The Platform PM
Primitive 04

Rules and policy

What logic decides outcomes, and who can change it?

Outcomes in regulated industries are decided by rules: who qualifies, what it costs, what gets approved, where it goes. The product question is less "what's the rule?" than "who owns it, and how fast can it change?"

What it is

Rules are the logic that decides outcomes. Some you set (your pricing), some your customers set (their approval thresholds), and some are imposed on you (a tax table, a carrier's content policy).

Who owns a rule matters more than how complex it is, because ownership decides who has to approve a change and how fast you can ship one.

Where you'll see it

IndustryRules that decide outcomes
PayrollTax tables, overtime rules, benefit eligibility
TelcoCarrier content policies, country sender rules, rate plans
PaymentsFee schedules, risk rules, card network rules
HealthcareCoverage and eligibility rules, prior authorization
InsuranceUnderwriting rules, rating tables

Questions to ask in week one

  1. Which rules decide the outcomes customers care about most?
  2. For each one, who owns it: us, the customer, a counterparty or a regulator?
  3. How is a rule change requested, approved and rolled out?
  4. Are rules stored as configuration, or written into code?
  5. When a rule decides against a customer, can we explain why?

The trap

Hardcoding a rule you don't own. If someone outside your company can change it, it belongs in configuration with a date on it (see effective dating), not in a code deploy.

In the Field Guides

AI agents: orchestration platforms

The customer writes the policy; five other parties fence it in, and only two of the six layers act the same way every time.

Who setsRules that decide outcomes (as of Oct 2026)
Lab usage policyHuman review for high-stakes decisions at OpenAI, Anthropic (plus disclosure) and Google; Meta, Mistral, xAI not checked
Defaults75 iterations and a 1-hour timeout on one AWS harness; 10 loop iterations in Google's kit (snippet)
ProtocolMCP: a human should be able to deny any tool call; tool annotations are untrusted hints
Policy engineA yes or no on each tool call, outside the model; AWS charges a tiny fraction of a cent per decision
CredentialsScopes, step-up when a scope is missing, network allow-lists
CustomerApproval thresholds, spend caps, the prompt

A guardrail is a probabilistic check on content; a control limits what the agent can do. Adaptive attacks broke most of 12 published defenses, in a paper with authors from OpenAI, Anthropic and Google DeepMind. Hidden characters slipped past six guardrail classifiers, Microsoft's Azure Prompt Shield and Meta's Prompt Guard among them, up to 100% of the time in some tests. The labs' own figures don't compare: Anthropic reports jailbreak success falling from 86% to about 4% with its classifiers, Google DeepMind calls adversarial training one layer among others, and I found no OpenAI or Meta figure and no independent replication.

People tire too. Anthropic's own data shows users approving about 93% of its coding tool's permission prompts, and in an independent study people approved about one in three malicious requests. Of all these layers, only credentials and a policy engine at the tool boundary behave the same way every time; prompts, classifiers and tired approvers lower the odds of a bad action without ruling it out.

AI coding agents: writing, reviewing and shipping code

An agent's rules live in three places, and only two of them enforce anything.

LayerExamplesEnforced by
Instruction filesAGENTS.md, CLAUDE.md, Cursor rules, Bugbot's own fileThe model's reading of them, so advisory
Harness permissionsPermission modes, allowlists, firewalls, sandboxes, tool lists per automationCode in the agent product
Repo-host policyBranch protection, rulesets, required checks, code ownersThe system of record

Cursor's Bugbot reads its configuration from the base branch, so a pull request can't change how Bugbot reviews it. That's the pattern I'd copy for any gate: keep it out of reach of the change under review.

Company policy sits above them: approved tools and plan tiers, which data may enter a prompt, dependency rules, reviewer independence, the line between prototype and production, and spending limits (GitHub sets budgets down to the user; Kiro keeps overage off by default for enterprises).

