Automating Jobs-to-be-Done Analysis with AI

Jobs-to-be-Done analysis has always depended on what customers say. Teams record interviews and sales calls, read the transcripts and work out the job, its steps and the needs where customers struggle. Done by hand, one round of analysis can take weeks, and two analysts reading the same transcript often code it differently.
Large language models change the cost of that work. A model can read every call a team records and tie what customers say to a fixed structure: the job, the job steps and the customer needs. This post explains what that analysis has to produce, how thrv does it with Evidence, and where people still check the results.
What the analysis has to produce
Automation helps only if the output matches the method. A JTBD analysis produces four things:
- The job: the goal the job beneficiary is trying to achieve, stated without reference to any product.
- The job steps: what the job beneficiary has to do to get the job done, grouped into Plan, Execute, Assess, Revise and Conclude.
- The customer needs: within each step, the outcomes the job beneficiary wants, each written as an action and a variable.
- The evidence: the customer's own words, linked to the job, steps and needs they describe, with a judgment of how hard each one is for that customer.
A model asked to "find insights" in a call will return themes, sentiment and feature requests. Those are not jobs or needs. The model needs the job in front of it, and it needs to know what counts as evidence.
An example: preparing meals for the week
Here is a statement from a customer interview:
I usually start my week by thinking about what we have coming up: any events, late nights at work, kids' activities. Then I check what's already in the fridge and pantry. After that, I browse through my recipe apps and Pinterest for inspiration, trying to balance healthy options with things the family will actually eat. I make a grocery list based on what I decide to cook, but I often end up changing plans when I see what's on sale at the store.
The job is to prepare meals for the family for the week. Its steps include reviewing the week's schedule and taking stock of the food on hand (Plan), buying the groceries (Execute), comparing the plan with what the store has (Assess) and changing meals based on price (Revise). Needs in those steps include:
- Determine the nights with limited time to cook
- Identify the ingredients already on hand
- Determine the meals the family will eat
- Determine which planned ingredients are on sale
The last sentence of the quote is evidence for the need to determine which planned ingredients are on sale. The customer chooses meals before knowing prices, so the plan falls apart in the store. A link like that would carry a rationale and a struggle rating of "difficult", with a short explanation of why.
One interview proves little. The value comes from doing this across every call and seeing which steps and needs customers describe struggling with most.
How Evidence works in thrv
- Add the calls. Upload audio or video recordings (MP4, MP3 or M4A), import recordings from Google Drive, or import transcripts from Granola or Gong. thrv transcribes the recordings. You can also add a quote by hand.
- Sort the call. thrv labels each call as a research interview, a sales call, product feedback or an internal conversation, and you can change the label. Internal conversations give the team context, but thrv does not use them as customer evidence.
- Review the speakers and jobs. thrv decides whether each speaker is the interviewer or a participant, and records a participant's name, company and title only when the call states them. A participant whose email matches a customer record is linked to that customer. thrv also suggests which of your jobs the participants work on. A person confirms the speakers and jobs, or thrv approves the review on its own when the source already identifies the speakers or when you chose to analyze an upload automatically.
- Find the quotes. For each approved job, thrv reads the whole transcript with the job's steps, needs, metrics, root causes, constraints and emotions in view. It pulls the participants' statements that support a point of view about the job, word for word, and links each one to the items it speaks to. Interviewer remarks, small talk and filler are left out. Each quote keeps its speaker and its place in the call, so for a recording anyone can play the moment it was said.
- Rate the struggle. Every link carries a rationale. When the quote shows how easy or hard the work is, the link also gets a rating on a five-point scale, from very easy to very difficult, with an explanation. When it does not, the link has no rating. A question, a routine task or a mention of a competitor is not struggle on its own.
- Analyze earlier quotes against new jobs. Jobs change. When you add a job or rework its steps and needs, thrv can analyze every approved quote against it. It first checks whether anyone on each call works on that job. It then asks, for each quote, whether it is evidence for each item of the job, keeps the strongest links and checks them against the transcript before saving them.
The results sit where the work happens. Each step and need shows the quotes that support it. Each customer record lists the needs that customer found difficult or very difficult. Quotes can be filtered by type of call. And because thrv connects to the AI tools your team already uses, those tools can read the same quotes and transcripts when they answer questions about the job.
What changes, and what does not
Speed is the obvious gain. Calls are analyzed as they come in, not in a research project once a year.
Consistency is the less obvious gain. The same job, the same definitions and the same rules apply to the thousandth call as to the first. Human analysts drift.
The method does not change. Automation makes a sound method faster, and it makes a weak one faster too. If the job is defined as a product category, the quotes will be linked to needs about that product. If needs are written as feature requests, the evidence will support a feature list. Get the job map right before scaling the analysis.
Where people still check the work
- The job. Evidence links only to the job, steps and needs you define. Someone has to confirm that the job is stated independently of any product.
- The speakers. Quotes come only from participants. If an interviewer is marked as a participant, their opinions become customer evidence. In thrv, marking a speaker as the interviewer takes their quotes out.
- The links and ratings. A model can link a quote to the wrong need or rate it too harshly. Your team can reject a quote, add or remove a link and change a rating, and thrv records each correction.
- Who is on the calls. Sales calls come from buyers who are evaluating you, and research interviews come from people you chose to ask. Filtering by type of call shows whether the struggle holds across both.
Doing this with thrv
thrv maps your customer's job, job steps and customer needs, links what customers say to them and shows where competitors fall short. It then connects that data to the AI tools your team already uses. To see the job for your own customer, get your free market analysis.
Get Jobs-to-be-Done in your inbox
Practical JTBD ideas and case studies from the thrv team, sent to your inbox.