How Asset Managers Use AI Agents for 10-K Analysis and Quarterly Earnings Report Automation
Feb 9, 2026
How Asset Managers Use AI Agents to Process 10-Ks and Quarterly Earnings Reports
If you cover public equities, you know the feeling: a 10-K drops, the earnings release hits, the deck lands, and the call transcript follows. There’s signal everywhere, but it’s buried in long, inconsistent, time-sensitive documents where a single misread unit, period, or definition can throw off an entire note.
That’s exactly why AI agents for 10-K analysis are showing up in serious research workflows. Not because they magically “understand” filings, but because they can run a repeatable, multi-step process: ingest documents, extract structured data, verify it against source passages, compare against history, and produce research-ready outputs that analysts can review quickly.
In this guide, you’ll learn what asset managers actually process on earnings day, the end-to-end workflow that makes AI agents for 10-K analysis reliable, architecture patterns that work in practice, and the guardrails teams use to keep accuracy and compliance front and center.
An AI agent in investment research is a system that can take a goal (for example, “summarize the 10-K and identify material changes”), execute multiple steps using tools (document retrieval, parsing, extraction, comparison), and produce auditable outputs with built-in checks and human review.
What Counts as “10-K” and “Quarterly Earnings Materials” (and why it’s messy)
“Read the 10-K” sounds simple until you see the actual inputs your team has to triage. The reality is a bundle of formats and contexts, often with enough edge cases to break naive automation.
Document types asset managers actually process
Most earnings workflows involve a mix of:
10-K and 10-Q filings (often with exhibits that matter more than the main body)
Earnings release (press release with GAAP and non-GAAP, plus reconciliations)
Investor presentation deck (where management frames the narrative)
Earnings call transcript, including Q&A (often the most time-sensitive “why”)
XBRL versus HTML text extraction, plus embedded tables
PDFs and occasionally scanned exhibits or images
AI agents for 10-K analysis need to handle all of the above, because coverage work isn’t a single-document task. Even when the filing is the anchor, the call and deck can change what matters in the next 12 hours.
Where the highest signal lives by use case
Not every team reads the same parts for the same reasons. A good 10-K parser or SEC filing extraction workflow starts by mapping sections to the decision you’re trying to support.
For long-term fundamental research, high-signal areas tend to be:
Item 7 MD&A for drivers, constraints, and management framing
Risk factors for emerging threats and shifting language
Segment discussion and footnotes for what’s actually changing
Liquidity and capital resources for durability and optionality
For near-term earnings and positioning, the “hot” areas are different:
Guidance and key assumptions (ranges, midpoints, puts and takes)
Mix shifts and margin bridge commentary
Backlog, pipeline, or demand indicators (where applicable)
Pricing and cost commentary with concrete qualifiers
Then there are soft signals that don’t sit neatly in a KPI table:
Tone shifts, hedging language, and repeated disclaimers
New competitive mentions or changed emphasis
Subtle changes in how the company talks about constraints
A key design reality: filings can be adversarial in the sense that they are written to satisfy disclosure requirements, manage risk, and avoid over-commitment. Formatting varies widely, fiscal periods can be tricky, and definitions change. That’s why agentic workflows in finance succeed when they treat normalization and verification as first-class steps, not afterthoughts.
The Core Agentic Workflow (End-to-End)
Most teams get value from AI agents for 10-K analysis when the agent behaves like a disciplined analyst-in-training: it follows a checklist, documents its work, and refuses to bluff.
Below is a practical workflow that mirrors what high-performing teams operationalize.
Step 1 — Ingest and normalize documents
Start by making the documents “machine-legible” in a consistent way.
Common sources include:
SEC EDGAR feeds (direct or via vendors)
Transcript providers and market data vendors
Internal research repositories (SharePoint, OneDrive, shared drives)
Analyst models and prior notes stored in a knowledge base
Normalization should handle:
Convert HTML and PDF into clean text or markdown
Preserve tables as structured objects where possible (not flattened prose)
Perform section-aware chunking for filings (Item 1A, Item 7 MD&A, footnotes)
Attach metadata that prevents period confusion:
Ticker, CIK
Filing date and fiscal period
Document type (10-K, earnings release, transcript, deck)
Version or amendment status
This step often determines whether downstream summaries are trustworthy. If your “chunking” breaks a table or merges unrelated sections, your model will sound confident and still be wrong.