AI evals and observability: how teams measure AI quality

The rules that bite are internal: the rubric (what "good" means for each failure mode), pass thresholds, which checks block a merge, who signs off, and who may override or lower a threshold. I'd add who may change the judge, since that changes every score after it.

Practice varies on strictness. Anthropic wants regression evals near 100%; Hamel Husain, a consultant who writes widely on evals, says the acceptable rate is a product decision. Will Larson found the tests for an internal agent useful as "directional context" rather than a blocking gate, because runs were flaky; one fix is to block only when a case fails three tries out of three.

After incidents, labs wrote stricter rules for themselves. OpenAI said it would treat model behaviour issues "as launch-blocking like we do other safety risks". Anthropic added a per-model eval suite for every Claude Code system-prompt change, soak periods and gradual rollouts. xAI said it would publish Grok's system prompts and require review after an unauthorised prompt edit in May 2025.

Contracts set the outer rule. Lab terms disclaim the accuracy of outputs: Anthropic's commercial terms, for example, provide outputs "as is" and make the customer responsible "to evaluate whether Outputs are appropriate". Testing quality is contractually the builder's job.

AI governance and model risk

The policy stack runs from an AI policy and an acceptable-use policy down to tiering rules, an approval matrix and vendor standards; in a bank it sits next to the model risk policy. The operational pieces lag: in a 2026 survey by Schellman, which sells ISO 42001 audits, 44% of US companies with 500 or more staff had incident procedures specific to AI.

Tiering combines an outside screen with an inside score. US bank guidance measures materiality as exposure plus purpose, with inherent risk on top; E-23's inherent risk rating sets review intensity, documentation, the approver and monitoring; the EU tiers by use.

The binding rule often sits in a consent order or a contract instead of a statute. Massachusetts' settlement with Earnest, a student lender, requires an AI inventory, risk assessments, testing, documentation and an oversight team, which reads like a bank model risk program. Labs' usage policies require human review in credit, employment, insurance, housing and health, and customer contracts add their own conditions on top.

Agent identity and delegated access: how AI agents act for people

The rules sit in four places:

  • The company's identity provider: Entra ships Conditional Access templates for agents (block high-risk agents; separate rules for agents acting for someone and autonomous ones), Okta applies policy at the Cross App Access exchange, and Google uses org policies and IAM rules.
  • The API or MCP server: OAuth scopes are plain strings and coarse. Rich Authorization Requests can describe a single action, and Auth0 uses them with CIBA, but I found no MCP proposal that does. MCP's tool annotations mark tools as read-only or destructive, as hints only, so the real check is the server's own permission rules.
  • The agent's maker: Anthropic's browser agent has site permissions and asks before purchases, publishing or sharing personal data; OpenAI's agent asks before sending messages or payments; Supabase added read-only and project-scoped modes for its MCP server.
  • The website: robots.txt, Cloudflare's Content Signals, and per-crawler Allow, Charge or Block rules.

No binding rule says how narrow an agent's scope must be or that agent and user must be told apart in logs. After an incident, the yardsticks will be NIST IR 8587, the MCP spec's "must" rules, CISA's joint guidance and OWASP's 2026 list for agentic applications, which ranks identity and privilege abuse third.

B2B payments: cross-border payouts

The provider writes the routing policy, but six other parties set rules around it.

Who setsRules that decide outcomes (as of Oct 2026)
Sender's lawSanctions; money transmission licences; travel rule from $3,000; Canadian EFT reports from CAD 10,000; tax forms and withholding
Destination lawSalary currency and pay dates; purpose codes; FX controls; tax receipts (CFDI, FIRA)
Partner bankCountry and industry lists, KYB depth, reserves; can change without notice
Local partnerIts own KYC, purpose-code checks, limits
Rail or networkCaps; Fast Funds within 30 minutes; issuer limits on cross-border pushes
ClientCharge option (OUR, SHA, BEN), funding time, pay dates
ProviderRouting, quote window and size, cut-offs, deposits, late fees

When rules conflict. Canada's Foreign Extraterritorial Measures (United States) Order, 1992, bars Canadian corporations from complying with US Cuba-embargo measures, and the EU Blocking Statute does the same. A Canadian payout firm still screens OFAC's lists in practice, because its dollars clear through US banks, but must not apply the Cuba measures. Going with "the stricter rule wins" is a business choice; no law says so.

