Multi-Step Research Workflows and AI Agents

A single AI prompt is a tool. A chain of prompts that takes a fresh idea and produces a finished research memo is a workflow. The most productive analysts are not the ones with the cleverest single prompts — they are the ones with reliable multi-step workflows that they run weekly.

This lesson covers the chained-prompt workflows that turn AI from a chatbot into a research assistant.

What You'll Learn

• How to chain prompts so each step builds on the prior output
• Four reusable analyst workflows: pre-meeting prep, post-print, screen-to-thesis, and M&A diligence
• How AI agents differ from chained prompts and when each is the right tool
• How to know your workflow is working

Why Chains Beat Single Prompts

A single prompt asks AI to do everything at once. It usually does each step poorly. A chain breaks the work into steps where each one is a focused prompt with focused output.

A simple comparison. To produce an earnings preview note:

• Single prompt: "Write an earnings preview for NVDA Q4."
• Chain:
  1. List the three to five drivers analysts should track this quarter
  2. For each driver, summarize what consensus expects and the range
  3. For each driver, identify what would be a positive surprise vs negative surprise
  4. Identify the data we will get during the call that will resolve each driver
  5. Combine into a one-page preview with the structure: thesis, key metrics, scenarios

The chained version produces a better output. Each step is also short enough that you can review and correct before passing to the next step.

Workflow 1: Pre-Meeting Prep

You have a meeting with management or a sell-side analyst tomorrow. The workflow:

Step 1 — Refresh recent news:

Search for any news, regulatory filings, or research notes on
[COMPANY] in the last 30 days. Summarize the top 5 items by
significance to an investor. Include sources.

Step 2 — Identify open questions from the model:

My current model for [COMPANY] assumes [X]% revenue growth and
[Y]% EBITDA margin in the next 12 months. Which assumptions are
most uncertain and what data would help resolve them?

Step 3 — Generate questions:

Based on the recent news and the open assumptions, generate 10
questions I should ask at tomorrow's meeting. Order by importance.
For each: state what answer would confirm or invalidate my model.
Flag any question that could be perceived as MNPI-seeking.

Step 4 — Anticipate management's likely answers:

For each of those 10 questions, predict the most likely management
response based on their public statements. Identify the question
where their answer would be most surprising.

Output: a one-page meeting prep doc with questions, scoring criteria, and predicted responses. 20 minutes of work for what used to take an afternoon.

Workflow 2: Post-Print Update (Earnings Day)

The first hour after a print is the most valuable. Your workflow:

Step 1 — Compare actuals to model:

Here is the press release and my model output. Compare actuals to
my estimates line by line. Flag any line that diverged by more than
5%. Categorize as positive, negative, or neutral surprise.

Step 2 — Read the transcript:

[Paste full transcript.] Identify changes in management tone vs
prior quarter on: guidance, margins, capital allocation, competitive
environment. Quote exact phrases.

Step 3 — Update the thesis:

Given the surprises and tone shifts, what is the impact on:
1. Our base case revenue and margin assumptions
2. Our 12-month price target
3. Our rating

For each, state the directional impact and what additional data I
need before updating the model.

Step 4 — Draft the client note:

Draft the 250-word post-print update note in our house style. Lead
with the headline takeaway, then the three most important details,
then the implication for our rating and PT. End with what we are
watching.

By the end of step 4 you have a draft note that needs editing and verification — but not authorship.

Workflow 3: Screen-to-Thesis

You ran a screen and found 12 names you want to investigate. Workflow:

Step 1 — Triage the screen:

I screened for [CRITERIA] and got these 12 names. Quickly rank
them by which deserves the deepest dive based on: data quality,
business model clarity, valuation dispersion vs peers, near-term
catalysts. Output as a ranked table with one-line rationale.

Step 2 — Two-paragraph thesis per top 3:

For each of the top 3 names from the prior step, draft a tentative
two-paragraph long thesis: (a) why we might want to own it,
(b) what would invalidate the idea. Use the most recent 10-K and
last 4 earnings transcripts (which I will paste).

Step 3 — Pick the winner:

Across the 3 candidates, which is the strongest pitch for the next
investment committee? Defend your choice in 100 words. Be opinionated.

Step 4 — Build the diligence plan:

For the winning name, build a 2-week diligence plan. Daily output:
what we read, who we call, what we model. End-state: a 5-page
investment memo ready for IC.

The screen-to-thesis flow now takes a day. The pre-AI version took a week.

Workflow 4: M&A and Strategic Review

Corporate development and M&A analysts run this regularly:

Step 1 — Strategic fit:

Our company [DESCRIPTION] is considering acquiring [TARGET]
[DESCRIPTION]. Assess strategic fit on: product overlap, customer
overlap, geographic overlap, capability gaps filled. Rate each on
a 1-5 scale with rationale.

Step 2 — Sizing the deal:

What is the likely range of purchase prices for [TARGET]? Triangulate
using: recent comparable transactions, public market valuations,
private market multiples, target's last funding round if applicable.

Step 3 — Synergies:

List the realistic synergies for this acquisition. Group into:
revenue synergies, cost synergies, and tax/capital synergies. For
each, estimate magnitude and time to realize. Be skeptical — most
synergies are over-estimated.

Step 4 — Risks:

What are the five biggest risks to this acquisition? Cover: integration,
customer attrition, regulatory, talent, and financial risk. For each,
state mitigation cost and likelihood.

Step 5 — IC deck outline:

Outline the 15-slide IC deck for this acquisition recommendation.
Combine prior outputs. Order: strategic case, valuation case,
synergy case, risks, recommendation.

A workflow like this used to take an associate two weeks. With AI it is a focused week of analyst work with much higher quality output.

Chained Prompts vs. AI Agents

Chained prompts mean you run each step manually, reviewing between steps. AI agents try to do all the steps automatically.

In analyst work, chained prompts almost always win. The reason: each step needs verification before the next, and AI agents currently cannot reliably know when their output is wrong. An agent that confidently invented a comp multiple in step 2 will produce a confident report in step 8 — and the bad number will be invisible.

Use AI agents only for steps where errors are low-impact: web search, formatting, document collation. Use chained human-supervised prompts for anything with quantitative output or analytical judgment.

The current AI agent landscape — ChatGPT's deep research mode, Claude's computer use, Perplexity's agentic search — is impressive for research tasks but still requires verification. Treat them as faster first-draft generators, not as autonomous workers.

How to Know Your Workflow Works

A workflow earns its keep when:

1. You can re-run it in under 30 minutes from scratch. If it takes longer, the steps are too big.
2. The final output requires editing, not rewriting. If you keep rewriting from scratch, the prompts need more context.
3. You can hand the workflow to a junior teammate and they get a similar quality output. If they cannot, your prompts are too dependent on tacit knowledge.
4. You actually use it weekly or quarterly. Workflows that sit unused do not justify their setup cost.

Building Your Workflow Library

Treat your prompts like code:

• Keep them in a single document (Notion, Confluence, or a Markdown file in your team Drive)
• Version them — track which prompts changed and why
• Document the prerequisites — what files you need to paste in, what context to load
• Note the typical run time and quality

A team with 5-10 well-maintained workflows out-produces a team with no AI workflows by a factor of 2 or 3, often without using AI for anything especially exotic.

Key Takeaways

• Chain prompts so each step has focused input and output — single mega-prompts produce worse results
• Build four core workflows: pre-meeting prep, post-print, screen-to-thesis, M&A diligence
• Treat AI agents as fast first-draft generators, not autonomous workers
• A workflow works when it is fast to re-run, requires editing not rewriting, and can be handed to a teammate
• Maintain a workflow library — version your prompts like code