iCoT - a new technique to employ AI collaboratively in investment research workflows

This is how to set up prompts that kick-off machine + human collaborative research work, that gets the most out of AI

The following post is part of our Best Practice AI Investment Research series that teaches how both individual investors and professionals buy- and sell-siders can employ AI seamlessly into your fundamental process driven investment workflows. Other primers can be found under the Techniques tab of our Substack site.

---

Human + Machine Q&A

Investing is an experience game, as is investment research. You can’t expect a new analyst or investor to perfectly analyze unfamiliar situations - and neither will your AI. AI models are trained on pattern matching from prior data. This makes them excellent at repetitive tasks with a defined range of potential outcomes. Investing however, is far more complex with ever changing situations, and with slight nuances in the context potentially being very important to the conclusion. No two investment situations are ever exactly alike, even if they rhyme.

Research has shown that the best outcomes, at least with the current generation of models, are where human and machine work in concert to bring the learned experience of the investor to bear in guiding the AI’s advanced reasoning and processing capabilities. The key is not to treat artificial intelligence as a magic box that produces infallible answers, but as an assistant who is highly capable, tireless but sometimes prone to error and whose true potential depends on how you guide its work.

Most users of AI aren’t aware that you can set up prompts to automate investment research (or any other) tasks through a pre-defined Q&A between the AI and yourself, executing a structured workfow.

This is a more advanced, guided form of chain of thought prompting that allows an investor to set up complex but repetitive investment workflows into a framework where, instead of just battling through yourself as you would have done in the past, you are assisted at every step by AI. You retain the same flexibility of analysis direction because the process is interactive and iterative. However, you are assisted by a neural network brain that speeds the work, checks for information you may have missed, reasons in a bias-free way to clean up preconceptions and executes diligently and fast - all the while coming back to you for confirmation and guidance every step of the way.

You are the conductor and the AI is your orchestra.

The Problem With “Single-Pass” Prompts

Most investors today use AI in the simplest way: by giving a single instruction - “Summarize this 10-K” or “Explain Microsoft’s valuation risk”. The worst I’ve seen is “Give me 5 stocks to buy”! This resembles asking a junior analyst to write an entire report, sight unseen, with only the most basic guidelines and only checking once they place it on your desk.

Such methods have obvious weaknesses:

• Shallow reasoning. A one-shot prompt produces a surface-level answer. The AI does the “doing” while you do the “thinking.” If the model misses nuance, you don’t notice until the end or its hidden in detail that leads to a sub-optimal conclusion.

• Do-overs. When errors appear, you must repeat the entire task, wasting both time and context.

• No learning loop. Your domain expertise, that is, knowledge of competitor updates, real-time information that is not online, management credibility, sector dynamics, or local regulation, never makes it into the analysis unless you explicitly spell it out.

Chain-of-thought (CoT) prompts are a partial improvement. They guide the AI to structure its reasoning, taking into account insights built at each stage but the reasoning is still one-sided: all machine, with some but still limited human course correction midstream.

Working With an AI Assistant

Contrast this with how you might train a junior analyst. You rarely give them the entire assignment and vanish. Instead, you:

1. Set up a process for them to follow.

2. Review progress at each step.

3. Correct assumptions as they arise.

4. Add your judgment where it matters most.

This iterative supervision ensures the final report reflects both the analyst’s labor and your experience at the first pass.

AI can be used in precisely this manner - through interactive Q&A prompts. Instead of issuing one sweeping command, you design the model to interview you while following a set of broad steps in the pre-defined task. At each step, the model does the processing however presents the output to you for confirmation and correction before proceeding. It allows you both to check assumptions, and refine outputs step by step, virtually eradicating hallucination risk if the investor is as diligent as they would be guiding their assistant.

An Example: Consensus Forecast Achievability

Consider a stock approaching its next earnings report. A common task for an investor is to review consensus forecasts as they may have shifted and make sure confidence remains of a meet or beat. If you cover 20 or more stocks, all of whom report within a limited timeframe, this is a all-consuming task. However with AI and an interactive CoT prompt (I refer to as “ iCoT”), this is much faster and perhaps even more reliable.

Here’s a simplistic iCoT prompt set built to test whether Robinhood ( HOOD 0.00%↑ ) could realistically meet its next earnings consensus. The AI didn’t simply return a conclusion. Instead, it asked:

• Which stock are we analyzing?

