Skip to content
Articles

Repeatable Supplier Checks: What Must Stay Fixed

Repeatable supplier checks hold four things fixed between runs: the entity identifier, the sources, the scope and the evidence format. Here is the spec.

By ApioskAgent workflows7 min read

Repeatable supplier checks need four things held fixed between runs: the entity's identifier and register, the named sources, the scope of questions, and the evidence format. If any of them drifts, a difference between two runs may come from the check and not from the supplier. Write them down once as a specification, then run that specification on a schedule or a trigger.

  • Use this when a supplier relationship outlasts onboarding and someone has to know whether the records changed.
  • Fix the subject by register and registration number. A name alone is a search, not a subject.
  • Give every run an explicit end state. A failed or partial run is never recorded as "no change".
  • Changing the sources or scope is allowed, but it starts a new baseline. It is not a finding about the supplier.
  • Skip the schedule when nobody is assigned to read the output.

A repeatable supplier check is a specification

Two people asked to "check the supplier again" will run two different checks unless the check is written down. The specification is the unit you repeat. A run is one execution of it.

A minimal specification has these fields:

  • Subject: the legal name as registered, the jurisdiction, the register, and the registration number stored as text. This is the same anchor described in supplier identity verification before onboarding.
  • Sources: each source by name, with the question it answers. "Sanctions screening" is not a source. A named list from a named authority is.
  • Scope: the questions asked, and the ones deliberately left out. Status, registered address and directors are three questions, not one.
  • Match rules: how a name is normalized and what counts as a candidate match, so that run twelve applies the rule run one applied.
  • Evidence format: the fields kept from each source, the retrieval time per source, and a link or reference to the record.
  • Cadence and triggers: the schedule, plus events that force a run outside it.
  • Budget: a maximum spend per run, and what happens when a run would exceed it.
  • Owner: the person or role who reads a changed result.
  • Version: a number that increases whenever any field above changes.

Store each run with the specification version it executed. A reviewer comparing March with June can then see whether both runs asked the same question.

Each source changes on its own clock

A single cadence for the whole check is a convenience, not a property of the data. Each source has its own update behavior, and some records tell you when they last changed.

Three documented examples:

  • Sanctions lists. OFAC states that the SDN list is updated frequently with no predetermined timetable. A monthly screen therefore has a known blind spot of up to a month, which is a policy choice to record, not a technical detail.
  • LEI records. The GLEIF LEI data format defines a LastUpdateDate, a NextRenewalDate and a registration status that becomes LAPSED when the record is not renewed by that date. A lapsed LEI still identifies the entity, but its reference data has not been recently re-certified.
  • Company registers. Update timing depends on when the company files and when the register processes the filing. Company registry data freshness separates the retrieval date, filing date and reporting period.

Keep the retrieval time per source, not per run. A run that completes on Tuesday may contain a register record fetched on Tuesday and a list file last published the week before. Where a source exposes its own last-updated date, keep that beside your retrieval time. The two dates answer different questions.

Access limits also shape a schedule. The Companies House API, for example, allows 600 requests within a five-minute period and returns HTTP 429 after that. A portfolio of several thousand suppliers checked at 09:00 on the first of the month will hit a limit like this unless runs are spread out. Other sources publish different limits or none at all, so check each one.

Compare runs field by field, and keep failures distinct

A repeat run is only useful if its result can be compared with the previous one, and comparison requires that both runs ended in a known state.

Give each run one of these end states:

  • Complete: every source in the specification answered.
  • Partial: at least one source answered and at least one did not. Name which.
  • Source unavailable: a timeout, rate limit or outage prevented the run. Nothing is established.
  • Identity unresolved: the identifier no longer returns the expected entity, for example after a dissolution, a merger or a re-registration. This goes to a person, not to a retry loop.

Then classify each difference against the previous complete run:

  • Source change: the same source returns a different value for the same field. This is the result you are running the check for.
  • Check change: the specification version differs. Do not report this as a supplier change.
  • Coverage change: a source that answered last time returned nothing this time, or the reverse. This is a gap or a gap closing, not a fact about the supplier.

