through local issuers | About 3 million worldwide; rare among small businesses |
| D-U-N-S | Dun & Bradstreet | Proprietary |
An ownership link needs more than a percentage. Give each one a type (shares, votes, a right to appoint the board, a trust role, a nominee), a date and a source. Thresholds differ by a word: the US, Canada and the EU say 25% "or more"; the UK says "more than" 25%. An owner with exactly 25% is in scope in Toronto and out of scope in London.
On the people side, synthetic identities and account takeover both break the model. A synthetic identity breaks the assumption that a set of details belongs to one real person. Account takeover breaks the assumption that the person logged in is the person you verified.
Login and access for apps: authentication and authorization
The records to keep apart: the person, the organization (tenant, workspace or account, depending on the product), the membership linking them with a role, the connection to that organization's directory, each credential (password, passkey, authenticator seed, recovery codes), each session and token, and each non-human identity: service accounts, registered OAuth clients, API keys and agents. One human often holds several identities, such as a work login through SSO, a personal Google login and an old password account at the same domain.
Key every identity on a stable ID from its source: issuer plus subject in OpenID Connect, issuer plus name ID in SAML, the directory's external ID in SCIM. Email changes and gets recycled, and linking accounts because emails match is how attackers set up a takeover before the real user signs up.
Most B2B apps use one global pool of users with memberships (the Slack and GitHub shape), so a person can belong to several organizations; separate pools per customer isolate harder and cost more to run.
Models and inference: what AI runs on
"The model" has at least three parts: a family (Sonnet, gpt-6.1-sol), a snapshot ID (the exact version), and the host or deployment that serves it. On Azure you call a deployment name you created, and on Bedrock the same Claude model has its own ID and its own retirement date. Open-weight models add the precision and serving stack, which is why OpenRouter lets you filter hosts by quantization. The billing identity is a tree of organization, workspace or project, and API key, and rate limits attach to the organization (Anthropic, OpenAI) or the project (Google), per model, so teams in one organization share its headroom.
The law names its own entities. Under the EU AI Act the "provider" places a model on the market and the "deployer" uses it, and a company that fine-tunes with more than a third of the original training compute becomes the provider of the result. An IP indemnity keys on plan, service and model together, so one API key can be covered for one model and not another. US export rules now follow a company's ultimate parent, wherever its subsidiary is registered (guidance of May 31, 2026).
Observability: metrics, logs and traces
In telemetry the unit of identity is the resource: the service name, its version, the environment, the host, the container, the Kubernetes pod, the cloud account and region. OpenTelemetry standardizes these attribute names, and a trace adds a trace ID and span IDs that follow a request across services.
Vendors bill on entities too, and they don't agree on which entity counts:
| Billing entity | Who uses it | Rough list price (Oct 2026) |
|---|---|---|
| Host | Datadog, Splunk Observability | $15-75 per host a month, by product tier |
| Memory per host-hour | Dynatrace Full-Stack | About $58 a month for an 8 GiB host |
| User | New Relic, Grafana, incident tools | $8-349 per user a month, by role |
| Event | Honeycomb | From $150 a month for 50 million events |
| Active metric series | Grafana Cloud | About $6.50 per 1,000 series |
| Custom metric | Datadog | $5 per 100 a month |
| Gigabyte | Everyone, for logs | Cents to dollars, depending on indexing |
Containers, serverless functions and autoscaling make "how many hosts" a moving number, so vendors bill on percentiles of hourly counts or "host-equivalents". The identity most often missing is ownership. Which team owns checkout-api? That lives in a service catalog if anywhere, and an alert without an owner pages whoever happens to be on call.
"Agent" has a double meaning, often on one product page: the software on each host that collects telemetry, and an AI agent. I'd write "host agent" and "AI agent" every time. Telemetry also carries other people's identities, such as user IDs, IP addresses and emails in URLs, which is where the privacy rules come in.
Phone numbers, messaging and voice
People mix up the number (an E.164 address, the international format of country code plus national number, allocated to a carrier and reaching you through resale), the sender (whatever the recipient sees) and the business (the legal entity a gatekeeper has checked) all the time.
| Sender | Unit of identity | What proves it |
|---|---|---|
| 10DLC | Brand → campaign → numbers (one campaign per number) | TCR matches legal name, EIN and address to official records; Auth+ emails a named representative (public companies, since Aug 2025); optional vet score |
| Toll-free | Each number | Verification, with a business registration number required for new submissions since early 2026 |
| Short code | Each program, per carrier | Registry vetting (name, tax ID, legal history), then each carrier's review |
| Alpha abroad | Sender ID × country × carrier | Proof of the brand link: India registers entity, header, template and consent; Spain wants a trademark, trade name or domain |
| RCS | Agent, owned by the platform as Google's partner | The brand's contact answers Google's verification email, once per agent |
| Portfolio → WABA → number → display name | Meta business verification, display name review | |
| Voice | The number on the call | The signing provider's STIR/SHAKEN attestation |
A 10DLC sole proprietor is a person or business without an EIN (an LLC with one isn't), verified by a one-time code to a US or Canadian mobile; providers say it gets one campaign and one number. And Auth+, today required only of public companies, will extend to new brands of every type except sole proprietors from January 21, 2027, according to providers.
In every case someone else issues and checks the identity, at a different level for each sender type, so I'd model brand, campaign, sender and number as separate records, each with its own history.
Voice AI: real-time voice agents
Three identities meet on every call: the business the agent speaks for, the voice (synthetic, cloned or human) and the phone number. The caller's own number can be spoofed. STIR/SHAKEN attestation, when present, says how much the originating carrier vouches for it: an A means the carrier knows its customer may use the number, and says nothing about the voice or the words.
Inside the stack, one call collects IDs from the carrier, the communications platform, the voice platform, the model session and the trace. Joining them is real engineering work, and you need it before you can debug a bad call or reconcile a bill.
The law names its own entities, and one platform can be several at once:
| Legal role | Law | What it triggers |
|---|---|---|
| The "maker" or initiator of a call | TCPA | Consent, caller identification, opt-out |
| A third party listening in | California's wiretap law | $5,000 per violation if callers weren't told |
| Provider or deployer | EU AI Act | Watermarking for the provider; disclosure for both |
| Business associate | HIPAA | A BAA, and BAAs with every sub-vendor |
| A collector of voiceprints | Illinois's biometric law (BIPA) and others | Written consent and a retention policy |
A voiceprint is a biometric template that identifies a speaker. A plain transcription agent generally doesn't build one, though courts are testing that line in cases about meeting transcription tools.
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.