Job-to-Be-Done AI Email Personalization Pipeline
Infer each lead's job and match it to the most useful available content
- Difficulty
- Advanced
- Time to result
- ~months to results
- Steps
- 7
- Confidence
- 99%
The pipeline begins when a lead converts on educational content and supplies signals such as a company URL, email address, company size, and the specific resource requested. An LLM combines those inputs with subsequent website behavior to summarize the job the person is probably trying to accomplish. It then imagines the ideal course or resource for that job without being constrained by the existing catalog. That ideal is sent to a vector database containing available courses and their relationships, which returns similar candidates. The LLM selects the candidate that best fits the inferred job and generates an email explaining how it can help, personalized with relevant person and company details. Performance is validated through open, click-through, and downstream conversion rates.
Origin
Built by HubSpot's marketing technology and email automation specialists to improve its first-conversion nurturing flow. Shared on Marketing Against the Grain.
Core principles
- 01Personalize the recommendation before personalizing the wording
- 02Infer intent from multiple behavioral and company signals
- 03Define the ideal resource before searching the existing catalog
- 04Use vector retrieval to find the closest available content
- 05Let real-user feedback guide model iteration
- 06Measure value through engagement and conversion outcomes
How to run it
- 1
Collect Intent Signals
Gather consented data available when the lead converts, including the requested offer, company URL, company attributes, and relevant website actions. Use multiple signals rather than relying on a single download.
Pro tip Favor recent actions that clarify what the lead is trying to achieve now.
Watch out Do not use sensitive or unexpectedly collected data for personalization.
- 2
Infer the Lead's Job
Ask the LLM to summarize the outcome the individual is likely trying to accomplish. Frame this as a best estimate supported by observed signals, not a known fact.
Pro tip Require the model to ground its inference in the supplied behavior and business context.
Watch out Overconfident intent inference can make the resulting email feel intrusive or irrelevant.
- 3
Design the Ideal Resource
Have the model describe the course or content that would perfectly help with the inferred job, regardless of whether that asset exists. This creates a semantic target for retrieval.
Pro tip Specify the desired outcome and subject matter rather than inventing a catchy title alone.
Watch out Searching the catalog too early can constrain the recommendation to familiar but weak matches.
- 4
Retrieve Available Matches
Send the ideal-resource description to a vector database containing the organization's content library. Return a shortlist of semantically similar assets.
Pro tip Store rich descriptions and relationships for each asset to improve retrieval quality.
Watch out Poorly indexed or outdated content will produce weak candidates.
- 5
Choose the Best Candidate
Give the shortlist, lead summary, and inferred job back to the LLM. Ask it to select the resource most likely to help the individual accomplish that job.
Pro tip Let the model reject candidates when none are sufficiently relevant.
Watch out Do not force a recommendation solely because an asset is commercially convenient.
- 6
Generate the Connecting Copy
Create email copy that explains why the selected resource is useful in the lead's context. Personalize names and company details only where they naturally support that explanation.
Pro tip Make the recommendation's utility the central message.
Watch out Highly personalized wording cannot rescue an irrelevant recommendation.
- 7
Release, Measure, and Tune
Test the system against the existing nurturing flow and monitor open, click-through, and conversion rates. Use real interactions to improve prompts, retrieval, data quality, and model behavior.
Pro tip Inspect recommendation accuracy separately from copy quality.
Watch out Do not optimize only for opens if the recommended content fails to create downstream value.
In the wild
A lead from an organic cold-brew coffee retailer downloaded an influencer-marketing resource and later showed interest in content organization and planning. The system inferred that she might be preparing seasonal promotions or product launches, searched HubSpot's course catalog, and recommended its content strategy course with copy tailored to the coffee business.
→ The recommendation connected observed intent to a concrete next step rather than relying on a broad marketing-persona segment.
HubSpot applied the pipeline to the high-volume email flow used after someone first downloads an educational offer. Compared with its long-optimized cohort-based approach, the AI system generated individual recommendations and contextual email copy at a scale that would have required thousands of human-written messages.
→ Conversion rate improved by 82%, while open rate improved by nearly 30% and click-through rate by more than 50%.
Common mistakes
Personalizing Copy Before Relevance
The team's first hypothesis emphasized personalized email wording, but the major conversion gain came from accurately inferring the job and recommending content that helped complete it.
Treating Inferences as Facts
The system is making evidence-based guesses about intent. Copy that states those guesses too definitively can feel invasive or simply be wrong.
Launching Without Domain Expertise
An AI specialist alone may not understand the nurturing workflow, personas, and automation constraints needed to produce a useful system.
Is it for you?
Best for
It is best for high-volume marketers with first-party lead signals and a substantial library of useful content or courses.
Not ideal for
It is not ideal for organizations with sparse consented data, a tiny content catalog, or high-stakes decisions requiring deterministic recommendations.
From the transcript
“our hypothesis was if we could use AI to actually guess very accurately what that job was and then surface the best content in our…”
“the first step in this process is after a person converts and we see the website we know what they're converting on and we're able…”
“the thing that's really most important is being able to accurately guess what the job is this person is looking to do and then give…”
From the episode
The AI Strategy That Doubled His Email Conversion Rate (Step By Step)