Keep the previous value, the new value and both retrieval times. "Director changed" is weaker evidence than two dated records showing who was listed on each date. The same states and gaps apply at onboarding, as the supplier onboarding evidence checklist describes.

One edge case deserves a rule: a new candidate match on a screening list. A repeat run can surface a name match that did not exist before. That is a candidate for review, not a hit, and the specification should say who reviews it.

Retries, cost and agents

Scheduled checks fail in ordinary ways: a timeout after the request was sent, a worker restarted mid-run, a source that answers slowly. If the check costs money, a naive retry can pay twice for one result.

The standard pattern is an idempotency key: one key per intended run, reused on every retry of that run. The IETF HTTP API working group's draft on the Idempotency-Key header describes the behavior: a retry with the same key should get a response consistent with the completed original, and a key must not be reused with a different payload. It is an Internet-Draft, not a published standard, so treat it as a description of common practice and read each provider's own documentation. A workable convention is to derive the key from the specification version, the subject and the scheduled run date, so that a restarted worker produces the same key and not a second purchase.

This matters more when an AI agent runs the check. An agent that re-plans the check from a prompt each time will pick sources and phrasing that vary between runs, which defeats repeatability. Give the agent the stored specification and have it execute that. Apiosk's published agent guidance says the same about paid actions: reuse the idempotency key when retrying the same paid action, and never invent identifiers or prices. Apiosk for AI agents describes connecting an agent through MCP or API with a maximum per request and per rolling day.

Cost scales with suppliers multiplied by sources multiplied by runs per year. A per-run ceiling stops one expensive run. It does not cap the annual total, so compute that before choosing a cadence.

When not to schedule a check

A schedule is the wrong tool in several cases:

  • One-off purchases. If the relationship ends with the invoice, a single documented check is enough.
  • Event-driven risk. A request to change bank details or a contract renewal should trigger a run immediately. Waiting for the next scheduled date defeats the purpose.
  • No owner. Results that nobody reads create a record showing you held information and did not act on it.
  • Unstable specification. If the scope is still changing every month, you are still designing the check. Settle it first.
  • Questions that sources cannot answer. Delivery quality, financial health and who controls a bank account are not established by repeating a registry lookup.

Whether your organisation is required to monitor suppliers at all, and how often, depends on your sector and jurisdiction. That is a question for your own policy and advisers, not something a check specification settles.

Where Apiosk fits

Apiosk recurring checks repeat a check you have already approved. You choose the days and time and a maximum spend per run, and each run returns a sourced result and a record of its cost, delivered to Slack or Gmail. Change alerts deliver only what changed compared with the previous result.

Coverage varies by country and source. The specification fields above are a way to organise your own process and are not a description of an Apiosk export format.

To put an approved check on a schedule, start from the recurring checks page.

This article was prepared with AI assistance using the linked OFAC, GLEIF, Companies House and IETF sources and the current Apiosk product pages. It is not legal advice.

Frequently asked questions

What makes a supplier check repeatable?

A written specification that fixes the entity identifier and register, the named sources, the questions in scope and the evidence format. With those held constant, a difference between two runs can be attributed to the supplier's records. Without them, the difference may only reflect a different search term, source or reviewer.

How often should a supplier check be repeated?

There is no universal interval. Sources change on their own schedules, and OFAC, for example, states that the SDN list has no predetermined update timetable. Set the cadence per question from your own policy and risk, and add event triggers such as a change of bank details or a contract renewal.

Is a failed run the same as no change?

No. A run in which a source was unavailable, rate limited or returned nothing establishes nothing about the supplier. Record it as partial or failed, keep the previous result marked with its original date, and retry or escalate under a rule you set in advance.

Can an AI agent run repeatable supplier checks?

Yes, provided the agent executes the stored specification and does not re-plan the check from a prompt on every run. Give it the fixed identifier, sources and scope, a spending limit, an idempotency key for retries and a rule for what goes to a person. The decision on a changed result stays with your team.