Picture a Toronto payroll provider whose US bank asks it to stop paying staff at a Cuban-owned company's office in Spain. Following the request could breach Canadian law; refusing could cost it the bank.

The tests differ. OFAC's March 31, 2026 guidance calls its 50% ownership rule a floor and points at control; Canada already asks whether property is "owned, held or controlled, directly or indirectly".

Routing policy is where it all lands: a rail per payee by amount, speed, cost and destination limits, a quote window and size cap, and who absorbs slippage.

B2B payments: order to cash

Rules come in layers, and every party on the sale sets some of them.

Who setsRules that decide outcomes (as of Oct 2026)
SellerCredit policy and approval authority, terms by segment, hold rules, dunning cadence, late fees, auto-write-off thresholds for small deductions (CRF median under $25 in 2012), accepted rails, card fees
PayerPayment-run cadence, portal and remittance formats, deductions, compliance manuals, often its own terms (Net 60 to 90 or longer at large payers)
Insurers and lendersMaximum terms and per-customer limits; eligibility, concentration and dilution limits
Card networksVisa US: surcharge credit only, at most 3% and no more than the cost of acceptance, 30 days' notice, disclosure at entry and checkout. Mastercard: reported as the lower of 4% and the seller's discount rate
Rail operatorsNacha return windows and thresholds (0.5% unauthorized, 3% administrative, 15% overall); Payments Canada's Rule H1 for PADs
LawSurcharges: banned in Connecticut and Massachusetts, capped at 2% in Colorado and 2.4% in Canada (since October 2022), banned on Quebec consumer sales. Terms: produce is due in 10 days unless agreed in writing (PACA); Ontario construction owners pay within 28 days of a proper invoice. Interest: US late charges need prior agreement; in Canada, a monthly rate without its yearly equivalent caps interest at 5% a year

Terms are a price. "2/10 net 30" means giving up 2% to be paid 20 days early, about 37% a year; a longer term is a discount nobody books. Say a packaging supplier offers 2/10 net 30 on a $10,000 invoice: a customer that pays on day 10 sends $9,800, so the supplier paid $200 to get its cash 20 days sooner. The seller sets its terms only until a large payer, an insurer or a lender sets them instead, which is partly why sales and credit pull against each other.

The other big rule is who starts the payment. In a pull (ACH debit, PAD, card) the seller does, and carries authorization and return risk; in a push (ACH credit, wire, RTP, FedNow, check) the payer carries fraud and misdirection risk. On card fees, the seller decides, the network caps, and state or provincial law can forbid.

B2B payments: procure to pay

The buyer writes most of the rules, within limits others set.

Who setsRules that decide outcomes (as of Oct 2026)
BuyerApproval matrix and delegation of authority; match tolerances; terms policy; payment-run calendar; rail policy ("ACH by default, card under a set amount"); bank-change verification; segregation of duties
LawUS Prompt Payment Act for federal buyers; Canada's federal construction prompt payment; 1099 and backup withholding; T4A; input tax credit documentation
Rail operatorsNacha's 2026 fraud monitoring for originators; the narrow ACH reversal grounds
BanksSecurity procedures under UCC 4A, payment limits, Positive Pay terms
Card networks and issuersInterchange tiers, surcharge rules, rebate schedules in the card contract
SuppliersWhich rails they accept, whether they surcharge, the terms on the invoice

Segregation of duties is the core buyer rule: vendor master, invoice entry, approval and payment release sit with different people, and auditors test it under SOX 404 at US public companies. Say one clerk can both add a new vendor and release the payment run: they could set up a fake supplier and pay it, and nobody else would see it until the audit.

