River
Y CombinatorBacked by Y Combinator
FREE TEMPLATE

Jobs to Be Done Interview Analysis

Every job statement arrives with the interviews behind it, the interviews that contradicted it, and the total, so nine of fourteen stops reading like one.

Free download  ·  No account needed

Fourteen interviews go in and a one-page portrait comes out. The trait eleven people described and the trait one participant described at length are in the same box, in the same font, and nothing on the page records which was which. That is usually sold as clarity. It is the reason a persona cannot be argued with, and Nielsen Norman Group's own account of why personas fail leads with the failure it produces: personas that were created and then never used.

So every job statement in this pack carries three numbers: how many interviews described the job, how many described doing something else, and out of how many. Nothing gets deleted for being unpopular. A statement at 1 of 14 stays, labelled 1 of 14, which is all a roadmap conversation needs. Statements are written as jobs rather than as features on purpose, because the underlying job does not go away when the product that served it does.

Then it writes the anti-persona: who the product is explicitly not for, by segment, with the evidence, which is the only artifact in this space that can actually decline a request. Coding the transcripts first is research synthesis, unsolicited feedback out of a support queue is feedback triage, and the jobs that survive get ranked against everything else you owe in the prioritization pack. Or download the whole pack as Word and CSV files.

The study, with the denominator left attached

Every count in Evidence Register traces to interview ids in Interview Index. Support Rate and the airtime ranks are computed, not entered.

Evidence Register

Illustrative rows for a fictional scheduling and payroll product for independent restaurant groups, Harborline. Fourteen interviews: five single-site owner-operators, six small groups of two to six sites, three enterprise chains above eighty sites.

IDJob statementSupportAgainstRateTarget segmentsMinutesAirtimeVerdict
S-01Cover a shift when someone calls out without triggering overtime13 of 14093%11 of 11313rdCore job
S-02See today's labour cost against today's sales before the shift ends11 of 14179%11 of 11264thCore job
S-03Show an employee exactly how their tip pool number was calculated9 of 14064%8 of 11225thCore job
S-04Get a new hire onto the schedule in the same week they are hired8 of 14257%6 of 11147thCore job
S-05Reconcile the published schedule against the hours actually worked7 of 14150%5 of 11118thSecondary
S-06Forecast next week's labour need from the same week last year3 of 14621%0 of 11382ndEnterprise only
S-07Apply one labour standard across every location3 of 14421%0 of 11441stEnterprise only
S-08Sync completed hours into an HRIS through an API2 of 14014%0 of 11196thEnterprise only

The two statements that took the most airtime, 44 and 38 minutes, have three supporters each and zero in the segments this product serves. Together with S-08 they took 101 of the 205 minutes this study spent on jobs. Half the study, from three participants who all arrived through inbound demo requests rather than the customer list. That is the roadmap the readout would have produced by accident.

Segment Behaviour Comparison

Each statement checked against what the segment actually does. Interviews say what people report; this says what they run.

BehaviourSingle site
1,840 accts
2 to 6 sites
610 accts
80+ sites
4 accts
TestsReading
Shift swaps per site per week, median1192S-01Confirms. The core job happens five times more often per site in target accounts.
Opens labour against sales daily64%58%8%S-02Confirms. Opened more often than the schedule itself.
Tip disputes logged per site per quarter2.41.90S-03Confirms. Enterprise tip handling sits in a payroll system this product does not touch.
Days from hire to first shift, median3619S-04Confirms, and explains the two contradictions: 19 days is a central onboarding queue.
Has ever opened the forecast view4%9%100%S-06Contradicts. Every account using the feature is in the excluded segment.
Has a written labour standard1%4%100%S-07Contradicts. The job title that owns one does not exist below forty sites.
Has ever issued an API token0%0.3%100%S-08Contradicts. Two of 2,450 target accounts, each once.

A feature at 100 percent reach in one segment and single digits in every other is not adoption, it is a segment boundary. All three of those rows land on exactly the statements the anti-persona note goes on to name, which is what makes declining them a count rather than an opinion.

Statement Candidates

Every candidate drafted, including the ones that did not survive, with the test each failed and where the job underneath went.

