Data-and-Prompt AI Prospecting System
Combine intent data with modular prompts to scale personalized sales outreach
- Difficulty
- Advanced
- Time to result
- ~months to results
- Steps
- 7
- Confidence
- 98%
The system begins by separating inbound demand from pure cold outbound because their signals and conversion dynamics differ. It then chooses the appropriate channel design: smaller accounts may be served through automated email, while larger accounts generally require coordinated email, calling, and social outreach with a salesperson in the loop. High-quality intent, firmographic, behavioral, and conversational data provide the context for personalization. Instead of relying on one monolithic instruction, the team creates modular prompts for subject lines, introductions, body copy, and calls to action. Those prompts are reverse-engineered from successful representatives and refined against real outcomes. Humans review prompt performance and exceptions rather than approving every word, allowing the system to increase coverage while preserving a deliberate quality-improvement loop.
Origin
Extracted from Marketing Against the Grain, where Kieran Flanagan described HubSpot's extended experimentation with AI-assisted inbound prospecting and meeting booking.
Core principles
- 01Generic data produces generic outreach
- 02Inbound and outbound prospecting require different designs
- 03Channel choice should reflect account complexity
- 04Prompts are production assets that improve through iteration
- 05Automate execution while reviewing system-level quality
How to run it
- 1
Classify the demand source
Determine whether the person has produced an internal signal or is a completely cold outbound prospect. Design separate workflows for the two cases.
Pro tip Begin with inbound demand because it offers stronger signals and was described as the larger opportunity.
Watch out Combining inbound and outbound prospects obscures the context AI needs.
- 2
Choose the channel pattern
Use automated email where a simple buying process supports it. For larger organizations, coordinate email, calls, and social contact while keeping a representative involved.
Pro tip Treat company complexity as a routing signal rather than assuming email alone is universal.
Watch out The hosts reported that only a small share of mid-market and corporate meetings were booked through email alone.
- 3
Assemble differentiated context
Provide the system with behavioral triggers, account details, buyer needs, and other relevant data. Exclude data that does not improve the sales decision or message.
Pro tip Test conversion changes as additional data sources are introduced.
Watch out Base-level enrichment usually produces outreach that resembles every other automated message.
- 4
Reverse-engineer top representatives
Study emails and sequences from successful sales representatives to identify their structure, tone, judgment, and calls to action.
Pro tip Use real winning messages as evaluation examples.
Watch out Do not merely instruct the model to sound personalized without defining what effective personalization looks like.
- 5
Build modular prompts
Create distinct prompts for subject lines, introductions, message bodies, calls to action, and supporting channel actions. Test each module independently before combining them.
Pro tip Modularity makes weak components easier to diagnose and replace.
Watch out A single oversized prompt makes failures difficult to attribute.
- 6
Iterate with operators
Review outputs and outcomes with the sales team, annotate quality issues, and revise prompts repeatedly. Allow a longer learning horizon when results are improving slowly.
Pro tip Evaluate prompt versions against meeting conversion, not writing preference alone.
Watch out A rigid six-week deadline may kill a workflow before its prompts mature.
- 7
Scale with bounded autonomy
Let AI generate and send appropriate outreach while humans focus on prompt quality, exceptions, calls, and qualified conversations.
Pro tip Provide representatives with generated emails, call context, and suggested social outreach.
Watch out Requiring approval of every word removes much of the productivity benefit.
In the wild
The HubSpot team iterated with sales representatives for six months before its AI workflow began outperforming the traditional process. Once the data and prompts improved, it reported sustained gains over the following eight to ten months rather than treating the initial slow progress as proof that AI prospecting could not work.
→ Email-only inbound prospecting reportedly increased month-on-month meetings by about 35%.
An inbound prospect from a 200-person company views pricing and requests a guide. AI drafts a contextual email, prepares call notes, and suggests a social touch. A representative checks the account context, makes the call, and edits only the parts requiring judgment.
→ The representative covers more accounts without reducing the interaction to generic automated email.
Common mistakes
Personalizing with generic data
Basic enrichment inserted into a standard prompt produces messages that sound like every other AI-generated approach.
Using email alone for complex accounts
Larger buying processes often require calls and social contact, so email-only automation leaves potential meetings unconverted.
Ending experiments too early
The described program needed six months of prompt and workflow iteration before it showed gains over the traditional process.
Is it for you?
Best for
It is best for organizations with meaningful inbound signals, reliable customer data, and enough outreach volume to support iterative improvement.
Not ideal for
It is not ideal for low-volume teams without usable intent data or for organizations unwilling to tolerate a sustained learning period.
From the transcript
“the majority of companies I talked to are failing with AI for prospecting and the reason they're failing with AI for prospecting is because the…”
“And then prompt is your product.”
“Break it into multiple prompts.”
From the episode
AI Just Broke Marketing (And What You Need to Do Now)