Law steps in only where governments chose to protect suppliers. A US federal agency pays against "a proper and valid invoice", owes interest if it's late (4.75% for July to December 2026) and may take a discount only inside the window. Canada's federal construction law, in force since December 9, 2023, makes the Crown pay within 28 days of a proper invoice and gives each subcontractor tier 7 days to pay the next. I found no general private-sector prompt-payment law in either country: terms are contractual. So one supplier can be paid in 10 days by one customer and 90 by another, and both are within the rules.

B2B payments: real-time payments and open banking

The stack, from the top: law (Article 4A of the Uniform Commercial Code for US business transfers, Reg E for consumers, the EU's Instant Payments Regulation), regulators' rules, scheme rules (RTP's Operating Rules of October 4, 2026, FedNow's Operating Circular 8 of April 1, 2026, the European Payments Council's rulebooks), each bank's limits and holds, then the customer contract.

Banks set tighter limits than the networks: FedNow's $100,000 default sits under its $10 million cap, and UK banks set theirs below the £1 million Faster Payments cap. Rules also decide who may charge whom. Brazilian banks can't charge individuals to send a Pix, EU banks can't price an instant payment above a standard transfer, and US open banking has no rule on data fees while the courts hold the 2024 rule.

FedNow gives banks fraud controls they set themselves: a negative list of accounts at other banks, activity limits by account and time window, a Network Intelligence API (launched April 28, 2026) showing what FedNow has seen of a receiving account, and fraud reporting. Checks that would run past the 20-second clock are skipped.

B2B payments: spend management

The company writes the spend policy; five other parties fence it in.

Who setsRules that decide outcomes (as of Oct 2026)
Card networkProduct rules (a purchasing card must be credit); category codes; interchange qualification; dispute windows; zero-liability exclusions; payments by AI agents
Sponsor bankCredit box, prohibited industries, KYB depth, collateral, final say on each card and transaction
Program managerUnderwriting inside the bank's policy, limits, rewards tiers, settlement terms
CompanySpend policy (categories, limits) and expense policy (receipts, per diem caps, approvals)
Tax lawUS accountable plans and substantiation; per diem and mileage rates; 50% meals; Canada's reasonable allowances and tax credit documents
Employment lawState reimbursement laws; the FLSA floor; the Canada Labour Code

The accountable plan. US reimbursements are tax-free only if the plan passes three tests: business connection, substantiation (proof of what was spent and why) and return of any excess (Treas. Reg. 1.62-2). Paying whether or not the employee spends fails the business-connection test, so a flat car allowance with no log is wages. The safe harbor: advances at most 30 days ahead, proof within 60 days, excess back within 120. Fail, and the amount goes on the W-2 with withholding and payroll taxes. Most employees can no longer deduct unreimbursed expenses (made permanent in 2025), so the plan is the only tax-free route.

Law versus policy. US law wants a receipt for lodging and anything else from $75; most policies ask for more. Illinois lets a written policy cut off late claims; California's rights can't be waived. Travel and client meals are 50% deductible, convenience meals 0% since 2026, so the GL needs separate codes. In Canada, a kilometre allowance paid alongside a fuel card for the same use becomes fully taxable, and that comes from how the company designed the plan.

Canadian sales tax. The evidence for an input tax credit grows with the amount: under $100, supplier, date and total; from $100, the tax and the supplier's GST/HST number; from $500, also the buyer's name, a description and terms. Reimbursements can use a factor (4/104 for GST, 12/112 to 14/114 for HST) if 90% of the total bore one tax. Meals get half the credit: on a $226 team dinner in Toronto, $26 of it 13% HST, the company claims back $13. The company-card shortcut needs the cardholder to be "solely or jointly and severally liable", so a pure corporate-liability program may not qualify.

