LeadRefinery — How the Platform Works

The seventeen checks before each dial

Compliance on this platform is not a policy document. It is a gate: a set of seventeen checks that runs before each dial. The checks run in a fixed order. The first failed check stops the call. The software has no path around the gate.

The gate runs at the moment of the dial

Three facts apply to the gate:

  • The gate runs immediately before the dial — not at the time when the call was scheduled. A lead can pass the checks at 9 a.m. and opt out before 2 p.m. The gate at 2 p.m. sees the opt-out and stops the call.
  • Each dial goes through the same gate. This includes campaign calls, single calls that a person starts from the console, and the demo calls that a person requests on the demo screen. A demo call passes the same seventeen checks. It differs only in the facts it brings. The demo request counts as the number's consent, the calling hours come from where the number is, and the number is treated as already screened. If a check says no, the phone does not ring, and the demo screen tells the person why in plain words. No type of call has a special path.
  • The gate is one indivisible operation, and it records each verdict. The full evaluation is one question to the database, and the gate cannot pass in part. The gate answers in less than one second. It writes its verdict — pass or block — to an audit trail before it answers.

A note about the name: the plans call this “the twelve checks”. During the build, one check got a sub-step: a check that the outside phone services have spare capacity. Before all of them, the gate checks that the lead, the campaign and the caller number belong to one client, and that the campaign itself is still running. In the middle, the gate added a check that the agent has a real name to greet the merchant by. And after all the others, the gate added a check for a free line platform-wide, so twenty held calls cannot all start at the same second. Thus the built system has seventeen checks. (An eighteenth, a hold from an early-warning screen on the number and the business, exists in the code but is switched off for every client today; see what is still being finished.)

A scheduled call moves along a single rail through fifteen numbered checks. The checks sit in five groups: call permission, call time, phone resources, the emergency stop, and a free line. An amber arrow next to each check shows the result of a 'no'. The possible results are: a permanent block, a wait, a change of caller number, or a pause with a retry. The first failed check stops the dial. The phone rings only after all fifteen checks pass. A note at the bottom states that the gate writes each verdict to the audit trail before it answers.
After the client and campaign checks, the other fifteen checks sit in five groups: call permission, call time, phone resources, the emergency stop, and a free line. The first failed check stops the dial before the phone rings. The gate writes each verdict, pass or block, to the audit trail before it answers.

The seventeen checks

The list below shows the checks in the order in which they run. The checks sit under the four questions that the gate asks. A “no” is not always permanent. Some checks block a lead permanently. Other checks delay the call to a later time.

Is this the right client?

  • 0. One client, end to end. The lead, the campaign and the caller number all belong to the same client. This check runs first, so one client can never dial with another client’s data. On a “no”: a permanent block.
  • 0a. The campaign is still running. A dial made for a campaign is refused the moment that campaign is killed or archived, or the lead has left it. A paused or finished campaign still passes this check, by design — pause and stop do not recall a lead already released. This check exists because an operator once ended a campaign by editing its record directly, bypassing the normal kill button, and leads on it kept dialling for a day. On a “no”: a permanent block, the same end a kill gives.

Is this call allowed at all?

  • 1. Correct stage of the journey. The lead is active in its calling schedule, or the lead holds a scheduled slot. On a “no”: a permanent block.
  • 2. Not on a suppression list. No suppression record exists for the lead, for the number, or platform-wide. On a “no”: a permanent block.
  • 3. Consent on record. Evidence of consent exists for this exact number, with its source and its date. On a “no”: a permanent block.
  • 4. Recent screening. The compliance screening of the number is not more than 21 working days old, and not more than 31 calendar days old. Both limits must hold; the platform uses whichever limit is reached first. On a “no”: the platform screens the number again, then puts the call back in the queue.
  • 5. The client's own exclusion list. The person is not on the exclusion list of the client. On a “no”: a permanent block.
  • 6. Permitted region. The person is not in a region that this client excludes. On a “no”: a permanent block.
  • 7. A real name to greet the merchant by. The lead's own first name is on file, and the client's own business name is a real name, not a placeholder still holding its own blank slot. This check exists because the agent used to greet the merchant by name only once the phone had already rung, so a missing name hung up on a person who had already answered. On a “no”: a missing lead name ends the lead, because no wait fixes it; a placeholder business name instead waits until the next calling day, because fixing the client's own setup can repair it.

Is this an acceptable time?

  • 8. Legal calling hours. The local time of the person is inside the permitted hours, with a safety margin. On a “no”: the platform tries again in legal hours. A phone number can sit in more than one possible time zone — an area code from one part of the country, held by a person who lives in another. When this happens, the platform lists every possible zone for the number, checks the permitted hours in each one, and dials only when every zone on the list allows the call.
  • 9. Not a holiday or a Sunday. Today is not a federal holiday, anywhere the platform calls. A state can bar calls on further days of its own; today is not one of those days either. On a “no”: the platform tries again on a different day.
  • 10. This client is inside its frequency limits. The limits are: a maximum of 3 calls in 24 hours, and a minimum of 3 hours between calls. The platform also does not call in the same part of the day two times. These limits count the actual person, not the lead record: the same merchant entered twice under one client shares one budget, not two. On a “no”: the platform waits.
  • 11. The full platform is inside its frequency limits. The limits across all clients together are: a maximum of 3 calls in 24 hours, and 12 calls in 7 days, counted on the number dialled. Separately, a small number of states set their own cap on how often any client may call the same person, counted across every number that person has and every client calling them. On a “no”: the platform waits.

