How to send your AI agent's data to Attify
There are four ways: connect the customer's own help desk, CRM or booking system; connect the agent platform or observability account the agent runs on; stream each interaction through the Ingest API or the Python and TypeScript SDKs; or point an OpenTelemetry exporter at Attify. The first is connected data, and so is a platform account the buyer owns; a vendor's own platform account and the streamed paths are witnessed data, which counts toward a verification once the customer consents and reconciles.
1. Connect the customer's systems#
The customer connects its own system through its own authorization, from its Attify account or a vendor's invitation. Attify reads every record of the period, with the vendor's claim from the platform's own field (for example Intercom's resolution state on Fin conversations). This is the strongest provenance and needs no code.
2. Connect a platform account#
A vendor creates a deployment with "Attify reads my platform", chooses the platform and pastes a key: Attify reads the account every 15 minutes as witnessed data, and the first read, which reaches 90 days back, is counted but never scored as witnessed, because it arrives late. A buyer connects its own account like a help desk, from its dashboard or a vendor's invitation, as connected data. Whose account it is decides the provenance, and Attify records it when the account is connected (details).
3. Stream with an SDK#
Create a deployment for one customer in Agents (/deployments), attesting that you may share its interaction data, and copy its key: Attify shows it once.
import os
from attify_ingest import Attify
attify = Attify(key=os.environ["ATTIFY_KEY"])
chat = attify.interaction("conv_123", channel="chat", customer="cust_42", external_ref="ticket_9")
chat.message("customer", "Where is my order?")
chat.message("agent", "It ships tomorrow.")
chat.claim("resolved")
chat.end("completed")
attify.close()
Install it with pip install attify-ingest; the TypeScript SDK (npm install attify-ingest) has the same shape. Both sign each request, number events per stream, buffer and retry. Any HTTP client works too. Raw HTTP: POST /v1/ingest/events with Proven-Key, Proven-Timestamp and Proven-Signature (HMAC-SHA256 of <timestamp>.<body>) headers. Full reference: /docs/ingest.
4. Send OpenTelemetry traces#
OTEL_EXPORTER_OTLP_TRACES_ENDPOINT=https://api.attify.app/v1/otlp/v1/traces
OTEL_EXPORTER_OTLP_TRACES_HEADERS="Authorization=Bearer pdk_..."
Attify maps the gen-ai semantic conventions (v1.41.1; OpenTelemetry) to interactions and messages. The agent's claim needs one attribute, proven.outcome; handoffs proven.handoff; the customer's ids proven.customer and proven.external_ref.
What every path needs to carry#
- The claim: the agent's own outcome, in the job type's words.
- Handoffs: to a person (
human), another AI agent (agent) or another channel (external). Onlyhumanandexternalfail a claim. - Who wrote each message:
customer,agent,human(a person on the customer's side),system,tool. - The customer's ids:
customerandexternal_ref, as the customer's own systems know them. - Explicit resolution answers:
feedbackof kindresolution(1 yes, 0 no). Satisfaction ratings go ascsat,thumbsornpsand never count as confirmation.
What Attify checks on arrival#
Events more than 5 minutes late are accepted and flagged; an interaction with a late event is never scored as witnessed. Missing sequence numbers and reused ids show on the deployment's health. Above 0.1% gaps or 1% late interactions in the scoring period, the deployment stops counting until it recovers; for support resolution, so does leaving more than 2% of the conversations the AI handled without a claim either way.
Questions
Which path should I use?
- If the agent works inside a help desk, CRM or booking system Attify reads, connect that system: it is connected data. If it runs on Vapi, Retell AI, ElevenLabs, Bland, Voiceflow, Langfuse or LangSmith, connect the account. Otherwise use an SDK, the Ingest API or OpenTelemetry.
How do I report the agent's outcome?
- Send an outcome.claimed event with the job type's vocabulary (support: resolved or not_resolved; sales meetings: meeting_booked or no_meeting; bookings: booked or not_booked). In OpenTelemetry, set the proven.outcome attribute.
Why do customer and external_ref matter?
- They are the ids in the customer's own systems. Attify uses customer to find repeat contacts in the customer's help desk, and external_ref to find the ticket, meeting or appointment a spot check looks up.
Does streamed data count toward an audited score?
- Only after the customer consents in Attify and reconciles each month. Until then it feeds the vendor's private report.