Step 2 — Extract structured data and key passages
This is where financial document intelligence becomes more than summarization.
A strong 10-K parser or SEC filing extraction agent typically targets:
Core KPIs: revenue, gross margin, opex, operating income, EPS, FCF
Balance sheet and cash flow points that drive thesis
Share count, buybacks, and dilution commentary
Segment metrics and segment mapping across periods
Two areas deserve special attention:
Unit normalization Filings and earnings materials constantly shift between thousands, millions, and billions. Non-GAAP adjustments can change definitions without obvious warnings. The agent should normalize units and store the original representation alongside the normalized number.
Table handling A lot of the most important numbers live in tables. Treat tables as their own extraction problem: identify row headers, column headers, periods, and footnotes, then validate that “Net sales” isn’t being matched to the wrong year.
The best workflows also extract key passages that explain the numbers. That makes the final output more useful and easier to review quickly.
Step 3 — Summarize with grounding and traceability
This is where many systems fail: they generate clean prose without being able to prove where it came from.
A reliable approach for AI agents for 10-K analysis is a strict “cite-or-abstain” policy:
Every numeric claim must trace back to a specific source passage or table cell
If the agent can’t find the source, it should say it cannot verify
It also helps to produce multiple summaries, because different readers need different depth:
Executive brief (10 bullets) A fast read that says what happened and what changed.
PM version (thesis-impacting deltas) Focus on drivers, guidance changes, risks, and surprises versus expectations.
Analyst version (deep sections with supporting quotes) More detail, especially around MD&A analysis with AI and footnote nuance.
This layered approach makes quarterly earnings report summarization actually usable across the team, instead of forcing everyone into one generic output.
Step 4 — Compare against history: “What changed?”
The killer feature isn’t “summarize this document.” It’s “tell me what’s new.”
Change detection typically includes:
Diff against prior quarter and prior year filing sections
Detect new or expanded risk factors and removals
Compare accounting policies and key definitions (especially non-GAAP)
Compare guidance language across calls and releases
A good output looks like a change log:
What changed (quote or excerpt)
Where it changed (section reference)
Why it matters (one sentence)
Materiality score (rules-based or model-assisted)
For PMs, this is often the highest leverage artifact from AI agents for 10-K analysis because it compresses hours of scanning into a focused set of deltas.
Step 5 — Generate downstream research artifacts
Once you have normalized text, extracted KPIs, and a change log, the agent can draft real deliverables:
Earnings notes draft with verified KPIs and context
Guidance tracker updates
Risk register updates (new risks, reworded risks, “escalated” language)
Watchlist alerts for recurring themes (inventory, pricing pressure, margin headwinds)
Updates to your internal knowledge base for future retrieval
This is where investment research automation becomes tangible: the agent doesn’t stop at “analysis,” it produces the artifacts your process already runs on.
A 5-step workflow to process 10-Ks and earnings materials:
Ingest and normalize documents
Extract KPIs, tables, and key passages
Summarize with grounded, traceable claims
Diff versus prior periods to identify what changed
Draft research outputs and push updates into your systems
High-Value Use Cases Asset Managers Prioritize
Not every workflow is worth automating first. Teams get the best ROI by starting with outputs that are time-critical, repetitive, and easy to evaluate.
Earnings day acceleration (minutes matter)
This is often the first production use case for AI agents for 10-K analysis because the time pressure is real and the workflow is predictable.
High-impact outputs include:
Rapid transcript and press release ingestion
Guidance extraction and normalization (ranges, midpoints, key assumptions)
Q&A tagging: recurring analyst questions, new disclosures, evasive answers
A one-page “what changed since last quarter?” brief
Teams often discover that AI agents for earnings call analysis are especially useful in the Q&A, where nuance is high and the value is in quickly locating the few moments that will matter for the next conversation.
Risk factor and litigation monitoring
The goal here isn’t to “summarize risks.” It’s to catch the changes.