The company writes the policy, but the bank, the network, the tax code and labor law can each overrule it, and only the card controls act at the moment of purchase.

B2B payments: stablecoins and blockchain payments

Public chains have no rulebook like Nacha's for ACH: no return codes, no debit authorization standard, no dispute window. The rules that apply sit in contracts.

  • Issuer terms: when the issuer may freeze, who may redeem, and that transfers are final. Circle says it "does not have the ability to reverse or recall any transaction".
  • Orchestrator and ramp terms: Bridge's terms for customers outside the US and EU let it freeze funds "without prior notice" and reverse transactions it suspects, and limit its liability to gross negligence, intentional misconduct or fraud.
  • Network rules: Circle Payments Network's rules say participants act "at their own risk", and card networks set the rules for banks that settle in coins.
  • Chains and banks: a chain's validators set its fees, and for tokenized deposits the bank's account terms apply.

The company's own policy fills the rest. I'd write down which coins, issuers and chains it accepts, how many confirmations count as final on each chain, the screening threshold for inbound addresses, how fast coins are converted, and the most it will hold. The coin list needs dates, because the US rules on permitted issuers start in January 2027 and tighten in July 2028.

Card issuing: launching a card program

Rules reach a program in three layers: network rules, then the bank's policy, then your own. The bank's is usually the one you work to day to day, and it often repeats the network's rules with stricter numbers.

What Visa writes onto the bank, which the bank passes to you:

  • The bank must control approval of cardholder applications, and can't approve an agent just because the contract limits the bank's liability.
  • For prepaid, the bank needs Visa's approval before launch, must hold all card funds and agent reserves in an account it controls, and keeps reserves for program losses. It completes a prepaid risk self-assessment at entry and every year, and Visa can review the program at the bank's cost.
  • An agent's contract must give Visa audit rights, and the agent must hand cardholder data to Visa or regulators within seven business days when asked.
  • Visa can block a BIN used for a purpose it didn't approve, and can refuse an agent registration "with or without cause".

The bank adds prohibited program types, how deep identity checks go, approved vendors, reserves, minimums, and approval of every product change. Chime's filing puts the balance plainly: the banks approve each program, and Chime designs it, negotiates terms and picks vendors, "subject to the requirements of the banks and applicable card networks".

Some policy choices are yours, and they're product decisions. Reg E lets an issuer withhold error resolution and limited liability on a prepaid account until it has finished verifying the customer's identity, so when you run identity checks decides which protections a cardholder has on day one. The checks themselves are in Identity and trust.

When a requirement lands on your desk, I'd ask whether it's a law, a network rule or the bank's risk appetite. That tells you who can change it, and usually only the last is open to negotiation.

Cloud security (CNAPP): posture, workloads and permissions

Policy comes from CIS benchmarks (prescriptive technical checks; AWS added version 5.0 of its AWS benchmark to Security Hub in October 2025, with 40 automated controls), the providers' own benchmarks, mappings to PCI or FedRAMP, custom rules, and preventive guardrails such as AWS service control policies and resource control policies, Azure Policy and Google organization policies. Detection dominates prevention in practice: in Datadog's 2025 data, 40% of AWS Organizations users had service control policies and 6% had resource control policies.

The metadata fix is my favourite example of a cheap rule left unapplied. Version 2 of the EC2 instance metadata service needs a session token, which blocks the request-forgery trick behind the Capital One breach. Datadog found it enforced on 49% of EC2 instances in 2025, up from 7% in 2022, while 82% of instances had used only version 2 in the previous two weeks and could have enforced it without breaking anything.

Regulators and insurers write rules too. New York now requires MFA for any individual accessing a covered firm's systems, and insurers increasingly reject text-message MFA and ask for phishing-resistant keys for administrators, endpoint detection and backups.

Identity and trust: KYB, KYC and ongoing monitoring

Law is only one of the kinds of rules that shape every decision.

