State and lifecycle
What states exist, and what moves an entity between them?
Every entity moves through states, and each move has a trigger. Map the states and the triggers and you have the skeleton of the business: where the money moves, where the risk sits and where customers get stuck.
What it is
A lifecycle is the list of states an entity can be in and the events that move it between them. Money, risk and regulation usually attach to the transitions, not to the states.
If you can draw the lifecycle on one page, you can usually explain where the business makes and loses money.
Where you'll see it
| Industry | A lifecycle that matters |
|---|---|
| Payments | Authorized → captured → settled, with refunds and chargebacks branching off |
| Healthcare | Claim submitted → adjudicated → paid or denied → appealed |
| Telco | Number available → active → porting out → released → aging |
| Payroll | Hired → active → on leave → terminated → final pay |
| Insurance | Quoted → bound → in force → renewed or lapsed |
Questions to ask in week one
- What are the states, and which ones are final?
- Who or what triggers each transition: the customer, an employee, a counterparty, a clock?
- Which transitions can be undone, and how?
- How long does each state normally last, and what happens when it lasts too long?
- Which states does the customer see, and which are internal?
The trap
Treating the lifecycle as one status field. In practice several lifecycles run in parallel on the same entity (a phone number can be live for calls while its messaging registration is still in review), and the bugs live where they disagree.
In the Field Guides
AI agents: orchestration platforms
An agent has four state machines, and they end at different times.
- Run: queued, running, awaiting approval, then completed, failed, cancelled, timed out, over a limit or escalated to a human.
- Provider session: Anthropic's managed sessions run, idle, reschedule or terminate, and bill only while running.
- Release: draft, evaluated, canary, live, superseded, by house convention.
- Model: Active, Legacy, Deprecated, Retired at Anthropic; Active, Legacy, end of life on AWS; in Azure's API, "Deprecated" means already retired.
State moved to the labs in 2025 and 2026, each its own way:
OpenAI
- Stateful agent API
- Responses and Conversations
- Stored state kept
- Responses 30 days; conversations and files until deleted
- Long runs
- Background mode
- Stateful agent API
- Interactions (GA June 2026)
- Stored state kept
- 55 days paid, 1 day free
- Long runs
- Background flag; no explicit caching yet
Anthropic
- Stateful agent API
- Managed Agents
- Stored state kept
- Session transcripts until deleted
- Long runs
- Sessions billed by the running hour, a few cents an hour
OpenAI's older Assistants API shut down on August 26, 2026, a year after its deprecation notice. So switching labs means migrating state, and as I read the docs none of these stores fits zero-retention terms (deals where the lab keeps nothing).
Resume is replay. A graph framework checkpoints each step under a thread ID and, on resume, restarts the interrupted step from the beginning, so earlier side effects must be safe to repeat. Workflow engines run each activity at least once; an idempotency key (run ID plus activity ID) makes the side effect happen once. Runtimes also cap the clock, as in the vendor onboarding above.
Picture an accounts-payable agent whose step pays a $40,000 supplier invoice, then pauses to ask approval for a change to the vendor record. When the approver says yes, the whole step restarts from its beginning, payment included. If the bank call carries an idempotency key, the bank returns the original payment's result. Without one, the supplier gets paid twice. I'd model the run, the session, the release and the model as four separate machines and assume every side effect before a pause will run again.
AI coding agents: writing, reviewing and shipping code
Several state machines run at once.
- Task: queued, running, waiting for input, pull request open, changes requested, merged or abandoned. On GitHub, moving a draft to "ready for review" is a human step the agent can't take.
- Pull request: opened, CI running, reviewed, merged or closed. Agent pull requests stall waiting for review and get abandoned, so I'd count abandoned ones as waste.
- Session: capped by the clock (59 minutes on Copilot), which forces work that can resume.
