AI Agents

Designing a phone-number check an agent can act on without a human in the loop

Removing the human from the loop does not just remove a click, it removes the judgement that human was quietly applying to ambiguous results. A phone-number check an agent runs on its own needs three things a typical human-facing integration can get away without: a response it can branch on with no interpretation, an explicit no-data state instead of a confident guess, and a per-call cost that scales with however many times the agent actually decides to call it.

What breaks when a human stops reading the result

A risk analyst looking at a lookup with carrier: null knows, without being told, to fall back to manual review. An agent executing a decision does not have that instinct unless the fallback is written into its branching logic ahead of time. Every ambiguous case a human used to absorb silently becomes a case the agent needs an explicit rule for, which means the design work moves from writing a dashboard to writing a decision tree, before a single line of integration code exists.

An explicit unknown beats a confident guess

An API that never admits it does not know something forces whatever consumes it, human or model, to either treat missing data as a negative result or invent a best guess, and both are unsafe in a fraud decision. Telebase's response field reference returns UNKNOWN for SIM swap detection in GB, DE, NL and FR today, while carrier registration in those markets completes, rather than a fabricated SWAPPED or NO_SWAP. An agent built around this pattern treats the absence of a signal as exactly that: absence, not evidence either way.

A worked example: verifying a payee number before executing a transfer

An agent handling a payee-change request can run a lookup the moment the new number is submitted, before the transfer step is reachable at all. If numberType is a VoIP type, it routes the request to manual review instead of proceeding, since a disposable-style number on a newly added payee is a known fraud pattern. If active is false, it does the same. If the SIM swap field, once live in that market, returns SWAPPED, it holds regardless of everything else. If it returns UNKNOWN, the agent proceeds on the strength of the other fields rather than blocking on the absence of a signal that was never promised to exist yet. None of this requires a person to look at the case first; a person only sees the ones the branching logic actually routes to them.

Designing for a call volume that is not fixed in advance

A human-triggered integration has a roughly predictable call volume: one lookup per sign-up, one per password reset. An agent's call volume is set by its own decision path through a conversation or workflow, which can vary a great deal from one run to the next. A flat cost per successful lookup, with no annual commitment and no monthly fee, matches that usage pattern directly: the bill tracks decisions made, not a licence bought in advance for a headcount that has nothing to do with how often the agent actually needs to check a number. Start with a hundred free queries to prove the pattern out before any cost is involved at all.

What still needs a human

None of this removes human review from fraud operations altogether. The value is removing the lookup step, and the routine branch of the decision, from the path that needs a person, not eliminating oversight for the whole process. Disputed and borderline cases still belong in a human queue; the agent's job is making sure only those cases arrive there; see giving AI agents phone intelligence tools for how the call itself gets wired into an agent framework in the first place.