An AI agent should not decide whether an uploaded image, video, or voice recording is genuine just by looking at its content. Put a specialist verification step before any consequential action. Let the detector return a score and evidence, route uncertain cases to a person, and allow the agent to continue only when the result satisfies a policy your team chose in advance.

That creates a useful division of labor:

ScamAI provides deepfake and forgery detection across images, video, audio, documents, and live calls. Its Eva model returns a probabilistic score, a tiered verdict, and a readable summary of what it found. The ScamAI integration in Aident Loadout brings its unified image, video, and audio detection endpoint into the same governed capability layer as the other Apps and Actions an agent uses.

You can start with Aident-managed ScamAI access without supplying your own API key. Use the Create Detection Action with saving disabled, inspect the price before running it, and keep consequential downstream Actions behind review. Loadout records the execution in Audit; the ScamAI result supplies the verdict and evidence used by your routing policy. ScamAI and Aident remain separate products, and this article is a workflow design rather than a joint performance benchmark.

The verification gate belongs before extraction and action

Suppose a claims agent receives a damaged-property photo, a short site video, and a voice note. A weak workflow starts extracting details immediately, updates the claim, and drafts a payout recommendation. If one input was manipulated, the agent may turn questionable media into clean-looking structured data before anyone notices.

A safer sequence is:

1. Receive the file without treating it as true. Record its source, case ID, media type, and capture context.

2. Run the appropriate verification check. Send the media to a dedicated detector before using its contents to drive a decision.

3. Normalize the result. Store the verdict, score, evidence summary, detector version, and time checked beside the case.

4. Apply a fixed route. Continue, request another source, or send the case to human review according to written thresholds.

5. Only then let the agent act. Extract data, update a system, draft a response, or call another tool within its approved permissions.

The same pattern works for a recruiting team checking an interview recording, a lender reviewing a recorded verification call, or a trust-and-safety team triaging reported media. The files and consequences differ, but the boundary stays the same: verify first, automate second.

Do not make the language model the detector

A general-purpose model can summarize a file, compare names, and spot obvious inconsistencies. That does not make it a forensic detector. A polished explanation can still be based on a manipulated input, and the agent may sound most confident when the visual or audio evidence is hardest to judge.

Specialist detection is also not certainty. ScamAI's API returns one of three verdicts: LIKELY_REAL, ALERT, or LIKELY_AI, together with a score and summary. That is the right shape for workflow design because it preserves uncertainty instead of hiding it behind a binary answer.

Keep three other limits visible:

• A file that appears authentic can still contain a false claim.

• A synthetic file can have a legitimate use, such as an authorized illustration or translated voice track.

• Detection does not prove identity, ownership, consent, or business eligibility.

The detector answers a narrower question: which manipulation signals are present, and how strongly did the system flag them? Your policy decides what happens next.

Use three routes, not one pass-fail rule

Map the detector result into a routing table before the agent sees a real case. For example:

Do not copy these bands from a vendor demo or a headline accuracy figure. Test them against your own media, capture conditions, countries, and fraud patterns. Measure false positives and false negatives separately. A threshold that is tolerable for moderating a low-reach post may be unacceptable for rejecting a job applicant or freezing a financial account.

For each route, define the allowed next Actions. A likely-real result might permit extraction and a draft. It does not automatically authorize payment, account closure, publication, or another irreversible change.

Give the reviewer a compact evidence packet

An alert without context merely moves the bottleneck. When the agent pauses, it should present the reviewer with a small, consistent packet:

• the original source and how it entered the workflow;

• the file type and capture context;

• the detector's score, verdict, and evidence summary;

• the policy rule that caused the pause;

• the exact downstream Action the agent wants to take;

• the target account or record that Action would affect;

• what changes if the reviewer approves or declines it.

Store this as a normalized internal record rather than assuming every detector uses the same response schema. Keep the vendor's original result available for audit, but make your policy depend on fields your system controls. That makes it easier to change models, compare versions, or add a second check without rewriting every downstream workflow.

If the input contains personal or regulated data, minimize what the agent receives. ScamAI's Trust Center says uploaded media is scored and discarded while results are retained. Its Eva page describes cloud, on-premise, and on-device deployment options. The API workflow here sends media to ScamAI; it is not on-device detection. Confirm the applicable deployment, retention, and access terms before sending production media.

Run the check through the ScamAI integration

The ScamAI integration follows the provider's OpenAPI contract. Its unified detection Action accepts one supported image, video, or audio input as a file or public media URL. A successful result includes the media type, model, verdict, score, readable summary, and provider credits used. Video results may also include per-frame scores; audio results may include scored or abstained segments.