- Billing: trial, active, paused at a limit, overage on, downgraded. When GitHub credits run out and overage is off, Copilot pauses until the next cycle. Anthropic's and OpenAI's individual plans add rolling five-hour and weekly windows on top.
AI evals and observability: how teams measure AI quality
A change moves through proposed, evaluated, rolled out, released and monitored, with blocked and rolled back as exits. Test cases have states of their own: candidate, active, quarantined for flakiness, retired when stale or saturated. In Anthropic's guidance, capability evals start with low pass rates and give the team something to climb, then "graduate" into regression evals, which should sit near 100%, once they pass reliably.
Models carry states set by the provider (active, legacy, deprecated, retired), with slightly different meanings at each lab and cloud. Judges move between calibrated, drifting and recalibrated, and vendors between free, paid, acquired and sunset. OpenAI's guide says its hosted Evals product goes read-only on October 31 and shuts down on November 30, 2026.
AI governance and model risk
A use case moves through discovered, proposed, in assessment, approved with conditions, live, under review or suspended, changed and re-validated, retired and archived. The April 2026 US guidance is silent on retirement; Canada's E-23 has an explicit decommissioning stage. Vendors run their own states on their own clocks (active, legacy, deprecated, retired, meaning slightly different things at each lab and cloud), and those states drive yours.
Laws have lifecycles too: Colorado's AI Act was enacted in 2024, delayed, paused by a court stipulation, then repealed and replaced in May 2026. Most AI lawsuits in 2026 sit between discovery and class certification, and an ISO 42001 certificate lasts three years, with a surveillance audit each year.
Agent identity and delegated access: how AI agents act for people
An agent goes registered, given an owner, credentialed, active, restricted, disabled and deleted. The state most systems lack is orphaned, an agent whose owner has left, which is why Entra now moves sponsorship automatically. A web agent goes submitted to a gatekeeper's registry, accepted or rejected, verified, and sometimes delisted.
Timers define behaviour: Google's agent certificates last 24 hours, Visa accepts an agent's signature for 8 minutes, and NIST wants access tokens to last an hour or less.
Failures cluster in a few places. Tokens outlive their purpose: Klue's breach began with a credential made for an abandoned prototype and never switched off, and GitGuardian, which sells secret scanning, found 64% of the valid secrets it detected in 2022 still active four years later. Approvals decay into "always allow". And revocation is coarse, so whole integrations get switched off because nobody can pick out the bad tokens.
B2B payments: cross-border payouts
A payout moves through three separate state machines, and each reports "done" at a different moment.
- Money: quoted, funded, converted, sent, credited or returned (in local currency, at a new rate).
- Compliance: screening hit, request for information, purpose-code hold, then released, rejected or blocked (reported to OFAC within 10 business days).
- Payroll: draft, cut-off, invoiced, funded, calculated, paid, receipted (the CFDI), remitted, reconciled.
| Who says "paid" | What they mean |
|---|---|
| Customer | Funded the provider or paid the pre-funding invoice |
| Provider | Instructed the partner |
| Swift gpi | Beneficiary bank confirmed the credit |
| Card network | Issuer approved; usable within 30 minutes at Fast Funds issuers |
| SPEI | Banxico settled; the bank credits within 30 seconds unless approved to delay |
| Worker | Local currency is usable |
| Tax authority | CFDI issued (Mexico), or the bank's FIRA exists (India) |
Most support tickets come from the gaps between these. Swift says 75% of payments reach beneficiary banks within 10 minutes, but reaching the bank doesn't mean the worker has been credited. A card push outside Fast Funds can take days, and a large SPEI payment can be settled and still not credited. I'd only promise a state you can prove; on many rails "sent" is the last thing you know for sure, while the worker only cares about "usable".
B2B payments: order to cash
Six lifecycles run on the same sale, and they disagree by design.
- The account: applied, approved, active, on hold, cash in advance, closed.
- The order: entered, credit-held, released, shipped, delivered, invoiced.
