eazyware
Playbook·June 17, 2024·10 min read

Investment research AI: copilots for analysts

Earnings transcript analysis, filings summarization, thematic research. The workflow patterns that actually save analyst time vs the demos that don't.

KR
Kushal R.
Engineering lead

Investment research copilots are one of the clearer wins for AI in financial services. Analysts spend significant time on information-gathering work that AI does faster without losing quality. The key is matching the copilot to the analyst workflow, not building a separate tool the analyst has to context-switch to. This post is the copilot patterns that stick and the ones that get abandoned.

Analyst workflow
Investment research copilot workflow Analyst ask question Retrieval filings + calls Synthesis cited draft Analyst review edit + validate Note ship Saves hours per note when done right All claims cited to source documents (filings, transcripts, datasets) Analyst reviews and owns the final — AI is the research assistant Never auto-publish; analyst sign-off required for every note Disclosure language: AI-assisted research (regulators increasingly require)
Analyst question retrieves filings and transcripts, synthesizes cited draft, analyst reviews and ships note. AI is the research assistant; analyst owns the conclusion.

What analysts actually do

Sector analysts cover 15-50 companies. For each: quarterly earnings reviews, model updates, thematic notes, client calls, conference presentations. The job is 50% information gathering and 50% judgment.

AI can materially compress the first 50%. Read transcripts, pull financial line items, summarize competitor commentary, identify theme shifts. Analyst focuses on judgment and client relationships.

AI cannot do the second 50%. Judgment, client trust, professional accountability for recommendations — these remain human.

Copilot use cases that stick

Earnings season preparation. For each company reporting, copilot pulls last quarter's transcript, current quarter's consensus estimates, historical guidance vs actuals. Analyst walks into earnings call prepared in 15 minutes instead of 90.

Thematic research. What are semiconductor companies saying about inventory? Copilot scans recent transcripts, filings, industry commentary. Produces cited summary. Analyst validates, edits, builds their thesis.

Competitor analysis. Pull comparables, normalize metrics across companies, highlight relative strengths. Frees analyst to focus on why rather than what.

Meeting prep. Before client calls, copilot reviews prior conversations, current portfolio, recent research published. Analyst doesn't re-discover context.

Use cases that get abandoned

Autonomous research generation. Copilot produces a full research note. Analyst doesn't trust output, rewrites from scratch, ends up using copilot for background only. The autonomous framing is wrong.

Stock picking. Copilot recommends buys/sells. Analyst's reputation is on the line; they can't stake it on AI recommendations. Rejected immediately.

Replacement of human client interactions. Clients want relationships with named analysts; AI-generated commentary doesn't have the same weight.

Architecture that works

Always-cited output. Every claim the copilot makes has a source link. Analyst can verify immediately. Unsupported claims erode trust fast.

Inside the existing workflow. Copilot integrates with Bloomberg, FactSet, Office, email — wherever analysts already work. Separate destination tools get abandoned.

Short responses over long. Analysts have 50 things to do in a day. A 2-paragraph summary they can glance at is more useful than a 10-page analysis they won't read.

Multi-source retrieval. Transcripts, filings, news, alternative data, internal research. The more sources the copilot can pull from, the more value it adds. See RAG patterns and hybrid search.

Regulatory considerations

Research reports published externally must comply with SEC/FINRA rules. Disclosures: if AI was material to the research, disclose. Data sources, model used, human review process.

Many firms treating AI output as draft, with human analyst taking full authorship and responsibility. Regulatory position is still evolving; watch for guidance specific to AI-assisted research.

MiFID II in EU specifically mentions AI in research. Firms operating across jurisdictions need to satisfy the strictest applicable standard.

Rollout patterns

Pilot with one sector team (tech, healthcare, energy). 3-4 analysts using the tool daily. Measure: time to first draft of earnings notes, meetings per week, client satisfaction signals. 8-12 week pilot before wider rollout.

Training matters. A 2-hour session covering copilot capabilities, limitations, best practices. Analysts who understand the tool use it well; analysts who figure it out alone often use it poorly.

Feedback loop. Analysts feedback on quality, missing capabilities. Weekly review in pilot phase. Monthly in production. Tool improves over time.

Vendors vs build

Vendors: AlphaSense, Bloomberg AI, Kensho (S&P), Factiva (Dow Jones). Mature category. Most firms start here.

Internal build for proprietary analytics. Firm-specific methodology, custom data sources, proprietary research IP all live in internal builds.

Hybrid common: buy commodity capabilities; build differentiated ones. Most banks and asset managers run both simultaneously.

Read next
Capital markets AI: trading desks, research, and risk
Read next
Wealth management AI: where it ships and where it does not
Read next
B2B sales copilots that salespeople actually use
Tags
researchequity researchanalystscopilots
/ Next step

Want to talk about this?

We love debating this stuff. 30-minute call, no pitch, just engineering conversation.

~4h
avg response
Q2 '26
next slot
100%
NDA on request