Networks and counterparties
Who sits between us and the outcome, and what do they want?
Almost nothing in a regulated industry happens end to end inside one company. Name every party between you and the outcome, what each one earns, and which ones can say no.
What it is
Counterparties are the organizations your product depends on to deliver an outcome: networks, intermediaries, registries, partners, regulators. Each has its own incentives, timelines and power to block you.
Your roadmap is partly written by people who don't work for you.
Where you'll see it
| Industry | Who sits in the middle |
|---|---|
| Payments | Card schemes, issuing banks, acquirers, processors |
| Telco | Mobile carriers, aggregators, registries, Meta and Google |
| Healthcare | Payers, providers, clearinghouses |
| Payroll | Tax authorities, banks, benefits providers |
| Legal | Courts, opposing counsel, e-billing platforms |
Questions to ask in week one
- Who are the parties between us and the customer's outcome?
- What does each one earn from the transaction, and what does it risk?
- Who can block, delay or reverse an outcome, and on what grounds?
- Which relationships are direct, and which go through an intermediary?
- Where are we the counterparty in someone else's chain?
The trap
Treating counterparties like vendors you can switch. The ones that matter most are often networks or monopolies you can't route around, so their policy changes land on your roadmap whether you planned for them or not.
In the Field Guides
AI agents: orchestration platforms
An agent run is a chain of other companies' systems, and each can change the agent's behaviour, price or lifetime without a release on your side.
| Party | What they control | What they earn |
|---|---|---|
| Model lab | Behaviour, prices, caching, rate limits, retention, retirements | Tokens, hosted tools, runtime |
| Cloud (AWS, Google Cloud, Microsoft) | Hosting, its own model dates, identity, network, marketplace | Resold tokens, runtime, memory, logs |
| Orchestration platform | Releases, state, approvals, traces, evals | Licence, seats, consumption, outcomes |
| Systems of record | The data, its permissions, API limits | Seats, credits, conversations |
| MCP server authors | Tool definitions, input checks | Often nothing |
| Identity providers | Who the user and the agent are | Licences |
| Third-party websites | Terms that can forbid agents | Nothing from the agent |
Some of this surprised me:
- One model, several lifecycles: a lab's model on a cloud can follow the cloud's retirement dates (OpenAI's model on Bedrock is an exception), and Anthropic bills usage bought through AWS or Azure marketplaces in small consumption units, so a cloud commitment can pay for it.
- Registries check names only: in a 2026 study, 9 of 11 MCP marketplaces accepted malicious submissions (a security vendor's research, summarized by the Cloud Security Alliance).
- Websites push back: Amazon sued Perplexity over its shopping agent (see Liability allocation).
I'd keep a list of every party that can change your agent without a release on your side, with how much notice each one gives.
AI coding agents: writing, reviewing and shipping code
| Party | What they control | What they earn |
|---|---|---|
| Agent vendor | The harness, the routing, the session log | Seats, credits, a margin on tokens |
| Model lab | Quality, prices, rate limits, retirements, supply | Tokens, subscriptions, its own agent |
| Repo host and CI | The pull request, approvals, required checks | Seats, CI minutes, security add-ons |
| Package registries | Which code can be installed | Usually nothing directly |
| Review and security vendors | Comments, checks, findings | Per developer or pull request author |
| Observability vendors | What to fix next | Ingest and seats |
| Cloud | Compute, marketplace billing, its own IDE | Resold tokens, commitments drawn down |
One company can sit in many of these roles. Microsoft owns GitHub, Copilot, Advanced Security, npm and Azure, writes the indemnity, trains its own coding model and offers both OpenAI's and Anthropic's models in Copilot. Google sells Gemini Code Assist, Antigravity and Jules, and owns Wiz and Mandiant. Amazon sells Kiro, runs AWS and invests in Anthropic. Every scanner or bot in CI is also one more token holder: a poisoned scanner, Trivy, became the way into other companies' pipelines in March 2026 (see Cloud security (CNAPP)).
