A trigger catalog for reliable AI agent work
Turn schedules, messages, files, record changes, thresholds, and handoffs into work with an owner, a destination, and a completion check.
A trigger is the recorded event that starts work without a fresh instruction from a person. FidelicAI's current working catalog uses six event types: schedule, inbound message, document arrival, record change, threshold crossing, and human handoff. The label matters less than the full trigger record: source, event, exclusions, work product, owner, destination, and completion check.
The AI agent versus chatbot guide separates a role that remains responsible for work from a tool that answers the current request. Triggers are one way a role continues the job between conversations.
That continued work replaces some manual watching, routing, status checking, record gathering, and follow-through that people perform today. People still decide which events matter, approve consequential actions, resolve unfamiliar cases, and remain accountable for the business result.
Define the event and the finish line together
Six trigger types for recurring work
Each event starts a specific work product and retains a person at the approval boundary.
| Trigger type | Example event | Required record |
|---|---|---|
| Schedule | Monday at 8:00 a.m. in the buyer's time zone | Time zone, source window, work product, destination, owner |
| Inbound message | A reply enters an approved sales or support queue | Channel, intent rule, exclusions, response window, approver |
| Document arrival | A contract, policy, invoice, or brief enters an approved folder | File type, source location, duplicate rule, reviewer, result |
| Record change | A project, listing, pipeline, or policy record changes | Object, fields watched, previous value, new value, resulting work |
| Threshold crossing | A measured value moves beyond an approved boundary | Measure, source, comparison value, observation window, owner |
| Human handoff | A person marks an item for review, correction, or personal handling | Person, reason, current state, required response, stop rule |
These are design categories. The current public role still controls whether a trigger is offered and which work product it may start.
Schedule triggers start dated work
A schedule should name a time zone and source window. “Weekly brief” is incomplete if the role cannot determine which week, which records, or which owner receives it. ALEK, the AI chief of staff, can maintain a regular owner brief from approved priorities, decisions, commitments, and risks. The owner still sets direction and approves consequential moves.
Done when: the result displays the covered period, source records, missing inputs, decision owner, destination, and next scheduled run.
Message triggers require an intent rule
A new message should not start the same action for every sender or phrase. VYRA, the outreach and SDR lead, can classify replies and prepare approved follow-through within the current campaign boundary. The trigger record needs included intents, excluded senders, duplicate handling, urgency rules, and the replies that require personal handling.
Slack's Events API illustrates the underlying event loop: an event is delivered to an app, the app acknowledges it, and business logic decides what happens next. OAuth scopes and channel membership still limit which events the app can receive.
Document triggers need a duplicate and version rule
A newly arrived file may be a revision, duplicate, irrelevant attachment, or incomplete record. ETHA, the AI insurance manager, can keep policies, certificates, applications, and evidence aligned. VELA, the AI compliance manager, can maintain approved compliance records. Neither role acquires licensed or legal authority from the arrival of a document.
Done when: the file has a stable identifier, received time, version relationship, approved source location, required work product, review owner, and unresolved missing evidence.
Record-change triggers need field-level boundaries
A change to one field should not imply permission to read or rewrite the whole system. A webhook (a system-sent notice that a record changed) can notify an app without granting permission to read or rewrite the whole source. Airtable's Webhooks API can report record or field changes. Asana's webhook endpoint can report changes to a resource. The role must still state which object and fields matter and what checked result follows.
For search work, FARO, the AI SEO strategist, can record changes in indexation, local facts, reviews, cited sources, and sampled AI answers, then prepare a prioritized correction plan. The owner's team makes public changes.
Threshold triggers need a denominator
“When performance drops” is not executable. Record the measure, unit, source, comparison value, observation window, minimum sample, and owner. A customer-response delay over two hours, for example, requires a rule for business hours and which conversations count. An alert without a denominator invites noise.
Human handoffs start work and stop work
A person may reject a draft, flag an exception, approve a safe action, or take ownership of a sensitive reply. The event must state what the AI role should begin and what it must stop. “Needs review” without a reason leaves the next worker guessing.
From event to accepted result
A reliable trigger keeps the event, work, review, and retry state connected.
- 1
Observe one approved event
Capture the source, event identifier, time, and values needed by the rule.
Owner: Connected system
- 2
Apply exclusions and duplicate checks
Reject out-of-scope events and link retries to the existing work record.
Owner: Fidelic agent
- 3
Create the role-specific work product
Use only approved sources and preserve missing or conflicting evidence.
Owner: Fidelic agent
- 4
Report the result or block
Send the result, source links, approval request, or failure state to the agreed destination.
Owner: Fidelic agent
- 5
Accept, correct, or escalate
The designated person records the decision and any correction needed before the next event.
Owner: Buyer reviewer
Follow the connected questions
Define finished work includes this decision and the questions that usually change it.
What is a workflow for an AI agent?
A workflow is a repeatable path from an input or event to a named work product, with checks and ownership at each consequential step.
See workflows end in work products →How does an AI agent do a multi-step job?
It reads the current state, chooses the next allowed action, uses an approved system, checks the result, records what changed, and continues or stops.
What belongs in an AI agent work order?
Name the assignment, source set, work product, checks, approval point, delivery window, and what happens when an input is missing.
What should you do next?
Choose one recurring workflow from the current AI agent directory. Write the event source, exact condition, exclusions, duplicate rule, work product, destination, reviewer, completion check, retry limit, and stop condition. Use the work-environment guide to decide where the result and approval belong. Compare the current rate board only after the trigger can be tested with harmless sample records.