- The invoice: issued, accepted, approved, scheduled (payer side), paid (in the payer's eyes), applied (in the seller's), closed.
- The payment: initiated, received, settled, funds final.
- Cash application: unidentified, unapplied (payer known, invoices not), partially applied, applied, on account.
- Collections: current, past due by aging bucket, dunning, promise to pay, placed, written off, recovered.
"Settled" and "final" mean different things on every rail:
| Rail | Settled when | Final when (as of Oct 2026) |
|---|---|---|
| Check | Provisional credit on deposit | After the paying bank's midnight deadline, but altered or counterfeit check claims can surface months later |
| ACH credit | Settlement date | Effectively on settlement; a return request or R17 is still possible |
| ACH debit | Settlement date | After 2 banking days on a business account; 60 days if it turns out to be a consumer's |
| Card | Funded net, 1 to 2 days later | After the chargeback window: 120 days for Visa card-absent fraud |
| Wire, RTP, FedNow | On settlement | On settlement |
| Canadian business PAD | Next day | After 10 business days; 90 calendar days with no agreement |
Bankruptcy can undo even a final payment: payments received in the 90 days before a customer files (a year for insiders) can be clawed back as preferences, unless a defense such as ordinary course of business applies. Because a paid invoice can reopen, every reversal has to reverse the application, re-age the invoice and restart dunning, or your customer ends up chasing the wrong people.
B2B payments: procure to pay
Four lifecycles run side by side, and they disagree by design:
- Requisition and PO: requested, approved, open, partially received, closed.
- Invoice: received, captured, matched or in exception, approved, posted, scheduled, paid, cleared.
- Payment: proposed, approved, released, sent, then settled, returned or cleared.
- Virtual card: issued, then charged by the supplier (or expired uncharged), then settled.
The word that causes the most trouble is "paid":
| Who says "paid" | What they mean |
|---|---|
| Buyer's AP | Released in the payment run, or the check printed |
| Buyer's bank | The ACH settled, or the check was presented and honored |
| Card issuer | The supplier charged the virtual card |
| Supplier | Cash received and applied to the invoice |
Discount and late-fee fights live in these gaps. A buyer that "paid" on day 10 by mailing a check and a supplier that got it on day 14 will disagree about the 2% in good faith.
I'd model "released", "settled" and "cleared" as separate states. A check can show as paid in the ERP for weeks before it's cashed, and meanwhile be stolen, altered or go stale. A virtual card can sit uncharged while the buyer thinks the invoice is closed.
B2B payments: real-time payments and open banking
A business instant payment has more states than "sent" and "received": requested, approved, accepted, settled and final, available, applied, reconciled. A held payment breaks the habit of treating "settled" as "usable".
After settlement the states are return requested, frozen (Brazil lets a receiving bank hold a suspicious incoming Pix for up to 72 hours), returned voluntarily and reimbursed by rule. I'd model each as a state with its own timestamps, never as an edit to the original payment.
Clocks run everywhere: 20 seconds on FedNow, 10 seconds by law in the EU, under 4.6 seconds inside Brazil's settlement system for 99% of Pix, 12 months for most US data consents.
B2B payments: spend management
Four machines run on one purchase, and they end at different times:
- Card: unactivated, active (a virtual card counts once viewed), suspended, terminated.
- Authorization: approved or declined, held, increased or partly reversed, then cleared or expired (Visa allows 10 days online, 30 for lodging and car rental).
- Expense: draft, submitted, returned or approved, audited, exported, paid, archived.
- The employee's obligation: an advance or personal charge, then repaid, or treated as wages after the IRS's 120 days.
A transaction can outlive its card: refunds and late captures land after cancellation. A salesperson leaves on a Friday and her card is canceled; on Monday the hotel's final bill and a software renewal still arrive against it. And one issuer processor allows recurring authorizations on expired cards unless they're canceled. When someone leaves, cancel rather than reissue, or Visa Account Updater sends the new number to every merchant that stored the old one.