AI evals and observability: how teams measure AI quality
| Party | What they control | What they absorb when quality slips |
|---|---|---|
| Product team (PM, engineers) | The gate, the rollout, the bundle | Rework, rollbacks, the customer conversation |
| Domain experts | What "good" means | Their time; they're scarce and rarely budgeted |
| Model providers | Model versions, aliases, retirement dates, terms | Reputation when the regression is theirs |
| Eval and observability vendors | The viewer, the workflow, the stored history | Little; contracts don't cover missed regressions |
| Users | The real ground truth, through behaviour | The bad answers, usually first |
The domain expert is the counterparty I'd underestimate least. Husain and Shankar recommend one domain expert who sets the standard, and that person's labels are the most valuable asset in the stack. Their method comes from practitioners who also sell a paid course on it, so I treat it as well-tested advice rather than settled research.
Users are also, in practice, the monitor of last resort, and their feedback is biased: OpenAI said its GPT-4o update over-weighted short-term thumbs-up and thumbs-down signals, which is how flattery looked better in the numbers.
AI governance and model risk
Inside the company, banks use the "three lines" model, and other companies copy it loosely:
| Line | In a bank | Elsewhere |
|---|---|---|
| 1st: build and run | Model owner, developer, user | Product owner and the ML or platform team |
| 2nd: challenge | Model risk management: validation, policy, inventory oversight | AI governance office, often in privacy, risk or legal |
| 3rd: check the checkers | Internal audit, which evaluates the program without redoing validation | Internal audit if there is one; a certification body for 42001 |
Outside banks, privacy teams and legal or compliance teams each own AI governance in about a fifth of companies, ahead of IT, in the IAPP's 2025 survey. Around the company sit labs and clouds, SaaS vendors with AI inside, assurance firms, certification bodies and their accreditors, regulators, customers' procurement teams, plaintiffs' lawyers and insurers. Supply is concentrated: in the Bank of England and FCA's 2024 survey, the top three model providers made up 44% of all the model providers UK financial firms named.
Agent identity and delegated access: how AI agents act for people
Agent identity runs on several networks at once. Identity providers and the SaaS apps that trust them decide enterprise access, and Cross App Access moves the consent decision from each app to the provider. Content delivery networks verify agents at the edge for millions of sites. Card networks act as identity directories for agents that pay: Visa publishes keys for agents in its protocol, and Mastercard lets only registered agents transact. The FIDO Alliance started a working group on agentic authentication in April 2026, when Google donated its AP2 payments protocol and Mastercard contributed Verifiable Intent.
The counterparty that shows up most in loss chains is the integration vendor holding other companies' tokens: Salesloft's Drift, Klue, Context.ai. Each was a small tool with standing access to much bigger systems.
B2B payments: cross-border payouts
Any of the parties below can stop a payout.
| Party | What they control | What they earn |
|---|---|---|
| Sending bank | Cut-offs, FX if it converts, the correspondent | Wire fee, FX markup |
| Correspondent banks | Route, screening, deductions | Lifting fees, balances |
| Beneficiary bank | Credit timing, queries, FX | Fee; FX margin 0.7% to 1.1%, about 1.7% in Latin America |
| Payout provider | Routing, quotes, where prefunding sits, holds | FX markup, fees, float |
| Local partner (bank, e-money institution, FX broker, licensed remitter) | Rail access, licence, its own KYC | Per-payout fees, local FX, float |
| Rail operators, central banks | Rules, limits, hours, who participates | Cost-recovery fees |
| Card networks and issuers | Push rules, Fast Funds, limits | Per-transaction fees, FX |
| Partner EOR | The local employing entity | A wholesale fee |
Correspondents are disappearing: active correspondent banks fell about 22% between 2011 and 2019 while payment values grew. The Fed is leaving too: FedGlobal ACH, which credited pesos in Mexico the next banking day, takes its last forward items on November 20, 2026, citing "steep declines in transaction volumes".
Local access is gated. A foreign payout company needs a regulated Mexican participant to reach SPEI; Pix is for central-bank-authorized institutions, with a R$15,000 cap for the rest; India's FX runs only through authorized dealer banks. Some doors are opening: Faster Payments has almost doubled its direct participants since 2018, and RPAA-registered providers may join Payments Canada since September 29, 2025. But the Fed's May 2026 proposal leaves out state money transmitters, so US payout fintechs still need a bank.
That makes the local partner the dependency I'd watch most closely, since it holds the licence, the rail connection and your prefunded money.
