Source checking is verifying a claim by going back to where it came from, confirming the origin says what is claimed, and judging how far that origin can be trusted. It is the opposite of taking a summary on trust. For anyone reviewing an AI agent's work, it is the habit that separates real oversight from rubber-stamping.
Why does source checking matter in AI agent work?
Agents summarise, combine and restate information. Each step is a chance for a claim to drift from its source: a figure copied from the wrong line, a policy quoted from an old version, a customer's assertion presented as a fact. A reviewer who checks only the draft is checking the agent's reading, not the truth.
Not all sources are equal, either. On a governed platform, every fact reaching an agent arrives stamped with where it came from, strongest first: sealed, from the system of record itself, such as the core ledger; signed, from a named party who answers for it; declared, stated by a customer or person and not verified; and external, arriving from outside, such as an emailed invoice. An external claim cannot authorise a payment. Knowing these classes tells you which claims need checking hardest.
How does source checking work in practice?
Three questions, asked of each claim that decides an outcome:
- Where did this come from? Identify the actual origin, not the agent's summary of it.
- Does the origin say this? Open it and compare, especially figures, dates and names.
- Is the origin strong enough for this action? A customer's statement may support a reply but not a refund; an emailed document may start a process but not complete one.
Teams that do this well often keep a list of their sources and how much each can be relied on, along the lines of a source provenance register template.
What is an example?
An agent at a building society drafts an approval for a customer's request to increase a savings account withdrawal limit, citing "verified identity". A reviewer we will call Hamid checks the source. The identity check was not from the society's own records: it was a scanned document the customer emailed, which makes it external. The society's rules require a check against its own records for limit changes. Hamid holds the draft and asks for the customer to verify through the society's app. The agent's summary was accurate as far as it went; the source behind it was not strong enough for the action.
How is source checking different from "source of wealth" or checking a certificate at source?
"Source" appears in other terms you will meet. Source of wealth and source of funds are anti-money-laundering checks on where a customer's money came from, explained in what source of wealth is and applied in how to review source of wealth evidence with an AI agent. They are specific uses of source checking, applied to money.
Checking a certificate at source means confirming a credential with its issuer rather than trusting the copy you are shown. It is source checking applied to qualifications, covered in what checking a certificate at source means.
Which terms are related?
- Provenance: the recorded origin of a fact and how it can be trusted.
- Grounded answer: an answer built from approved sources.
- Citation: the pointer in an answer that lets you check its source.
For how employers test this skill, read AI skills for job seekers.
Questions people ask
Do I need to check the source of every claim?
No. Check the claims that decide the outcome: amounts, identities, eligibility, payment details. Background wording rarely needs it.
How do I show source checking skills in an interview?
Describe a case where the summary looked right and the source did not support it, what you checked and what you did next. Name the type of source involved and why it was or was not strong enough.