The word that causes the most trouble is "done":
| Who says "done" | What they mean |
|---|---|
| Employee | The money is in my account |
| Issuer | Cleared and posted |
| Finance | Receipt matched, coded and approved |
| Payroll and tax | Substantiated within 60 days, excess returned within 120 |
| Auditor | Receipt kept 3 to 4 years (US) or 6 (Canada) |
Legal clocks hang on these states. Illinois' 30 days to submit and the IRS's 60 days run from the expense date. Payment deadlines run from the claim: 30 days in Iowa, New Hampshire and federally regulated Canadian workplaces. In New York, an unpaid agreed reimbursement is a misdemeanor 30 days after it's due. Because the four machines end at different times, most of the bugs sit where one has ended and another hasn't.
B2B payments: stablecoins and blockchain payments
A stablecoin payment adds states that ACH doesn't have: broadcast, confirmed at some number of blocks, final, screened, held for review, converted and paid out. Everyone says "paid" at a different moment: the orchestrator when coins leave its wallet, the chain at the confirmation depth the receiver chose, the off-ramp when it starts the payout, the supplier when pesos are usable. On the chain that takes seconds; cash in a bank account comes later.
Then there are legal states: frozen by the issuer, blocked under sanctions, redemption deferred during stress, or stuck in an insolvency. Redemption normally takes up to two business days under the GENIUS Act; in an issuer bankruptcy, law firms expect at least two weeks, because holders must ask the court before reserves are paid out.
Chains stall too. Base, Coinbase's network, halted twice in June 2026, for 116 and 20 minutes, when its single sequencer (the one machine that orders transactions) failed; payments waited and no funds were lost. I'd only show customers a state I can prove, and on many routes the last thing the payer knows for sure is "sent".
Card issuing: launching a card program
Card states get all the attention: ordered, personalized, shipped, unactivated, active, suspended, terminated or expired, reissued. Program states matter as much, and few systems model them: contracted, go-live, launched, under order, winding down.
A program can be "launched" and frozen at the same time. When the Fed put Evolve under a cease-and-desist in June 2024, the bank couldn't take a new fintech partner, product or program manager without its regulators' prior approval. Programs on it stayed live but couldn't change, and their roadmaps waited on someone else's examiner.
The bank has a state too. Its Durbin status is measured on assets at the end of each calendar year, and a bank that crosses $10 billion must apply the cap by July 1 of the next year. That gives a program about six months to reprice or move once the year-end numbers are in.
Network rules move card states too: Visa's longer maximum expiry for prepaid cards, from July 25, 2026, changes reissue planning on long-lived payroll cards.
A reissue during a migration changes the card number, and cards stored at merchants break unless tokens or the network's account updater carry the new number over.
Cloud security (CNAPP): posture, workloads and permissions
A finding at AWS carries three separate states at once: a workflow status (new, notified, suppressed, resolved), a record state (active or archived) and a compliance result (passed, failed, warning or not available). They change independently. The detail I'd remember: a resolved finding goes back to new if the check fails again, while a suppressed one stays suppressed. A suppression is sticky and a fix is fragile.
Severity is a state too. A finding becomes critical when something next to it changes: a firewall rule opens, a role gains a permission, a bucket starts holding customer data. Graph-based tools recompute as the inputs change.
Credentials have a lifecycle that nobody closes. In Datadog's October 2025 data, 59% of AWS IAM users had an active access key more than a year old, and in the 2024 Snowflake breaches some stolen credentials dated back to 2020 and still worked. "Decommissioned" is the state most often missing.
Regulators add states of their own: an incident becomes "material" at the SEC or "major" under DORA, and a cloud provider becomes "designated" as critical in the EU and the UK.
Identity and trust: KYB, KYC and ongoing monitoring
One customer has a state in the registry, on your platform and in your alert queue, and these drift apart unless you connect them.