B2B payments: order to cash
An order-to-cash flow touches a dozen organizations, and the ones that matter most never sign your contract: the payer's AP team, your sponsor bank, and whoever insures or finances the receivables.
| Party | What they control |
|---|---|
| Payer AP team | The rail, the remittance channel, the portal, the payment run, deductions, often the terms |
| Sponsor bank (originates ACH and holds funds for a fintech) | Origination limits, reserves, KYB depth, return-rate tolerance; it can stop your origination |
| Seller's bank | Lockbox, cash reporting, ACH limits |
| Payer's bank, card issuer or AP platform | Virtual card programs that push card costs onto suppliers |
| Card networks and processors | Interchange, surcharge rules, data qualification, disputes, funding |
| Rail operators | Nacha writes ACH rules; the Fed runs FedACH, Fedwire, FedNow and check services; The Clearing House runs RTP and EPN; Payments Canada runs ACSS, Lynx and RTR |
| Bureaus, credit insurers, factors and lenders | The data behind limits; per-customer limits an insurer can cut; eligibility rules and remit-to control |
| Carriers, retailer compliance programs, agencies | Proof of delivery; fines for late or short deliveries; escalation and suit |
| ERP vendors | The invoice record, the cash application module, the integration marketplace |
So this behaves like a network business as much as SaaS: every new seller brings payers who have to change behavior, and every payer already has someone else asking it to pay a different way.
ERP vendors are partners and competitors: several mid-market ERPs now embed payments through partners, by their own account. I'd watch for the one that takes the "pay" button.
B2B payments: procure to pay
Two kinds of network compete to own the connection to the supplier.
- Buyer-side networks: supplier portals, issuers' card enrollment programs, AP payment networks. Each portal is a walled garden: in the Fed's words, accounting systems are "closed-loop and not interoperable", so suppliers juggle many portals.
- Open exchange networks: DBNAlliance (a US e-invoice network launched in 2023 by Business Payments Coalition pilot participants), Peppol and EDI value-added networks. DBNAlliance uses Peppol's four-corner model, where a business joins through a service provider. Neither country has a national Peppol authority; OpenPeppol fills that role.
Virtual card acceptance is the network that surprises people. One corporate-payments company's annual filing says each issuer negotiates acceptance with each supplier, so issuers' virtual cards aren't interchangeable. Visa's Supplier Matching Service checks a buyer's vendor file against Visa's merchant data monthly to flag suppliers that take commercial cards, and issuers use it to size a program.
| Party | What they control |
|---|---|
| Buyer's bank | Treasury agreements, security procedures, Positive Pay, limits, often the card program |
| Card issuer and network | Interchange, rebates, enrollment campaigns, straight-through programs |
| AP network or payment hub | The enrolled-supplier list, payment status, which rail each supplier gets |
| ERP vendor | The "pay" button and the integration marketplace |
| Rail operators | Nacha, the Fed, The Clearing House (RTP), Payments Canada |
If your bank's card team has enrolled 400 of your suppliers to take its virtual card and you move the program to another issuer, that enrollment stays behind, and the new issuer starts its own campaign from zero.
B2B payments: real-time payments and open banking
The US has two instant networks competing on rules: The Clearing House is owned by large banks, and the Fed runs FedNow as a public service, adding fraud tools and discounts to win smaller banks. Most small banks connect through core processors and more than 40 certified service providers (March 2026), and a company that isn't a bank reaches either network only through a bank.
Whoever owns the scheme decides whether it's a cost center or a business. Brazil's central bank and the Fed price to recover costs; The Clearing House, NPCI and the card networks answer to owners.
In open banking the bank holds the data and the sign-in step, so aggregators depend on its API and, since 2025, its fees. Wells Fargo and PNC are reported to be steering fintechs to Akoya, which Fidelity's parent, The Clearing House and 11 banks own. In the UK, Open Banking Limited runs the standard until a successor body takes over; its independent chair was named in September 2026.
Card networks have become a counterparty on bank payments too: Visa A2A launched in the UK in June 2025 with a dispute framework, and Mastercard's A2A Protect started there with Monzo, NatWest and Santander.