A reliable system flags:
New risk factors
Material expansions of existing risks
Significant language shifts that change the implied posture
Changes in legal proceedings disclosures
Over time, this becomes a practical monitoring layer across a coverage universe.
KPI and segment performance tracking
If you’ve ever tried to track segment KPIs across 30 companies, you know the pain: segments get renamed, reclassified, and restated.
AI agents for 10-K analysis can support:
Standardizing segment metrics through mapping rules
Identifying restatements or reclassifications
Building consistent “apples-to-apples” time series where possible
This is especially powerful when paired with a data layer that stores normalized tables and metadata by fiscal period.
Thematic research at scale (cross-company)
Once your filings and transcripts are normalized and indexed, the same system can answer thematic questions with evidence:
Which companies are calling out AI capex and why?
Who is seeing supply chain normalization, and what’s the timeline?
Where is pricing power being defended versus conceded?
How is China exposure being framed across a universe?
This is where RAG for financial documents becomes a research multiplier: instead of reading 50 documents, the agent finds the cited passages and assembles a grounded view.
Architecture Patterns That Actually Work (Practical Blueprint)
In high-stakes finance, the most useful systems behave less like improvisational chatbots and more like repeatable workflows with strong controls.
Pattern A — Workflow-first (the default for high-stakes finance)
This pattern is a deterministic pipeline: Ingest → Extract → Verify → Summarize → Diff → Draft outputs
Why it works:
Predictable steps make evaluation straightforward
Outputs can be logged and audited
Failures are easier to diagnose (parsing versus extraction versus summarization)
For AI agents for 10-K analysis, workflow-first designs are usually preferred because they align with how investment teams already work: structured review, clear provenance, and repeatability.
Pattern B — Orchestrator plus specialist agents
As you scale, you’ll often split responsibilities:
Retrieval agent: finds relevant sections and passages
Table agent: extracts and cleans tables, handles units and periods
Accounting-policy agent: flags revenue recognition, leases, and definition changes
Skeptic agent: verifies numbers, checks units, demands traceability
This approach reduces the chance that one model run tries to do everything at once. It also makes it easier to improve a weak component without rewriting the whole workflow.
RAG vs. long-context vs. hybrid
Teams often treat this as ideology. It’s better treated as a practical choice.
Long-context works well when:
You’re analyzing a single document
The task is narrow (for example, summarize Item 7 MD&A only)
You can tolerate higher cost for simplicity
RAG for financial documents becomes necessary when:
Filings are large and you have many associated documents
You need cross-document answers (10-K plus transcript plus deck)
You want to scale across a coverage universe
You need strong retrieval controls and repeatability
Hybrid is often best in production: retrieve the relevant slices, then run deep analysis in a longer context window. Regardless, chunking strategy still matters, because filings are structured documents where section boundaries convey meaning.
Data layer decisions that reduce downstream errors
If you want AI agents for 10-K analysis to be dependable, invest early in the data layer:
Store normalized documents with metadata
Index by section and fiscal period
Keep version history so you know what the agent saw at the time
Store extracted tables as structured objects, not just text
These choices pay off in evaluation, monitoring, and auditability.
Accuracy, Compliance, and Guardrails (What Can Go Wrong)
The fastest way to lose trust in AI agents for 10-K analysis is to ship one confident wrong number into a shared note.
Common failure modes
The biggest issues tend to fall into a few buckets:
Hallucinated numbers or invented comparisons
Mixed periods (Q4 versus FY, current year versus prior year)
Wrong units or sign conventions (parentheses, currency shifts)
Misattributed management quotes (call transcript versus filing text)
Missed restatements or quiet definition changes in non-GAAP metrics
These errors aren’t rare. They’re the default outcome unless your workflow forces verification.
Guardrails asset managers implement
The guardrails that matter most are simple, enforceable rules:
Citation requirement for each numeric claim and key assertion
Confidence thresholds that trigger escalation to human review
Refusal behavior when sources are missing or ambiguous
Tool-call validation and logging for an audit trail
It also helps to log structured intermediate outputs. If a KPI is wrong, you want to see whether the table extraction failed, the unit normalization failed, or the summary layer made an incorrect leap.