Are a caller number and the providers ready?

  • 12. A healthy caller number is free. The gate finds one of the client's own numbers with daily capacity that remains. The gate claims the number immediately, and thus two calls cannot use the same number at the same time. On a “no”: the gate changes to a different number.
  • 13. The providers have spare capacity. The voice service and the phone network can each accept one more call now. On a “no”: the platform holds back every dial to that provider for a short time, then tries one call to test whether room has freed up. Each further refusal makes the hold longer — about one minute, then five, then fifteen — so a large group of waiting leads does not retry at the same second and hit the same limit again. No attempt is spent during the wait.

Is the emergency stop off?

  • 14. No emergency stop is active. No stop is set at any level. If the gate cannot read the status of the stop, the gate does not dial. An unknown status means “no”. On a “no”: the dial stays blocked until the stop is off. Most stops are a deliberate click, but one kind sets itself: if the platform cannot confirm it saved a stop-calling request, it sets this stop for the whole platform on its own, because an unconfirmed stop is worse than a halted client.

Is there a free line right now?

  • 15. A live-call slot is free, platform-wide. Only a fixed number of calls can be in progress across every client at the same instant, a limit set by the phone network's own account. The gate claims a slot the moment it says yes, so two calls can never take the last slot together. On a “no”: the dial waits for a slot to free up. This check runs last, deliberately — a slot is worth taking only for a dial that every other check has already allowed.

The limit of 3 calls in 24 hours is the highest limit that federal and state law allow. LeadRefinery raised the limit from 2 to 3 in August 2026, after a compliance review found that a limit of 2 was refusing lawful, already-scheduled calls under the platform's own daily calling pattern. The platform now runs at the legal ceiling, with no spare room above it, so a miscount of even one call would be a breach rather than a near miss.

Seven compliance flags in a finished call

The gate controls calls before they start. After the call, the platform examines each finished call for seven compliance flags:

  • a do-not-call request;
  • revoked consent;
  • a wrong number;
  • a mention of an attorney or litigation;
  • a deceased person or a bankruptcy;
  • incapacity or a minor;
  • a language barrier.

Two independent methods find most flags. The first method is the platform's own artificial-intelligence (AI) analysis of the finished call, with the same fixed questions on every call. The voice service now sends only its standard summary of the call. The second method is a separate phrase check that reads the transcript, and this check is intentionally too cautious. The platform combines the two methods, and a flag from either method stands. Duplicated or garbled reports do not remove a flag that occurred one time.

A do-not-call request is held to a stricter rule than the other six flags. The platform only records it when the person's own words match a known stop phrase, or the agent itself logged the request during the call. The call analysis alone saying the person asked to stop is not enough: that case sends the call to a person for review instead of suppressing the number automatically. This rule exists because the call analysis has, in practice, read a complaint about call frequency as a stop request and suppressed a number nobody asked to suppress. For the same reason, the phrase check no longer counts “I don't understand” or “I can't understand you” as a language-barrier flag: English speakers say both on a bad line, and that flag alone used to be enough to mark a call as a do-not-call request. A call whose only flag is a genuine language barrier now goes to a person for review, rather than suppressing the number automatically.

A flag becomes the outcome of the call, and it outranks all other results from the call. The platform writes a suppression record on the same day, for the lead and for the number. The platform does not wait for the next monthly screening. This closes a known gap: without the same-day record, a person who says “stop” on day 3 gets calls until the screening on day 30.

A wrong number is handled more narrowly than the other six flags: the platform suppresses only that exact number, for this client alone, rather than the lead or every client. The lead itself is not suppressed; if it has another number on file, the platform moves on to that one. A misdialled or reassigned number is a fact about the number, not about the person who no longer answers it, so it must not block every client from ever calling that number's rightful owner.

Two disclosures on every connected call

Every call the agent connects to a person must state two things: that the caller is an artificial-intelligence voice, and that the call is recorded. After each call, the platform's own analysis answers, for each disclosure, whether the agent said it. A call with no answer is recorded as unknown, never as a pass. The platform records both answers and watches the failure rate on a live board, with a target of zero.