  • Law: the bank rules on customer identification and beneficial ownership in the US, Canada's PCMLTFR, the UK's money laundering regulations, the EU's AMLR from July 2027, and the INFORM Consumers Act for marketplaces.
  • Network rules: Visa requires a payment facilitator (a payfac, a company that signs up merchants under its own card acceptance account) to check every applicant against its list of terminated merchants, and bans some industries outright unless the payfac is registered for high risk. Mastercard requires a check against MATCH, its list of terminated merchants, before onboarding.
  • Your bank's policy: a prohibited and restricted business list, how deep KYB goes, which vendors you may use, how fast alerts must be worked, how often customers get refreshed.
  • Your own policy: risk appetite and auto-approval thresholds.

Many hard calls aren't data questions at all. What counts as "control"? Do you onboard a company that's delinquent but can cure it? Is a registered agent's address, a virtual office or a UPS store acceptable? How far up a foreign ownership chain do you go? Write those calls down as policy, separate from the data, because that's where two platforms with the same vendors end up differing.

Visa's high-brand-risk list is a set of merchant category codes plus activities such as cyberlockers, skill games like daily fantasy sports, and buying crypto. A marketplace may not sign those merchants at all; only payfacs registered for high risk can.

Thresholds trade false accepts against false rejects. You tune them; your sponsor bank approves them.

Login and access for apps: authentication and authorization

Permission models come in three families, usually mixed. Roles are simple and auditable until every customer wants a custom one ("role explosion"). Attribute rules (department, region, time) are flexible and hard to explain to an auditor. Relationships handle sharing, folders and nested teams, and need a store and modelling skill.

Where the check runs matters as much as the model. The gateway confirms a token is valid and has the right scope, the service decides whether this user may do this to this record, row-level security in the database filters by tenant as a last line of defence, and hiding a button controls nothing. OWASP's advice: deny by default, implement the check once and reuse it, enforce record ownership, and log failures.

OAuth scopes record consent for an app. A token with "files:read" says what the app may do for the user; the app still has to check what the user may do, and agents widen that gap. The other rules are customers' settings (require SSO for our domain, MFA for everyone, sessions under eight hours, these IP ranges only), which belong in the product as per-organization policy that admins set themselves.

Models and inference: what AI runs on

Access is rationed by rules the provider writes and can change:

Anthropic

Tiers
Evaluation, Start, Build, Scale, Custom
Monthly spend cap
$500 / $1,000 / $200,000
Example limit at the top self-serve tier
Opus 5.5: 10,000 requests and 10 million input tokens a minute; Fable 5.x: 4,000 and 4 million

OpenAI

Tiers
Free, Build, Launch, Grow (after $5, $100 and $500 of purchases)
Monthly spend cap
$100 / $500 / $5,000 / $200,000
Example limit at the top self-serve tier
gpt-6 Astra, Sol and Terra: 15,000 requests and 40 million tokens a minute

Google (Gemini API)

Tiers
Free, Tiers 1-3 (by spend and account age)
Monthly spend cap
$250 to $100,000 or more, plus a 10-minute spend limit
Example limit at the top self-serve tier
Per project, per model

The top model gets the tightest limit, and Anthropic calls its limits "maximum allowed usage, not guaranteed minimums". New organizations start lower "to prevent fraud and abuse", and fast ramps hit acceleration limits. Gateways add the customer's own policy (route by price or speed, cap the price per token, use only zero-retention hosts, set budgets per team), and the labs' usage policies restrict high-stakes uses without human review.

Observability: metrics, logs and traces

Most rules in observability are ones the customer writes: what to sample, redact and index, for how long, when to alert and who gets paged. They live in collector and alerting configuration, and they decide both the bill and what an investigator will find.

