Job-to-Be-Done Personalization Pipeline
Infer each contact's goal and recommend the closest useful resource at scale
- Difficulty
- Advanced
- Time to result
- ~months to results
- Steps
- 7
- Confidence
- 99%
When a contact converts, combine consented firmographic data, the company's website, the downloaded resource, and subsequent behavior into a concise profile of what that individual may be trying to accomplish. Ask an LLM to describe the ideal course or resource for that job without being constrained by the existing catalog. Use that ideal specification to retrieve several semantically similar assets from a structured or vector-backed content database. Give the candidates and original profile back to the model so it can select the closest useful resource. Finally, generate an email explaining how that resource supports the inferred objective, personalize appropriate details, and measure whether the recommendation advances engagement and conversion.
Origin
HubSpot developed this pipeline for first-conversion nurturing and reported an 82% conversion-rate improvement on Marketing Against The Grain.
Core principles
- 01Useful recommendations matter more than personalized decoration.
- 02Behavior and company context can reveal a contact's likely job to be done.
- 03An ideal resource provides a semantic target even when it does not exist.
- 04A rich content library increases the system's ability to deliver value.
- 05AI makes individual-level personalization economically scalable.
How to run it
- 1
Assemble relevant context
Gather the contact's submitted company information, business URL, triggering content offer, and other relevant first-party behavior.
Pro tip Prefer recent, directly observed signals over speculative demographic assumptions.
Watch out Use only data collected and processed with appropriate consent and privacy controls.
- 2
Infer the job to be done
Ask the model to summarize what the individual is likely trying to accomplish from the combined evidence.
Pro tip Require the model to ground its inference in the supplied signals.
Watch out Treat the output as a hypothesis, not a verified fact about the person.
- 3
Specify the ideal resource
Have the model design the course or resource that would best help complete the inferred job, regardless of whether it exists in the catalog.
Pro tip Describe desired learning outcomes and practical capabilities rather than only a topic label.
Watch out Do not constrain the ideal specification too early to existing inventory.
- 4
Retrieve available candidates
Query a structured or vector-backed content library for the resources most similar to the ideal specification.
Pro tip Store relationships among courses and assets to improve retrieval relevance.
Watch out Poor metadata or embeddings can cause the system to overlook the strongest resource.
- 5
Select against the original context
Return the leading candidates to the model and ask it to choose the resource that best fits the individual's inferred objective.
Pro tip Include both the original profile and the ideal-resource description during selection.
Watch out Semantic similarity alone does not guarantee practical usefulness.
- 6
Generate the recommendation email
Create copy that explains why the selected resource is relevant and how the recipient can use it. Personalize names and company details only where they support the message.
Pro tip Center the email on helping the recipient perform the job rather than showcasing the model's knowledge.
Watch out Overly specific speculation can feel intrusive or inaccurate.
- 7
Validate and iterate
Compare the personalized pipeline against the existing nurturing experience using engagement and conversion metrics. Inspect failures and refine inference, retrieval, and copy separately.
Pro tip Test recommendation relevance independently from stylistic personalization.
Watch out Do not attribute the result to personalized wording without isolating the recommendation mechanism.
In the wild
A person from a small online cold-brew coffee retailer downloads an influencer-marketing resource and later explores content planning. The model infers that she may be preparing seasonal promotions or a product launch, specifies an ideal learning resource, and searches HubSpot's catalog. It selects a content strategy course and generates copy connecting that course to the retailer's marketing objectives.
→ The recipient receives a recommendation aligned with her likely job rather than a generic asset selected for a broad marketing persona.
Common mistakes
Personalizing copy without improving relevance
Customized wording may sound impressive, but the reported step change came from accurately inferring the job and recommending useful content.
Treating an inference as certain
The model is making a best guess from incomplete behavioral evidence, so copy should avoid asserting speculative details as facts.
Neglecting the content library
Even strong inference cannot deliver value when the retrieval system lacks relevant, well-structured resources.
Is it for you?
Best for
It is best for organizations with substantial first-party context, a structured content library, and enough volume to justify automation.
Not ideal for
It is not ideal when data is unreliable, consent is unclear, or the available content cannot genuinely address inferred needs.
From the transcript
“And our hypothesis was if we could use AI to actually guess very accurately what that job was, and then surface the best content in…”
“We believe that AI can actually do one to one personalization at scale.”
“But 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…”
From the episode
The Ai Strategy That Increased Our Email Conversion Rate By 82%