Earnings-Triggered Research Automations: A Control Guide
A documented comparison of scheduled and event-triggered earnings research, with trigger design, failure handling, review gates, and vendor boundaries.
Published August 28, 2026 · Updated August 30, 2026
In this article
An earnings-triggered automation is safe only when the trigger, eligible sources, output schema, approver, and failure behavior are explicit. The research model is one component. The harder requirements are document versioning, entitlements, idempotent runs, alerts, and a record of what changed. Start with one narrow deliverable, make missing sources visible, and prevent generated output from updating a model or reaching a client without approval.
This public-source comparison draws on product pages and help documentation checked August 30, 2026. We did not run the products under shared conditions. We build one of the systems discussed, and the author is one of AllMind's founders. Competitor claims are vendor-reported and our own claims are first-party; contract-specific sources, pricing, and service levels remain unverified.
Clock schedules and document events solve different problems
| Trigger | Good use | Required guardrail | Common failure | Evidence status |
|---|---|---|---|---|
| Calendar schedule | Pre-earnings brief, weekly theme sweep | Check whether source set changed since last run | Repeats stale output | Workflow design recommendation |
| Release published | Initial results snapshot | Identify issuer, period, document type, and version | Fires on duplicate or corrected release | Workflow design recommendation |
| Transcript available | Call review and Q&A extraction | Mark live versus reviewed transcript | Publishes a quote from provisional text | Workflow design recommendation |
| 10-Q or 10-K filed | Source-linked model-update queue | Map fiscal context and filing accession | Applies values from wrong period or amendment | Workflow design recommendation |
| Alert or thesis condition | Escalation on a named change | Preserve rule and evidence that fired it | Model inference becomes an opaque trigger | Workflow design recommendation |
A calendar can prepare a preview before results. A document event can react after publication. Many desks need both, with separate templates and approvals.
What vendors publicly document
| Product surface | Trigger documented | Source boundary described publicly | Delivery | Material limitation from public source | Last checked |
|---|---|---|---|---|---|
| Quartr Automations | Schedule or new earnings/document event | First-party public-company information | Chat plus email, feed, or mobile notification | Product page does not publish contract pricing or a universal company cap | Aug. 30, 2026 |
| Marvin Labs scheduled reports | Schedule or filing, release, transcript, earnings-filing event | Filings, releases, and transcripts used by its agents | Saved and emailed report | Documentation says event behavior varies with local filing practice | Aug. 30, 2026 |
| AlphaSense scheduled custom agents | Daily, weekly, or monthly schedule | Selected AlphaSense sources and custom-agent scope | Email result | Documentation says scheduled agents are not real-time alerts and limits schedules to 10 | Aug. 30, 2026 |
| AllMind agents | Scheduled and monitoring workflows, per our own pages | Built-in licensed data plus buyer-controlled internal sources | Platform, alerts, and cited report surfaces | We do not publish a trigger SLA or a buyer-independent coverage benchmark | Aug. 30, 2026 |
| Build on SEC and vendor APIs | Team-defined schedule or event | Whatever the team contracts and integrates | Team-defined | Engineering team owns ingestion, permissions, retries, evaluation, and audit | Aug. 30, 2026 |
Quartr's August 2026 launch says tasks can run on a schedule or when earnings and new documents are released, over first-party company information. Marvin Labs documents four event types and a watchlist scope; its technical product documentation adds timing and plan detail. AlphaSense documents daily, weekly, and monthly custom-agent schedules, with specific limits and no real-time trigger claim.
Our Agent Studio provides recurring and monitoring agents across a research workflow. The agent does not begin with an empty workspace: AllMind's data layer spans 750M+ documents and 6,800+ premium data sources licensed from 100+ providers and partners. It includes filings, live earnings calls with real-time transcription and speaker-labeled processed transcripts, event calendars, licensed news and newswires, company guidance with revision histories, broker research, Expert Insights, FactSet fundamentals and Revere supply-chain data, LSEG estimates, S&P/Capital IQ market and index data, MSCI data, and exchange feeds such as CME. Buyer-controlled documents, models and warehouse data can be added to that built-in corpus. Our pages do not establish exact event-detection latency, peak concurrency, or source completeness, so treat those as unverified until we demonstrate them in your pilot.