- The registry's: formed, active and in good standing, delinquent (a missed annual report or franchise tax), administratively dissolved or revoked, reinstated within a window each state sets, or gone.
- Your platform's: started, submitted, auto-verified or pending review or documents requested, approved (with limits) or declined, active, under review, restricted, offboarded. Add states nobody plans for: frozen for sanctions, and "keep open", when law enforcement asks you not to close an account.
- The alerts': new, in triage, escalated, case open, report filed, closed with no action.
A person's verification has a lifecycle too. Documents expire, addresses change, and in the EU from July 2027 customer information must be refreshed at least yearly for high-risk customers and every five years for the rest. Assurance also comes in levels.
If the landscaping company misses its Delaware annual report in March, the registry moves it to delinquent while your platform still says "active". Nothing is wrong yet, but if a fraudster later reinstates the company and changes its officers, you'd want to have seen both events, so subscribe to the registry's state changes where you can.
Watch one more state: "closed automatically". A US regulator's April 2026 order against a payments-heavy bank criticized alerts that closed themselves. No alert should close without a human or a validated model giving a reason.
Login and access for apps: authentication and authorization
An account goes invited, active, suspended, deactivated and deleted, sometimes reactivated. A credential is enrolled, verified and revoked; a session is created, refreshed, then expires or is revoked; an enterprise connection goes draft, verified, active and disabled. SCIM moves users through create, update and deactivate, and real directories rarely delete.
The state most systems lack is "decommissioned": tokens, keys and test accounts nobody switched off. Microsoft's Midnight Blizzard breach in 2023-24 started with a password spray on a legacy test account with no MFA, then used an old test OAuth app with high privileges.
Timers define behaviour. NIST's 2025 guidelines suggest ending sessions at the middle assurance level after 24 hours, or one hour idle, and disabling an authenticator after at most 100 failed attempts in a row. Access tokens live minutes and refresh tokens days to months.
Models and inference: what AI runs on
A request goes from admitted to queued, prefilling, decoding and streaming, and ends complete, incomplete or failed. A model goes from preview to generally available, legacy, deprecated (still works, retirement date set) and retired. A prompt cache is written, stays warm while used, and dies when it expires or when the prefix, tools or reasoning settings change. An account climbs tiers with spend history and pauses at its spend cap until the next month. Capacity goes from contracted to built, energized and allocated.
Retirement clocks differ by vendor and are getting shorter at some:
| Vendor | Typical notice before retirement | Notes |
|---|---|---|
| Anthropic | At least 60 days; 60-62 days for every 2026 notice | Down from 181-189 days in 2025; Opus 4.1 lived exactly 12 months |
| OpenAI | At least 6 months for generally available models, 3 for specialized ones, about 2 weeks for previews | A restricted cyber model got 20 days in September 2026 |
| Google (Gemini API) | Dates are "earliest possible"; previews retire 2-4 months after release | Some previews run in production for months |
| Microsoft Foundry | 18-month lifecycle for most generally available models, at least 60 days' notice | Standard deployments auto-upgrade; "retirement dates aren't extendable" |
| AWS Bedrock | "End of life no sooner than" a stated date; Legacy for 6 months or 45 days | Legacy blocks new customers and provisioned throughput |
| Mistral | No formal policy; about 24 days observed for one model | |
| Self-hosted open weights | None | Hosted versions at clouds still retire |
The same model can retire on different dates at the lab and at each cloud. Products retire too: OpenAI ends its Assistants API, hosted evals and self-serve fine-tuning jobs between August 2026 and January 2027.