• Which quarter and what revenue/EPS estimates?

• Can you confirm these historical data?

• Do you want me to analyze for seasonality patterns?

• Lets calculate what needs to be true…

• Should we compare this to news, analyst reports and operational update data for contextual consistency?

At each stage, the investor provides any necessary corrections and context. The machine supplies structure, added reasoning and processing. The result is not the AI’s analysis nor the human’s alone, but a synthesis of human experience and neural network capabilities - the best of both worlds.

This technique is ideally suited for sophisticated analysis workflows:

• Assessing earnings

• Valuation tasks

• Financial Modeling

• Identification of new ideas

Setting Up an Interactive Q&A Prompt

Creating such a workflow is straightforward:

1. Define the end goal.
   Example: “Assess the achievability of the next consensus forecast.”

2. Break the task into stages.
   Example: Identify stock → Confirm period → Gather consensus estimates and historical data → Synthesize recent news, reports and company updates → Compare to what must be true → Draw conclusion.

3. Have the AI interview you at each stage.
   The prompt is structured so the model must ask questions before proceeding. For instance:

   “First, ask me for the stock name. Next, ask me to confirm the earnings period and consensus estimates. At each step, confirm my response before moving forward. Do not skip stages.”

4. Insert reasoning checkpoints.
   Require the model to summarize assumptions back to you.
   Example: “Repeat the revenue and EPS targets you have recorded, and ask me to confirm before analysis.”

5. Allow corrections without a full restart.
   If estimates change midstream, the AI should re-run only the relevant step.

In effect, you have built an assistant workflow, not a one-shot, hit or miss, answer bot.

The Payoff

An investor who simply queries the AI receives an answer that may or may not withstand scrutiny. An investor who works with the AI gains:

• Higher accuracy through real-time correction.

• Deeper integration of judgment (information recency, management credibility, competitive pressures).

• Efficiency, since mistakes are corrected along the way, not at the end.

• Transferable prompts, reusable across companies and quarters with minimal adjustment.

In this technique, the investor doesn’t abdicate judgment, they structure the process and work collaboratively. AI becomes most valuable not when it replaces your reasoning, but when it augments it through disciplined interaction.

---

A Simple iCoT Template

Here is a bare-bones template you can use as the system prompt to kick-off almost any investment workflow set-up:

You are my investment research assistant.  
Do not complete the full task in one go.  
Instead, guide me step by step with questions.  
At each step:  
1. Ask me for the needed input.  
2. Summarize back my response for confirmation.  
3. Wait for me to confirm or correct before moving forward.  

The overall task is: [INSERT YOUR GOAL, e.g., “Assess the achievability of the next earnings consensus forecast”].  

Break this task into stages. For each stage:  
- Ask me what you need.  
- Confirm and reason briefly.  
- Only then move on.

The stages of the task to go through are:
[INSERT NUMBERED STAGES]  

Do not assume missing information. Always ask me first.  
At the end, provide a structured analysis and conclusion that incorporates all confirmed inputs.  

Think of this as a checklist rather than a blank sheet of paper. AI walks with you through each stage, you correct them as needed, and together you build the final analysis.

Worked Example: Interactive Prompt for Forecast Achievability

Here’s how you could set up an interactive Q&A prompt with AI to test whether a company’s next consensus forecast is realistic.

The Prompt

You are my investment research assistant.  
Our task: Assess whether the next consensus revenue and EPS forecasts for a company are achievable.  

Do not complete the task in one go.  
Instead, guide me step by step, asking questions and confirming my answers before proceeding.  