SLOs matter most for signal. Google's SRE Workbook recommends paging on burn rate over two windows, so a short spike pages nobody and a real outage pages fast:

Page

Long window
1 hour
Short window
5 minutes
Burn rate
14.4x
Budget spent when it fires
2%

Page

Long window
6 hours
Short window
30 minutes
Burn rate
6x
Budget spent when it fires
5%

Ticket

Long window
3 days
Short window
6 hours
Burn rate
1x
Budget spent when it fires
10%

That's for a 99.9% target over 30 days. At a burn rate of 1,000, the whole month's budget is gone in about 43 minutes. The math breaks for low-traffic services: at ten requests an hour, one failure spends about 14% of the monthly budget, so teams add synthetic traffic, group small services or lower the target. An error-budget policy closes the loop: while the SLO is met, releases continue, and when the budget is spent, they pause.

Google's on-call guidance sets a bar few teams meet: every page actionable, at most two incidents in a 12-hour shift, and on-call capped at a quarter of an engineer's time. Vendors add their own rules (cardinality limits, rate limits, index quotas), which are product decisions about how hard to stop a customer's mistake.

Phone numbers, messaging and voice

Most rules that decide outcomes here are written by carriers and industry bodies, and they change by memo.

The US baseline is CTIA's Messaging Principles and Best Practices (May 2023, still the latest):

  • Consent scales with content: implied for conversational, express for informational, express written for promotional messages.
  • One opt-in per campaign, not transferable; no rented or bought lists.
  • Opt-out works in plain language, not only STOP, with one confirmation message.
  • No public URL shorteners and no snowshoeing (spreading traffic across many numbers to dodge filters).

CTIA's short code handbook adds HELP and STOP wording and the SHAFT rules (sex, hate, alcohol, firearms, tobacco): legal SHAFT content needs an age gate.

Carriers set throughput by rule, from the vet score. A retailer's campaign with a vet score of 40 gets 240 texts a minute at AT&T, so a 100,000-text flash sale takes about seven hours to go out; with a score above 75 it would take under half an hour.

CarrierHow 10DLC throughput is set
AT&TPer campaign per minute, by vet score: 4,500 SMS (75-100), 2,400 (50-74), 240 (1-49); 75 for low-volume mixed; 15 for sole proprietors
T-MobilePer brand per day, shared by all its campaigns: 200,000 (75-100), 40,000 (50-74), 10,000 (25-49), 2,000 (1-24); 1,000 for sole proprietors
VerizonNo published numbers; relies on filtering

Carriers don't publish these tiers; I took them from provider documentation in October 2026. What moves you up a tier is the vet score, and a standard vet costs around $40 at TCR.

Rich channels have their own rulebooks. Since April 9, 2025, Meta approves a WhatsApp "utility" template that reads like marketing as marketing. One line like "upgrade your seat for $29" at the end of an airline's boarding-pass template can turn a cheap utility message into a marketing one. Google holds each RCS agent to the use case declared at launch: OTP, transactional, promotional or multi-use. Abroad, countries add content categories and quiet hours: India's P/S/T/G header suffixes, the UAE's "AD-" prefix with no promotions from 21:00 to 07:00 (according to provider guides), and Colombia's advertising window of weekdays 07:00-19:00 and Saturdays 08:00-15:00 under Ley 2300.

Anything an outside party can change belongs in configuration with an effective date.

Voice AI: real-time voice agents

Turn-taking settings are policy decisions: how long a silence ends a turn, how eager the model is to answer, how long a sound has to last before it counts as an interruption. A longer timeout while the caller reads out digits protects accuracy and costs speed.

Guardrails behave differently by architecture. A chained stack can check the text before it's spoken. On a speech-to-speech stream the check runs alongside the speech, and an open issue on OpenAI's agent toolkit reports the agent kept talking after a guardrail tripped. Required disclosures and consent prompts are fixed text inside a probabilistic stream, and a chained or hybrid stack makes it easier to guarantee they're said word for word.

Above the builder's settings sit laws, card network rules, carrier analytics and the platform's and model provider's usage policies. Retell's terms, for example, require outbound agents to identify themselves and ban calls to emergency lines.

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.