B2B payments: spend management
A card program is a chain of contracts: network to bank, bank to program manager, processor to bank, bank to company, and bank to employee only under individual liability.
| Party | What they control | What they earn |
|---|---|---|
| Sponsor bank | Charter, BIN, cardholder agreement, credit box, collateral, termination | A few basis points; one fintech-focused bank about 8 (2025) |
| Program manager (the spend fintech) | Customer, product, credit model inside bank policy, rewards | Interchange after rebates, software, float |
| Issuer processor | Authorizations, card records, controls; may hold the interchange contract | Basis points; one public processor about 16 (2025) |
| Card network | Rules, interchange tables, incentives, uptime, misuse insurance | Fees both sides; Visa paid $15.75B in client incentives in FY2025 |
| Funders | Advance rates and covenants on facilities secured by card receivables | Interest |
| Merchant and acquirer | Category code, data; soon, whether to take commercial cards | Pay interchange |
| Travel agency or travel and expense suite | Booking channel, itinerary, duty-of-care data | Booking fees, commissions |
| Accounting and HR platforms | API access and price, the employee list, increasingly their own cards | API fees, bundles |
A few things here surprised me. One processor can be a single point of failure: on October 20, 2025, a change on Visa's side stopped North American traffic reaching one issuer processor, and stand-in rates stayed high until a fix by October 30. Migrations happen too; one public expense company moved its card to a new sponsor bank in 2024, keeping its processor and network. And the interchange contract may not be yours. One public processor's bank contracts give it all the interchange, which it shares with customers. The basis points in the table are my arithmetic from the companies' filings.
A spend fintech owns the customer but rents the license, the rails and sometimes the interchange contract, and any of those landlords can change the terms without asking.
B2B payments: stablecoins and blockchain payments
Any of these can stop or change a payment.
| Party | What it controls | How it can stop a payment |
|---|---|---|
| Issuer | Minting, redemption, freezes, fees | Freezing an address; delaying a redemption it finds suspicious |
| Reserve custodian | Where the reserves sit | Failing, as Silicon Valley Bank did with $3.3 billion of Circle's reserves in March 2023 |
| Chain operators | Which transfers get in, fees, uptime | An outage; on permissioned chains, the validator set |
| Orchestrator and ramps | Bank access, licences, price, routes | Declining, holding or repricing a payment |
| Banks at both ends | Dollars and pesos in and out | Closing accounts; refusing crypto-linked money |
The networks themselves are moving into the stack. Visa and Mastercard both validate on Arc, Visa also on Tempo, and both are founding partners of Open USD, a coin Bridge issues that went live on September 30, 2026.
The counterparty I'd watch most closely is the off-ramp: it holds the local licence and the pesos, and the payer usually never signs anything with it.
Card issuing: launching a card program
Most of a program's counterparties can stop it.
| Party | What they control | What stops if they fail |
|---|---|---|
| Sponsor bank | The licence, BIN, money, approvals, termination | New programs, changes and sometimes the whole program |
| Card network | Rules, rates, registration, token service, stand-in | Everything on that network |
| Issuer processor | Authorizations, card records, files | Spending, unless stand-in approves |
| Card manufacturer and personalizer | Plastic, chips, lead times | New physical cards |
| Wallets | Adding the card to a phone, wallet fees | Tap-to-pay on phones |
| Identity vendors | Checks at sign-up | New accounts |
| ATM networks, instant-payout providers | Surcharge-free cash, instant transfers | Cash access and instant loads |
Concentration runs in both directions. Chime runs on two banks. Marqeta said 64% of its 2025 volume settled through one bank, Sutton, and Block was 45% of its revenue. CompoSecure got 63% of its 2024 sales from two issuers, and The Bancorp's top three partners brought it 59% of its fintech fees. When one side leaves, the other side's numbers move: Galileo's segment revenue fell by about a quarter year on year after its large client left.
Early on I'd write down which single vendor failure stops new card issuance, and which stops cardholders spending; the second list is shorter and scarier.