Use this sequence in an agent that has Aident Loadout installed:

1. Open the ScamAI App in Loadout. Aident-managed access supports detection without a personal provider key; stored history and account features require personal credentials connected through Vault.

2. Find Create Detection and inspect its live input schema.

3. Validate one authorized file or URL, set save to "false" for Aident-managed access, and inspect the current USD quote before execution.

4. Run detection and preserve the returned verdict, score, summary, model, and credits_used value.

5. Apply the routing policy before calling any downstream Action.

6. Record the ScamAI result and each later Action as distinct audit evidence.

Start with a bounded request such as:

Use ScamAI to check this one authorized media URL with saving disabled. Show me the quoted cost before running detection. Return the verdict, score, summary, model, and credits used. Do not update a case, send a message, or take another downstream Action. If the result is ambiguous or unavailable, stop and tell me what needs review.

For example, after the verification route is resolved, the agent could use another authorized Action to create a review task, attach a bounded evidence summary, or update a case status. Before any write, inspect that Action's contract, validate its input, and use the narrowest approval scope. Loadout's Audit record tells you which Action ran, through which connection, whether it succeeded, and what it cost. It does not certify that the media was genuine or replace ScamAI's result.

This separation also gives you a clean failure policy:

• If verification is unavailable, pause rather than treating the file as safe.

• If ScamAI returns NOT_ASSESSABLE, route to review or request better media instead of treating the result as likely real.

• If the media type is unsupported, stop before extraction rather than sending it to a general-purpose model as a substitute detector.

• If the downstream App is unavailable, preserve the verified result and retry only the failed step.

• If the proposed target or input changes after approval, ask again.

Do not blindly resend a timed-out detection. ScamAI's contract says a timed-out run may still have completed and consumed credits. Inspect the Loadout execution record and any returned result first. With personal credentials, check the provider credit ledger or a saved detection record when available; Aident-managed access does not expose those account-history features. If completion remains unknown, pause for reconciliation instead of automatically resubmitting. The contract also identifies duplicate media sent by the same account within 30 days with zero_charge_reason: duplicate, but that is a metering result, not a general idempotency guarantee for the rest of your workflow.

Treat live calls as a separate path

Recorded files can wait at an intake gate. A live call cannot. ScamAI now presents Halo as available for Windows devices with Snapdragon X Series processors, with team setup arranged through a demo. Its public page describes on-device detection during meetings, with local processing rather than uploading the video stream. Halo is a separate product surface, not a live-call capability of the Loadout detection Action described here.

That signal should change the meeting procedure, not silently decide that a participant is fraudulent. A team might pause a sensitive request, verify the person through a second channel, or require another authorized approver. Write that response plan before the first alert. During an urgent call is the wrong time to invent it.

Evaluate the whole workflow, not one model score

Run a controlled pilot before connecting the gate to production Actions. Include genuine files, confirmed manipulations, poor-quality captures, unusual formats, and cases that are hard even for a human reviewer. Then track:

1. detection performance on your own case mix;

2. how often reviewers overturn each route;

3. time from alert to a resolved decision;

4. downstream Actions attempted, blocked, and completed;

5. duplicate work after timeouts or retries;

6. privacy, retention, and access behavior in the chosen deployment.

The workflow succeeds when uncertain media stops before it can trigger a consequential Action, reviewers receive enough evidence to decide, and cleared cases continue without unnecessary manual handling. A single top-line accuracy number cannot demonstrate that outcome.

Start with one low-volume intake path. Give the agent only ScamAI and the downstream tools it needs, keep consequential writes behind review, and test unavailable and ambiguous cases as carefully as the happy path. If Loadout is not installed, tell your coding agent exactly: Follow https://aident.ai/SETUP.md. Then run your first ScamAI detection with Aident Loadout. Use one authorized file, saving disabled, and no downstream writes.

For the next control, define which agent Actions need human approval. If you are choosing how ScamAI should connect to your runtime, compare MCP with a direct API and decide where policy and durable state will live.

Sources and scope

• ScamAI product overview

• ScamAI API documentation

• ScamAI OpenAPI 3.0 contract

• ScamAI Eva detection model

• ScamAI Halo

• ScamAI Trust Center

• ScamAI App in Aident Loadout

• Aident's ScamAI workflow walkthrough

Product details, the ScamAI Platform API v1.0 contract, and the production Loadout integration were checked on October 2, 2026. Document detection and Halo live-call detection are separate ScamAI surfaces and are not presented here as capabilities of the OpenAPI-based Loadout Action. This is not a hands-on benchmark or customer result. Recheck the live Action schema, supported media, result fields, pricing, privacy terms, and Halo availability before implementation.