AI Prospecting and Outreach SOP

Simple system for bankers and business teams

This SOP shows how to use AI tools to turn a broad market idea into a clean outreach package.

It was inspired by a live bank outreach demo, but the same process can be used for many other goals, such as:

prospecting for new business clients
building local market lists
preparing CRM-ready account briefs
creating outreach campaigns
making sales decks and banker prep materials
organizing research for other industries and territories

This process is for outreach, planning, and relationship development. It is not a credit approval process and it should not be used to make lending decisions.

What this SOP helps you create

By the end of this workflow, you should have:

a researched list of target companies
a source map that shows which sources support which companies
one grounded profile for each strong prospect
a clean folder of files ready for AI analysis
CRM-ready briefs
a campaign plan
outreach messaging
a next steps playbook
leadership-ready documents, slides, and spreadsheets

Tools used in this workflow

ChatGPT to help write strong prompts
NotebookLM to gather and organize research
Google Docs or Notes inside NotebookLM to save working outputs
Google Sheets or Excel to save tables
Codex to turn the research into banker-ready deliverables

Before you start

Make sure you know these five things:

your target industry
your target geography
your outreach goal
the kinds of banking conversations you want to open
where you will save your files

Examples of outreach goals:

business banking
cash management
merchant services
equipment financing
working capital
SBA lending
expansion financing
acquisition financing

The big idea

Do not jump straight from a broad idea to outreach copy.

First, use AI to collect research. Then, organize the sources. Then, create grounded truth files for each prospect. Only after that should you ask Codex to create briefs, campaigns, and banker materials.

This order matters because it helps the final output stay more accurate, more useful, and easier to trust.

Phase 1: Build the research prompt

Go to ChatGPT first.

Ask it to create a research prompt for a deep research agent. Tell it to keep the prompt concise but complete.

Your goal is to create one strong master prompt that you can paste into NotebookLM.

Use a prompt like this:

"Create a concise but comprehensive deep research prompt for an AI agent. I want to research [INDUSTRY] businesses in [LOCATION] for [OUTREACH GOAL]. I want the best prospects in the area, review signals, recent achievements, growth activity, leadership details, and business signals that may matter for outreach."

When ChatGPT gives you the prompt, review it and make sure it asks for:

company name
website
address
phone
type of business
owner or decision-maker, if public
reviews and review count
key praise and complaint themes
recent activity, awards, hiring, growth, or projects
business signals that matter for outreach
a ranked list or score
source links

Phase 2: Run deep research in NotebookLM

Create a new notebook in NotebookLM.

Give the notebook a clear name based on the project.

Open the Deep Research feature and paste in your master prompt.

Wait for the research to finish and for the sources to load.

Then do these cleanup steps:

Click Import all.
Wait for all sources to load.
Scroll through the source list.
If NotebookLM shows failed sources, remove them.
Keep the notebook clean before moving on.

At this stage, your goal is not final outreach. Your goal is to build a reliable source base.

Phase 3: Create a source map

Next, go back to ChatGPT and ask it to create a second prompt.

This prompt should tell NotebookLM to analyze all uploaded sources and build a source-mapping table.

The table should show which sources mention which companies and what each source confirms.

Use a prompt like this:

"Create a concise prompt for NotebookLM that maps uploaded research sources to the companies they mention. I need a table that shows which sources support which prospects, what each source confirms, any review or reputation signals, recent activity, banking-relevant insights, conflicts, and an overall usefulness rating."

Paste the finished prompt into NotebookLM with all sources selected.

Ask NotebookLM to create a table with fields like:

company name
source name
source type
source link
what the source confirms
key details found
review or reputation signals
recent projects, awards, hiring, or expansion
banking-relevant insights
source reliability
notes or conflicts
overall usefulness

When NotebookLM gives you the table:

convert the output to a note
convert the note to a source
export the table to a spreadsheet if possible
save those files in your project folder

This step is important because it becomes your research map for the rest of the workflow.

Phase 4: Find the right sources for each prospect

Now narrow the project from broad market research to company-level research.

In NotebookLM, select only:

the main overview source created by NotebookLM
the new source-mapping table you just created

Then ask NotebookLM which sources should be used for deeper research on each company.

Use a prompt like this:

"Using only the selected sources, tell me which sources are the best ones to use for grounded research on each company so I can create a strong prospect profile."

Save this output as a note.

This note will help you choose the best 3 to 6 sources for each company in the next step.

Phase 5: Create a grounded truth file for each company

This is one of the most important steps in the whole process.

For each strong prospect, manually select only the sources that actually apply to that company.

Do not leave every source selected.