Specify the automation as a contract
Use a one-page control record for every automated deliverable:
| Control | Decision to record |
|---|---|
| Identity and accountability | Unique automation ID, owner, purpose, permitted audience, and escalation contact |
| Scope and access | Company universe, entitled sources, and the behavior for restricted material |
| Event definition | Trigger, timezone, deduplication key, and expected source versions |
| Deliverable | Output schema, delivery channel, required citations, and approval gate |
| Recovery | Retry limit, backoff policy, and the behavior when a required source is missing |
| Records | Retention period, audit-log period, and the prompt, model, and workflow versions used |
The deduplication key matters. An 8-K, an investor-relations release, a corrected file, and the eventual 10-Q may all describe the same quarter. The system should update a versioned event record, not send four indistinguishable “new earnings” reports.
Build the output as an exception packet
For a release-triggered brief, require structured fields before narrative:
| Field | Output rule |
|---|---|
| Document status | Source URL, publication time, version, and event ID |
| Reported values | Value, unit, period, basis, passage |
| Guidance | Exact range, assumptions, prior range, passage |
| Variances | Approved inputs and deterministic formula |
| Changed language | Prior and current quotes side by side |
| Missing sources | Explicit list with expected arrival |
| Review queue | Items requiring analyst decision |
The generated note should use approved fields. If the transcript is absent, the call-review section remains pending. If consensus is stale or on a different basis, the variance does not calculate.
Failure handling is part of product quality
Test these cases before enabling unattended delivery:
- The same release arrives through two feeds.
- A company files an amendment or corrected exhibit.
- The transcript is live text, then changes after human review.
- A source times out after the trigger fires.
- One user can access a broker note and the recipient cannot.
- Forty covered companies report within an hour.
- A company changes its ticker or fiscal calendar.
- The model or prompt version changes between quarters.
- The requested metric is absent.
- The output destination rejects the report.
For each case, record expected state, alert, retry, and approver. A missing source should produce a visible incomplete status. It should never encourage the model to fill the gap from prior-quarter language.
Keep deterministic actions outside the language model
The model can classify documents, retrieve passages, extract candidates, and draft commentary. Code or a spreadsheet should calculate variances, validate ranges, enforce schemas, and apply approved model mappings. Identity and policy systems should enforce source permissions. When a source arrives through MCP, the protocol's authorization specification provides transport-level controls, but each vendor and buyer must still implement scopes, token validation, and logging correctly.
The SEC's EDGAR APIs can supply real-time submissions and XBRL facts for a build-it-yourself pipeline. That route offers control but creates operational responsibility: respectful access, issuer mapping, document normalization, retries, monitoring, data quality, and review remain with the team.
A four-week pilot
Week 1: replay ten historical events and build an answer key. Week 2: run in shadow mode, with no delivery beyond the project team. Week 3: deliver to covering analysts with mandatory approval. Week 4: expand to a small watchlist only if failure alerts and citations work.
Track operational measures:
| Measure | Definition |
|---|---|
| Trigger latency | Source publication to accepted event |
| Source completeness | Required sources present at run time |
| Duplicate rate | Duplicate runs divided by accepted events |
| Field correction rate | Analyst-corrected fields divided by reviewed fields |
| Citation validity | Claims with working passage links divided by checkable claims |
| Review time | Minutes from delivery to approval |
| Failed delivery rate | Reports not received at the approved destination |
Do not turn these pilot measures into generalized performance claims. State sample, dates, universe, and configuration.
Decision paths
Choose Quartr when the desired trigger and source boundary are company-published IR material and the desk wants a managed product. Marvin Labs is worth evaluating for document-event reports over its supported public sources. AlphaSense fits scheduled research over its licensed corpus, with the documented caveat that schedules are periodic and limited. Evaluate AllMind when the run must combine structured premium data, filings, transcripts, broker or expert research, live market context and buyer-controlled evidence across a coverage list, then deliver a cited artifact; verify the exact entitlement and trigger behavior in the pilot. Build internally when control outweighs the cost of operating the entire pipeline.
No route removes the analyst approval gate for thesis judgments, model changes, or client-facing conclusions.
Documentation and unverified items
The comparison uses official Quartr, Marvin Labs, AlphaSense, SEC, and Model Context Protocol pages, plus our own AllMind product pages, opened August 30, 2026. It records public trigger and workflow claims without scoring performance. We could not verify contract pricing, buyer-specific source rights, event-detection service levels, peak-load behavior, or failure-recovery time. Those items should be written into the pilot and procurement record, then rechecked before unattended delivery expands.