CandidateRaised inFailed testWhere it wentStatus
Have better visibility into the business9 of 14Not falsifiableThe most-mentioned phrase in the study, and no two people described the same task. Split into S-02 and S-05, which are what they described when pressed.Rejected
Stop using WhatsApp for shift swaps7 of 14Describes the workaroundRecorded as the incumbent solution on S-01. Seven people building the same workaround is stronger evidence than seven people requesting a feature.Rejected
Reduce staff turnover6 of 14Outcome, not a jobNobody performs it. Attached as the expected outcome on S-03 and S-04, where participants drew the connection themselves.Rejected
Integrate with Toast5 of 14Names a productThe job underneath is S-05. Toast recorded separately as the system of record five of eleven target accounts use.Rejected
Have a better mobile app4 of 14Solution, not a jobTrue, and a design constraint rather than a statement. Passed to design with the 88 percent mobile session share attached.Rejected
Comply with predictive scheduling ordinances2 of 14ConstraintBoth participants operate in the same city, which makes it a constraint on two accounts rather than a job in the market.Kept as constraint
Cover a shift when someone calls out without triggering overtime13 of 14Passed all fourS-01. Verb, object and context, no product named, same task at the same moment in everyone's week.Promoted

The top row is the finding most synthesis loses. Nine of fourteen participants said a version of have better visibility, which makes it the loudest signal in the study and the one thing nothing could be built or declined on. Keeping the rejection in writing is what stops it coming back as a theme next quarter.

What is in the pack

01

Evidence Register

Every job statement with support, contradictions, the study total, target-segment support as its own column, transcript minutes, and the interview ids behind each count so any row can be checked against the transcript.

02

Interview Index

One row per session with the column research repositories omit: how the participant was recruited. Customer list, inbound demo, churned account, panel. It is the most predictive field in the study, because a mixed channel almost always means a mixed segment.

03

Statement Candidates

Every candidate drafted, including the rejections, with the test it failed and where the job underneath went. Half of these are the phrases a readout would otherwise have led with.

04

Segment Behaviour Comparison

Each statement checked against usage data by segment: job frequency per unit, reach of the feature that already serves it, time-to-outcome, and whether the segment even has the prerequisite the job assumes.

05

Job Statements and Persona Briefs

The write-ups. Statements carry their counts and their incumbent workaround. Briefs carry jobs in support order, populations, and an acceptance criterion instead of a personality. No names, no stock photos, no invented quotes.

06

Anti-persona Note

Who the product is not for, by segment, with the interviews, the account count, the behavioural separation, and the specific evidence that would reopen the question.

How it works

  1. 1

    Send the transcripts

    Dovetail, Otter, Descript, Lookback, Grain or plain text, one file per session, plus usage data by segment if you have it.

  2. 2

    Draft wide, cut hard

    Candidates get tested four ways: job or outcome, job or solution, falsifiable or not, and whether the phrasing came from the participant or the guide.

  3. 3

    Count every statement

    Supported by, contradicted by, out of, and target-segment support split out, with transcript minutes recorded next to prevalence.

  4. 4

    Name the boundary

    Two or three personas the evidence supports, and the anti-persona that gives the set the power to decline something.

Frequently asked questions

How is this different from a persona template?

A persona template gives you boxes to fill. Every attribute you write into one carries the same implied confidence, so the trait eleven people described and the trait one described render identically. This pack keeps the count against every statement, which is the single thing that turns a portrait into something a room can check.

What does the anti-persona actually do?

It lets the persona set decline a request. Saying the product is for independent restaurant groups sounds like a decision and functions as none, because every inbound request comes from somebody who runs restaurants. The anti-persona names the segment, the account count, what they use instead, and what evidence would reopen it.

Why record how long each job was discussed?

Because a synthesis session run from memory weights airtime, and airtime is close to uncorrelated with prevalence. In the worked example the two statements with the most minutes have three supporters each, while the job thirteen of fourteen people described ranks third. Showing both columns is what stops one articulate participant setting a roadmap.

What happens to a statement only two people described?

It stays, labelled 2 of 14, with the interviews named. Nothing is deleted for being unpopular, because a two-person statement can be an early signal or a segment finding, and the count is what lets you tell later. What it cannot do is arrive on a roadmap looking the same as a thirteen-person statement.

Do I need usage data as well as interviews?

The counts work without it and get much sharper with it. Behaviour is what separates a job people report from a job people perform, and prerequisite measures settle the most arguments: a job that assumes a central labour team cannot be a job in a segment that has no central team. Send whatever exists.

How many interviews is enough?

Enough that the denominator means something, which in practice is around a dozen per segment you want to make claims about. The register tells you where you are rather than asserting a threshold: it shows which interview each statement first appeared in, so you can see whether the last three added anything new. The continuous discovery pack turns that into a running rate.

Does it replace Dovetail or my research repository?

No, and it starts where they stop. A repository stores and tags sessions. What is missing between the tags and the readout is the counted register, which no repository builds because it requires reading every transcript for one statement at a time. Code the corpus first with research synthesis if you want the extracts too.

Find out which findings survive a count

Send the transcripts however they exported. The first number back is how many statements have unanimous support in your target segment, and it is usually fewer than four.

Count my interviews