Then use a reusable prompt in NotebookLM to create a grounded profile for one company at a time.

Use a prompt like this:

"Using only the uploaded sources, create a grounded profile for [COMPANY NAME]. Do not use outside knowledge. Do not guess. If something is missing, say 'not found in sources.' Clearly separate confirmed facts from inferences. Include basic information, a source evidence table, reputation summary, growth and business signals, outreach relevance, red flags, open questions, and a short final summary with a confidence rating."

Run that prompt once per company.

For each finished profile:

convert the output to a note
convert the note to a Google Doc or exportable file
save the file in your project folder

By the end of this step, you should have one grounded file for each strong prospect.

Phase 6: Prepare the project folder for Codex

Once your grounded files are ready, create or open a project folder for Codex.

Add these items to the folder:

the main lead spreadsheet
the source-mapping spreadsheet
the grounded company files
any supporting notes you want Codex to use

Markdown files usually work very well because they are easy for AI tools to read, but Word files can work too.

Keep the folder clean. Remove extra clutter if it is not needed.

Phase 7: Add strong system instructions for Codex

Before asking Codex to create banker deliverables, give it clear standing instructions.

The system instructions should tell Codex to act like a bank business advisor for prospect outreach.

The instructions should make these rules clear:

separate facts from inferences
do not invent details
do not make credit decisions
do not use protected class information
create banker-friendly outputs
focus on outreach, relationship strategy, CRM prep, and campaign planning

This step helps the final package stay useful, professional, and compliant.

Phase 8: Ask Codex for the full outreach package

Now give Codex one clear package-building prompt.

Tell it to review the spreadsheet and grounded files, then create a full outreach package.

Your prompt should ask for outputs such as:

executive summary
CRM-ready company briefs
risk analysis
campaign strategy
outreach messaging
next steps playbook
PowerPoint decks
Excel workbooks

It is helpful to tell Codex:

which companies have full grounded files
that it may use subagents if useful
to label missing information clearly
to place weak or incomplete records into a Needs More Research group
to save the outputs into a named folder

Phase 9: Review the outputs before sharing

Do not skip quality control.

Before you use the files in a presentation or send them to a banker, check that:

every strong prospect has a full brief
claims are supported by source material
facts and inferences are clearly separated
messaging sounds professional and not pushy
no one is making a lending approval statement
files are easy to skim
spreadsheets are clean and filterable
decks are simple and executive-friendly

If needed, ask Codex for one clean revision pass.

Phase 10: Publish the SOP and package on your website

After the workflow is done, decide what visitors should download.

For a conference or live demo, a good setup is:

one QR code in the presentation
one landing page on your website
one download button for the SOP
optional extra buttons for sample outputs or a contact form

Keep the website page simple.

The page should answer:

what this SOP is
who it is for
what business problem it solves
what the visitor will download
how to contact you if they want help doing it

Best practices

Keep prompts concise but complete.
Use placeholders like [INDUSTRY], [LOCATION], and [OUTREACH GOAL] so the process is easy to reuse.
Use only company-specific sources when building grounded files.
Save each important output as soon as it is created.
Name files clearly so the folder stays easy to understand.
Treat weak or conflicting source material carefully.
Use banker-friendly language, not technical AI language.

Common mistakes to avoid

asking Codex to do everything before the research is grounded
leaving too many unrelated sources selected in NotebookLM
using long prompts that break or get cut off
mixing confirmed facts with guesses
creating outreach based on thin evidence
skipping the source-mapping step
saving files with vague names
forgetting to remove failed sources

Quick quality checklist

Before you call the project complete, make sure you can say yes to these questions:

Do I have a clean research notebook?
Do I have a source map?
Do I have one grounded file per strong prospect?
Do I have a lead spreadsheet and supporting files in one folder?
Did Codex create banker-ready deliverables?
Did I review the outputs for accuracy and tone?
Is the SOP saved in a format I can upload to my website?

How to reuse this SOP in other industries

The same workflow works for many outreach projects.

You can swap out the industry, geography, and goal.

Examples:

dentists in a target county for treasury and lending outreach
law firms in a metro area for relationship banking outreach
contractors in a region for equipment and working capital outreach
medical practices for merchant services and expansion conversations
local businesses for chamber partnerships or branch growth campaigns

The steps stay mostly the same. Only the prompts, scoring logic, and outreach angle change.

Final reminder

This SOP works best when you follow the order.

Build the research prompt.
Run deep research.
Create the source map.
Create grounded company files.
Move the files into Codex.
Generate the outreach package.
Review the work.
Publish the SOP and share it.

That order helps turn raw information into a process bankers can actually use.