The written record checks the analysis

Since 3 November 2026, the platform checks each answer against the agent's own lines in the written record of the call. The platform applies these rules, in this order:

  1. The approved sentence is in the agent's lines. If the agent's own lines hold the approved sentence, or the half of it for that disclosure, the platform counts the disclosure as delivered. This occurs whatever the analysis said. A real case: a co-owner talked over the last word of the notice. The agent finished the notice, but the analysis said it was missed.
  2. The caller hung up early. If the caller hung up before the agent had a second turn to speak, the platform records “the call ended before the disclosure”, not a miss. A real case: six callers answered the greeting with “I can't talk right now” and hung up. The analysis wrongly reported a miss each time.
  3. Every other case. The answer of the analysis stands. Thus a real miss still counts as a miss: the agent kept talking and never said the notice.

When a missed disclosure raises an alarm

A missed disclosure on a call of 20 seconds or more pages a person at once, with a top-priority alarm. It also fails the automatic campaign audit. A shorter call does neither. The reason comes from 316 real calls in the 14 days to 3 November 2026. The agent had said that it is an artificial-intelligence voice:

  • by 13.5 seconds on the typical call;
  • by 18.1 seconds on 95 calls in 100;
  • by 22.9 seconds on 99 calls in 100.

A shorter call seldom gives the agent time to say it. An example is a caller who swears and hangs up. The record of the call does not change: it still keeps what the analysis found. Only the alarm and the audit leave out the shorter calls.

A fast-moving merchant, asking for something such as the application link in the same breath as answering the phone, used to get it before the agent said either disclosure. The platform now reads the agent's own words on the call before it carries out a mid-call request. If the notice has not yet been said, the platform holds the request and has the agent say the notice first, then ask to continue. The request only goes ahead once the merchant agrees.

Claims the agent must never make

The agent must never quote a rate or a repayment term, name a funding amount, or promise approval. Only the client's specialist talks about these subjects. The after-call analysis reports any such claim. A reported claim has two results:

A claim must have its words in the agent's lines

Since 3 November 2026, a claim of a rate, an amount or an approval must have its words in the agent's own lines. These words are:

  • a figure or a percentage;
  • a dollar amount;
  • words such as “approved” or “you qualify”.

Without those words, the platform records no claim. Instead of a page, the call goes in front of a person for compliance review. A real case: the analysis reported a rate claim for the agent's line “That depends on the file — I can't speak to fees or rates”. In that line, the agent refuses to quote.

A claim of any other kind keeps the answer of the analysis, because words alone cannot decide it. An example: the agent tells the caller that the file is closed, but the platform did not close it.

Email has its own, separate opt-out rule

A call follows the rules above. A sent application-link email follows a different law, aimed at business email, not phone calls. Every one of these emails must carry the sender's true postal address and a one-click unsubscribe link; see how the platform sends and tracks one for both.

The rules apply where the person lives

The platform applies the calling-hour rules of the place where the person lives, not the place where the client operates. Federal law permits calls from 8 a.m. to 9 p.m. in the local time of the person. If the state of the person has stricter rules, the platform applies those rules. The platform stores each state rule with a citation of the source law.

  • Kentucky permits calls only from 10 a.m. — the latest start in the country.
  • Connecticut, Michigan and Minnesota permit calls from 9 a.m.
  • Alabama, Louisiana, Massachusetts, Maryland and Florida permit no calls after 8 p.m.
  • Alabama and Louisiana permit no calls on Sundays.
  • Florida and Maryland set a limit of three call attempts for each day.
  • Louisiana bars calls on state days, for example Mardi Gras, Good Friday and election days.

The platform excludes eight states fully. Maine is the one true legal exclusion: its law, in effect, bars automated calls to all possible contacts, and the platform’s own rule book says, almost word for word, do not try to configure around it. The other seven — California, Rhode Island, Washington, Arkansas, Montana, Mississippi and Wyoming — are excluded from the pilot as a client decision, mostly over litigation risk rather than a law the platform cannot satisfy.

Some states have more than one time zone, and some states do not use daylight-saving time. For these states, the platform finds the actual local time of the person, not one time for the full state.

The three scopes of a suppression record

A suppression record has one of three scopes: one lead, one number, or the full platform. The platform-wide scope is intentionally narrow. It holds only facts that are true for each client that dials the number:

  • federal do-not-call registrations;
  • flags for known litigators;
  • numbers reassigned to a different person.

The rule is simple: business data stays with one client. Safety data goes to all clients. One case follows this rule to its conclusion: two different clients can hold the same person's number at the same time. If the person tells one client to stop calling, the platform marks the number for every client that holds it, the same minute, not only for the client on the call. A "stop calling me" request is a safety fact, so it is never kept private to one client's business data.

A stop-calling request does not have to arrive on a call. An admin can record one by hand — from a letter, an email, or a complaint by phone — from a lead's own record or from a dedicated screen in the control room. A hand-recorded request takes the same platform-wide path as one the agent hears live, and it still works for a number with no lead uploaded yet: the number is refused the moment someone tries to bring that person in.

A stop-calling request the agent hears on a call it cannot match to a client — an inbound call, or a test call from outside the platform — is held on its own, refusing that number for every client until a LeadRefinery staff member decides which client it belongs to, or sets it aside as a false alarm.