Observability: metrics, logs and traces
A piece of telemetry moves through states: emitted, collected, processed, ingested, indexed or stored, tiered down, deleted. Each change of state is a pricing event, and regulators have started naming the states: PCI's card-data rules require logs to be "immediately available" for three months, and the US federal memo M-26-14 distinguishes "actively searchable" from "retrievable".
| Tier | What it means | Rough cost |
|---|---|---|
| Hot, indexed | Searchable in seconds, used by alerts and dashboards | Datadog: $1.06-2.50 per million log events, for 3-30 days |
| Warm or flex | Searchable, slower, priced separately | Datadog Flex: $0.05 per million events stored, plus compute |
| Archive | Object storage the customer often owns | About $0.02-0.03 per GB a month |
| Rehydrated | Archive pulled back into the index for an investigation or audit | A fee per GB, plus minutes to hours |
| Deleted | Retention expired, or an erasure request | Free, but hard in append-only stores |
Rehydration is fading. Since 2025, Datadog's Archive Search, Elastic's frozen tier and lakehouse engines let you query old data where it sits.
An alert goes from inactive to pending to firing, then acknowledged and resolved. An incident is detected, triaged, mitigated, resolved, reviewed and closed. Standards have states too: OpenTelemetry's traces, metrics and logs are stable, profiles reached alpha in March 2026, and the conventions for AI and LLM telemetry are still marked "Development".
The contract has states as well: free tier, annual commitment, monthly usage, overage, true-up, renewal. Seat reductions at Datadog take effect only at renewal, and overage history tends to become next year's commitment, so commitments ratchet up.
Phone numbers, messaging and voice
A number has one lifecycle; a sender has several running side by side, and the bugs tend to show up where they disagree.
The NANP number lifecycle, from the platform's side: spare → allocated to a carrier block → in your inventory → reserved → active → porting out or released → aging → spare again. Aging is set by law: a disconnected number must age at least 45 days, and at most 90 days for residential and 365 for business numbers. Toll-free runs its own clocks for reserved, assigned and disconnected numbers.
On top of the number, every sender layer keeps its own state:
| Object | States that matter | Who moves it |
|---|---|---|
| 10DLC brand | Unverified, verified, Auth+ pending or passed, vetted | TCR and vetting firms |
| 10DLC campaign | Pending review, rejected, approved, registered at each carrier, suspended | The DCA (direct connect aggregator, which reviews campaigns before carriers see them), then each carrier |
| Toll-free number | Text-enabled, pending verification, verified, rejected | Somos registry, the verification reviewer |
| Short code | Leased, approved per carrier, live, suspended, lapsed | CTIA's registry, each carrier |
| RCS agent | Per carrier: pending, launched, rejected, suspended, unlaunched | Google or the carrier |
| Number connected; display name approved or declined; templates pending, approved, paused, disabled | Meta | |
| 911 record | Validated, provisioned, stale | Platform and its 911 provider |
Abroad, add degraded states: delivered but labelled "Likely-SCAM" (Singapore) or "Unverified" (Australia), or delivered with your sender overwritten.
Take the same clinic chain. Its appointment-reminder campaign clears the DCA on Monday; it can be live at T-Mobile by Tuesday and still pending at AT&T on Thursday, so some patients get their reminders and others don't.
"Approved" in one system rarely means live everywhere, so show customers the state per carrier. Port-out and release should also clean up the campaign link, any hosted SMS route (texting on a number whose voice stays with another carrier) and the 911 record.
Voice AI: real-time voice agents
A call carries several state machines that move at different speeds.
- Turn: listening, caller speaking, end of turn pending (when the model starts early), thinking, tool running, agent speaking, interrupted, then back to listening.
- Call: ringing, answered by a person or a machine, in progress, transferring, ended (completed, no answer, busy, failed or cancelled).
- Model session: Google's Live API adds connected, "going away" (about 60 seconds' warning) and resumed, because a connection lasts about 10 minutes and an audio session about 15.
- Consent: none, prior express, prior express written, then revoked, and a revocation has to be honoured within 10 business days.
On outbound calls, answering-machine detection takes about 4 seconds after pickup with Twilio's defaults. A model upgrade changes turn-taking behaviour as well as answers, so I'd put timing in every regression test.
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.