Human-in-the-loop review design
Human-in-the-loop investment research is not “a person checks it at the end.” The best designs focus reviewer time where it matters:
What analysts review:
Guidance and key KPIs
Anything tagged as “material change”
New risk factors and legal disclosures
Any extracted number without strong traceability
What can be mostly automated:
Document ingestion and normalization
First-pass tagging and routing
Drafting a structured outline for notes
Populating dashboards in read-only mode
Good review UX patterns include:
Highlight the exact source text for each extracted point
Show extracted tables alongside the original table image/snippet
Display diffs in a familiar redline style
Require explicit approval before publishing or distributing internally
Implementation Plan (0→1 Rollout in 30–90 Days)
Most teams succeed when they start with a narrow pilot and make evaluation part of the build, not something bolted on later.
Phase 1 — Pilot (single sector, 20–50 names)
Pick a domain where the team already has repeatable processes and a clear definition of “good.”
Choose 2–3 must-win outputs, such as:
Earnings brief
Guidance extraction
Change log for MD&A and risk factors
Then build an evaluation set:
Known filings and transcripts
Expected KPI values and key changes
A review rubric for what counts as a correct extraction
Track a baseline:
Time-to-first-draft versus current process
Error rate on extracted KPIs
Percentage of claims with traceability
This makes the value of AI agents for 10-K analysis measurable, not subjective.
Phase 2 — Productionize
Once the pilot works, production is mostly about consistency:
Monitoring for data freshness (did the transcript arrive?)
Alerts for extraction failures or missing citations
Regression tests for common table templates and segment formats
Clear versioning when prompts, parsers, or models change
The goal is not perfection. The goal is controlled reliability.
Phase 3 — Scale across coverage and integrate with systems
Scaling works best when outputs land where people already work:
Research management system or internal knowledge base
Slack or Teams alerts for “material change” events
Read-only integration with portfolio and risk tools first
As scale grows, standardization becomes a competitive advantage. AI agents for 10-K analysis can widen coverage without forcing the team to lower standards.
A quick pilot checklist for an earnings and filings agent:
Defined outputs (brief, guidance, change log)
Normalization pipeline that preserves tables and metadata
Verification rule: every number must have a source
Human review flow with clear approval steps
Evaluation set and regression tests before scaling
Tooling and Vendor Landscape (Build vs Buy)
Most teams end up with a mix: vendor data feeds, internal storage, and a platform layer to orchestrate the workflow.
What to look for in an earnings and filings agent solution
Focus on what actually breaks in production:
Section-aware parsing and table extraction that holds up across issuers
A strong grounding experience so reviewers can verify quickly
Permissioning, audit logs, and retention controls that satisfy enterprise needs
Repeatable workflows, not just a chat interface
Common stack components (agnostic)
A typical build includes:
Document ingestion and parsing (HTML, PDF, transcript formats)
Retrieval and indexing (for cross-document analysis at scale)
Orchestration and tool calling (to run multi-step workflows)
Evaluation and monitoring (to catch regressions and drift)
Where StackAI fits (one practical option)
For teams building AI agents for 10-K analysis, StackAI is often used as the orchestration layer to design repeatable workflows, connect to enterprise systems, and ship an end-to-end agent with governance controls.
In practice, that means teams can:
Build workflows with a visual drag-and-drop builder
Connect internal repositories and knowledge bases for RAG-style retrieval
Add tool use for extraction and downstream actions
Enforce governance with access controls, logging, and approval flows
Deploy with interfaces beyond chat, including forms and batch processing, plus Slack or Teams
This is especially helpful for firms that want to prototype quickly, then harden the workflow with stronger evaluation, monitoring, and review gates as adoption grows.
Conclusion — The New Research Edge Is Speed and Verifiability
The firms getting real leverage from AI agents for 10-K analysis aren’t using them as replacements for judgment. They’re using them to compress time-to-first-draft, expand coverage, and create reliable change detection across filings and earnings materials.
The advantage compounds when you combine three things: disciplined workflows, rigorous data normalization, and human review designed around traceability. Do that well, and you get faster insights without sacrificing the standards your process depends on.
To see how a production-grade workflow for AI agents for 10-K analysis can work in your environment, book a demo: https://www.stack-ai.com/demo