Stages:  
1. Ask me for the stock ticker and/or company name.  
2. Search for the next earnings reporting date. Ask me to confirm or change the earnings period (quarter/year) I want to analyze.  
3. Search for the last four quarters revenue and adjusted (non GAAP ) EPS. Tabulate and check for confirmation of changes needed.
4. Search for the next consensus forecast revenue and EPS estimate from reliable sources. Append to table and ask for confirmation or adjustment.  
5. Summarize back these estimates and confirm with me.
6. Generate contextual update information: Gather, summarize news, analyst reports, operational update information from the company, company guidance and the financial impact of any acquisitions/divestments for the period and relevant macro data updates from the last 8 weeks. Synthesize a view of the signals from this information with reference to whether the environment has changed since the last earnings report. Ask whether I wish to add anything to the consideration? 
7. Seasonality: Analyze the stock for evidence of any repetitive quarter to quarter or period to period seasonality in revenues or margins. Summarize the seasonality impacts as they apply to the next earnings period.
8. Add that information and update your synthesis of the current environment and what it implies for the company's growth.
9. Calculate and add to the table, "what must be true" to hit consensus targets (YoY growth, QoQ growth). Append that to the table and show it in comparison with the recent YoY and QoQ growth achieved in recent quarters.  
10. Compare company guidance (if any) with analyst estimates, what must be true growth, recent growth rates and the contextual synthesis in a logical consideration of whether the targets appear to be able to be met, missed or exceeded. 
11. State your confidence and the most important inputs into your reasoning that drives this confidence level. 
12. At the end, provide a structured conclusion:  
   - Required growth rates vs historical trends  
   - Operational feasibility  
   - Probability that the forecast is met or missed  
12. Finally, ask if there is anything else the investor wishes to add to the analysis and update the report accordingly each time with future steps.

Do not assume missing information. Always ask me first.  

How It Plays Out in Practice (Indicative only)

AI: What stock are we analyzing today?

You: Robinhood, (HOOD).

AI: The next earnings period is Q3 2025. Reporting date is estimated as 29 October? Confirm?

You: Yes.

AI: I have found the following reported results for the last four quarters [TABLE]. Are you happy with these?

You: Change the EPS estimates to non GAAP as follows [ state period and EPS from database].

AI: I have not been able to find consensus estimates for Q3: Please provide?

You: Revenue $1,150 million; Non-GAAP EPS $0.51

AI: [TABLE UPDATED] Confirm the data before we proceed.

You: Confirmed.

AI: The following is a summary and insights from all updates, relevant news and analyst reports from the last 8 weeks I could find. They highlight the following trends…Do you wish to add any thing else to the consideration?

You: No.

AI: Understood. I have calculated the growth rates that must be true in order to achieve consensus forecasts for Q3 and added them to the table. Any further changes?

You: No.

AI: Consensus requires ~80% YoY revenue growth; EPS requires 95% YoY growth. On a QoQ basis, revenue requires 16% growth and EPS 2%. Based on the company’s operational update from August 2025, the following indicators were reported. [summary of indicators].

Year over year growth in revenue drivers such as assets under custody, growth in the margin book, securities lending and crypto trading volumes plus the Bitstamp acquisition impact, all support the required YoY growth rates as they are above 80%, in some cases materially.

On a QoQ basis, it looks more challenging. August was a soft trading month and it may require September to be stronger in order to achieve the 16% QoQ growth in revenues. Analyst forecasts have been revised higher since the August update and the stock has responded positively, both of which suggest that the YoY comparison may be more reliable as a basis to assess achievability.

The 80% YoY revenue growth needed for Q3 however represents an acceleration compared to the last 2 quarters (50% and 45% for Q1 and Q2 respectively). This is true also for EPS targets.

Conclusion:

• Revenue target ($1,150m): Requires ~80% YoY growth, above current trajectory (~45-50%) but supported by recent operational KPIs and the Bitstamp acquisition which should add approximately $25m to $30m revenue in the quarter (+5% incremental revenue growth on YoY and 3% on QoQ basis). QoQ comparisons and August trading suggests a more difficult target however my be less reliable / more noisy. Revenue appears challenging but achievable given estimate trends in the market but with limited upside given revised expectations.

• EPS target ($0.51): Achievable compared with the prior quarter (only 2% QoQ required despite revenue growth and operating leverage). Historically, HOOD’s incremental margins at the net income level have been 45% or higher. There appears more upside to EPS estimates than revenue.

• Assessment: Forecasts appears achievable although revenue remains a stretch target. Confidence is increased due to the analyst response to the August update. Estimates are fresh, not stale. Markets have been strong. Probability of miss <30%.

AI: Are there any other considerations you wish to build into the analysis?

---

Closing Thought

I encourage you to spend time experimenting with this technique. At The Inferential Investor we are setting up sophisticated and professional iCoT prompts to run financial modeling, valuations, consensus checks and even new idea identification. These will all be available via the prompt library for you to just drop into your chose AI model when you have a task to complete.

As always,

Inference never stops. Neither should you.

Andy West

The Inferential Investor