Cloud security (CNAPP): posture, workloads and permissions
| Party | What they control | What they absorb when it goes wrong |
|---|---|---|
| Security team (CISO, cloud security engineers, SOC) | Tool choice, policy, triage, exceptions | Blame and disclosure decisions |
| Platform or cloud team | Accounts, guardrails, the roles tools need, the commitment | Rework from bad guardrails |
| App teams | The code and most fixes | Engineering time |
| Cloud provider | Infrastructure, defaults, log tiers, the marketplace | Product changes, capped liability |
| Security vendors | Detection, the graph, routing | Reputation; their access can become the attacker's |
| Auditors (QSAs for PCI, CPA firms for SOC 2, 3PAOs for FedRAMP) | Which evidence and exceptions pass | Nothing directly |
| Insurers | Eligibility, price, exclusions | Covered losses |
Outside parties are now the usual shape of a breach. The 2026 Verizon report, as summarized by others, found partners or vendors involved in 48% of breaches, up from 30%. The chains I'd keep in mind run from a contractor's laptop to Snowflake (2024), from a chat vendor's stolen tokens to Salesforce data to AWS keys (Salesloft Drift, 2025), and from a poisoned scanner to a CI pipeline to AWS keys (Trivy, 2026). The tools are counterparties too: Datadog's 2025 data found 12.2% of vendors' integrations into customers' clouds dangerously over-privileged.
One company now sits in several seats. Google runs a cloud, owns Wiz, owns Mandiant (whose report traced the Snowflake breaches to customers' credentials), publishes threat data from its own incidents and sells Model Armor. That's a lot of the story told and sold by one party, which I'd keep in mind when reading their reports.
Identity and trust: KYB, KYC and ongoing monitoring
A business approval runs through a chain, and the platform sits in the middle of it.
| Party | What they control |
|---|---|
| Registries and tax agencies | Whether a company exists; who may check a tax ID |
| Data aggregators | Coverage, freshness and how records get matched |
| KYB and identity platforms | The workflow, decision rules and audit trail |
| Bureaus, SSA, phone carriers through aggregators | Whether a person's details can be confirmed without a document |
| Screening and monitoring vendors | List coverage, alert volume |
| Sponsor bank or acquirer | Policy, vendor approval, the right to stop new onboarding |
| Card networks | Merchant rules, MATCH, fines passed down to the payfac |
| Regulators and intelligence units | Rules, penalties, requests about specific people |
| Consortiums | Shared fraud data: Early Warning in the US, Cifas in the UK |
| Credential issuers and wallet makers | Who gets a digital ID, and whether apps can read it |
Information flows through the chain too. FinCEN sends banks lists of law-enforcement subjects every two weeks (a 314(a) request), and banks answer only with matches. Institutions with an AML program requirement can share information with each other under 314(b). A platform without one can't join, so if you aren't a regulated institution yourself, your bank is both your rule-setter and your only window into what other institutions know.
Login and access for apps: authentication and authorization
Login runs on federation, trust between systems that don't share a database. Customers' directories (Okta, Entra, Google Workspace) decide how enterprise users arrive and leave, and their quirks set your support load. Credential managers (Apple, Google, Microsoft, 1Password) hold passkeys. Phone carriers and email providers carry codes and reset links, so their weaknesses become yours. Outsourced help desks reset accounts on someone's behalf and are a favourite target: Marks and Spencer's 2025 breach and Clorox's suit against Cognizant both centre on one. Insurers write rules through the MFA questions on their applications.
The trust anchors are concrete: a customer's SAML signing certificate, an identity provider's published signing keys and the DNS record that proves domain ownership, each a way in if checked loosely.
Models and inference: what AI runs on
| Party | What they control | What they earn |
|---|---|---|
| Lab | Quality, prices, limits, retention, retirements | Tokens, subscriptions, enterprise deals |
| Cloud platform | Procurement, regions, its own SLA and dates | Resold tokens; commitments drawn down |
| Inference provider | Speed and price for open models | Tokens, dedicated endpoints, GPU rental |
| GPU cloud | Energized capacity | Multi-year take-or-pay contracts |
| Chip maker | Allocation of new chips | Hardware at about 74% gross margin (Nvidia) |
| Gateway | Routing and budgets | A small fee, or nothing (open source) |
| Grid operator and utility | Power and connection dates | Capacity payments, set by auction |
Money flows down the chain and equity flows back up it: Nvidia invests in OpenAI and in GPU clouds and providers, Amazon and Google in Anthropic, Microsoft owns about 27% of OpenAI and has invested in Anthropic, and AMD gave OpenAI warrants for up to 160 million of its shares. Amazon sells Claude on Bedrock, holds Anthropic notes, supplies it Trainium and sells its own Nova models, all at once.
Observability: metrics, logs and traces
| Party | What they control | What stops if they fail |
|---|---|---|
| App teams | What gets instrumented, log levels, labels | Telemetry quality; the bill, through cardinality and verbosity |
| Platform or observability team | Collectors, pipelines, budgets, vendor choice | Collection, sometimes for everyone |
| Observability vendor | Storage, query, alerting, pricing units | Visibility during incidents |
| Pipeline vendor | Routing, sampling, redaction | Data to every destination at once |
| Cloud provider | Compute, network charges, default tools, marketplace billing | Everything running on it |
| Incident tool | Paging, escalation, incident records | Waking the right person |
| Security team and SIEM | Detection rules on the same logs | Security monitoring |
| AI agent vendors | Read access to telemetry, sometimes write access to code | Automated investigation |
Concentration runs both ways. Datadog's largest customer, widely reported to be OpenAI, renewed at nine figures in 2026 but is cutting usage from July, and Datadog built that into its guidance. Coinbase's roughly $65 million a year Datadog bill was restructured in 2023 after Coinbase built a parallel open-source stack. The biggest customers can credibly leave, and the threat works best at renewal.
The pipeline is the new switching point. Whoever owns it decides which backend gets which data, which is why security and observability vendors have both bought pipelines since 2025.
I'd write down early which single failure blinds us during an incident. For many teams it's the collector tier: in OpenTelemetry's 2026 survey of collector users, 23% didn't monitor their collectors at all.
Phone numbers, messaging and voice
A US business text passes through at least four companies before it reaches a phone, and any of them can stop it.
The chain: brand → platform (registered at TCR as the CSP, campaign service provider) → DCA or aggregator → carrier → handset. Alongside sit TCR, the vetting firms Aegis and WMC (whose scores set throughput), and Campaign Verify for political senders (providers say it's required on toll-free and short codes too since February 17, 2026).
| Party | What they control |
|---|---|
| Carriers | Filtering, throughput, fees, fines, suspension of any brand, number or provider |
| DCAs | Whether a campaign reaches the carriers at all |
| TCR | Brand identity checks and fees; it does not approve campaigns |
| Somos and RespOrgs | Toll-free numbers and text-enablement (a RespOrg, the number's managing company, can reject an enablement it didn't authorize) |
| iconectiv | Porting data, routing data and SHAKEN token administration |
| Analytics engines (Hiya, TNS, First Orion) | The call labels behind the major US wireless carriers |
| Meta | Everything on WhatsApp |
| RCS partner admission and Google-managed launches, which need a carrier-managed launch in that country first | |
| Apple | Whether iPhones render business RCS (iOS 18 or later, with its own user toggle) |
Abroad, the mobile carrier is the gatekeeper and sets the termination price (what it charges to deliver to its subscriber). Colombia requires a registered integrator for short codes, Nigeria routes international A2P through a single licensed gateway, and Brazil plans a central directory that binds short code brokers by contract (announced January 2026; I haven't confirmed it's live).
Then there are counterparties nobody signs with: grey routes (traffic pushed through consumer SIMs in "SIM boxes" to dodge termination fees) and SMS pumpers. A lot of the rules in this guide exist because of them.
Voice AI: real-time voice agents
A call passes through the caller's carrier, transit carriers, a communications platform or SIP trunk, the voice platform, one to three model vendors (often on different clouds) and the business's own systems. Twilio counts at least ten network traversals per turn of a chained agent, and each boundary can add another encode, decode and buffer. Moving the pieces closer together is mostly about latency: Telnyx places GPUs next to its telephony sites, and speech-to-speech models collapse three vendors into one.
| Party | What they control | What they earn |
|---|---|---|
| Communications platform or carrier | Numbers, caller-ID attestation, the audio stream | Per-minute and per-number fees, plus carrier surcharges |
| Terminating carriers and analytics engines (Hiya, TNS, First Orion) | Spam labels and blocking | Branded-calling services sold to businesses |
| Voice platform | The turn loop, logs, transfers | A per-minute platform fee |
| Model vendors | Accuracy, voices, latency, retirements | Per token, minute or character |
The analytics engines decide whether an outbound agent is heard at all. An outbound agent dials many short calls from few numbers, which is the pattern that earns a "Spam Likely" label, and a caller-ID company's 2026 survey found 86% of consumers say they won't answer an unidentified call (a vendor survey). Registration and branded calling are in Phone numbers, messaging and voice.
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.