A prospect went quiet mid-process. The reasonable assumption was a lost deal.
The actual reason was a previous vendor. That vendor had sat in his company's internal interviews and then used what they heard to sell to a competitor in the same industry. Same sector, same operational problems, and his team's honesty had become somebody else's commercial advantage.
So when a new provider arrived asking to listen and to observe, he was not protecting data. He was protecting people who had trusted him once and been exposed for it.
The risk is commercial before it is technical
Most security reviews of a discovery vendor examine encryption, access control, hosting and certification. Those matter and they are not where this failure occurs.
What the previous vendor did required no breach. The data was accessed legitimately, under a signed agreement, by people who were supposed to have it. The question that was never asked, and therefore never answered, concerned what they were permitted to do with what they learned.
That gap sits between two things buyers usually check. A confidentiality clause typically covers the disclosure of information. A data processing agreement typically covers the handling of personal data. Neither necessarily covers a vendor developing sector insight from your operation and selling it as expertise to the company you compete with.
| Risk | Covered by a standard NDA | Covered by a standard DPA | Usually unaddressed |
|---|---|---|---|
| Disclosure of your documents | Yes | Partly | No |
| Breach or unauthorized access | Partly | Yes | No |
| Personal data misuse | Partly | Yes | No |
| Re-identification of individuals | No | Partly | Frequently |
| Aggregate sector insight sold onward | No | No | Yes |
| Your operation used as a sales reference | Partly | No | Frequently |
| Interview content used to train a model | No | Partly | Frequently |
The three rows at the bottom are the ones that produce the situation above, and all three are addressable in a contract that most buyers do not ask for.
Why employee interviews are a special category
Process interviews produce a type of record that is unusually sensitive and unusually easy to underestimate.
The content is candid by design. A discovery conversation works when people describe what actually happens, including the workarounds, the steps that should not exist and the decisions they disagree with. That candour is the deliverable, and it is also an account of the organization's weaknesses in the words of its own staff.
It is attributable even when anonymized. In a function of nine people, a description of one role's daily work identifies the person. Aggregate reporting does not solve this in small teams, and small teams are where the most specific findings come from.
It names counterparties and commercial terms. Operational conversations surface which suppliers are slow, which clients get exceptions and which contracts cause problems. That content is commercially sensitive independently of any personal data question.
It is durable. A process description stays accurate for months and an account of how an organization handles exceptions stays useful for longer. Unlike a point-in-time metric, it does not expire into harmlessness.
Information security frameworks treat this as a matter of classification and supplier control. ISO/IEC 27001 requires information security in supplier relationships to be agreed and managed, including for information the supplier processes on the organization's behalf. ISO 20252, the standard for market, opinion and social research, sets requirements for confidentiality and for how respondent information is handled, which is the closest formal analogue to a process interview programme.
Where the data goes
Six destinations, and which ones a buyer usually controls
| Destination | Legitimate | Controlled by a standard agreement |
|---|---|---|
| Your own findings and reports | Yes | Yes |
| The vendor's delivery team | Yes | Yes |
| Subprocessors and infrastructure | Yes, if disclosed | Usually |
| Model training or fine-tuning | Only if agreed | Rarely |
| The vendor's sector benchmarks | Only if agreed | Rarely |
| The vendor's sales conversations with your competitors | No | Rarely |
The bottom three rows are where the commercial risk sits, and all three are permitted by default in the absence of an explicit restriction.
The clauses that matter
Seven provisions, in order of how often they are missing.
1. Purpose limitation on derived insight
The strongest single clause. It restricts the vendor to using what they learn for your engagement, and prohibits the use of insight derived from your interviews, in any form including aggregated or anonymized, to develop or deliver services to named competitors or within your sector for a stated period.
Vendors with a benchmarking product will push back on this, and the pushback is informative. A negotiated version carves out genuinely aggregated statistics across many clients while keeping specific operational patterns out of scope.
2. No training without explicit consent
Interview content is not used to train, fine-tune or improve any model that serves other customers, unless separately agreed in writing. Silence here defaults to permitted in many standard terms.
3. Re-identification floor
A minimum group size for any reported figure, and a prohibition on reporting that would identify an individual in a small team. State the number. Without it, anonymized means whatever the vendor's template means.
4. Named sector restriction
Your competitors, named in a schedule, for a defined period. Broad non-compete language is usually unenforceable and a named list is both negotiable and meaningful.
5. Deletion and return on termination
What happens to raw transcripts, recordings and derived artifacts when the engagement ends, with a deadline and a confirmation. Raw interview content is the asset that matters here, and many agreements only cover deliverables.
6. Subprocessor disclosure and change notice
Which third parties touch the content, including model providers and infrastructure, with notice before that list changes.
7. Reference and case study consent
Whether your organization can be named, and whether a finding from your engagement can be described without naming you. Case studies frequently describe an operational problem specifically enough to identify the company to anyone in the sector.
What to ask in the vendor conversation
Contracts record the outcome. These questions reveal the practice, and the quality of the answer matters more than its content.
| Question | What a good answer looks like |
|---|---|
| What may you do with insight from our interviews beyond our engagement? | A specific restriction they already operate under, offered before being asked |
| Do you sell benchmarks built from client interview data? | A direct yes or no, with the mechanism if yes |
| Is interview content used to train models that serve other customers? | No, or a precise description of consent and isolation |
| What is your minimum group size for reporting a figure? | A number |
| Who are your current clients in our sector? | A clear position on whether they will say, and a stated conflict policy |
| What happens to raw transcripts when we stop working together? | A deadline and a confirmation process |
| Can an employee see what was recorded about their own input? | Yes, with the mechanism |
| Has a client ever asked you to restrict sector use, and what happened? | A real example |
The last question is the most useful one in the list. A vendor who has been asked before has a position. A vendor hearing it for the first time is forming one in the room.
The internal half of the problem
Confidentiality failures are not only a vendor question. Three internal conditions cause the same damage.
Findings circulated without the agreed scope. An interview conducted on the understanding that results go to the transformation team, and then presented to the function's management with identifiable detail, breaks the agreement the participant believed they had.
Performance use of operational input. A description of how long a task takes, given to improve the process, appearing in an individual's performance conversation. One instance of this ends participation for an entire function.
No answer to what happens to this. Participants asked to describe their work who receive no statement about who sees it will protect themselves, and the output is a defensible account rather than an accurate one.
Each of these is a design decision rather than an accident, which means each is preventable by stating the scope before the first conversation and holding it afterward.
Where Horizon fits
Horizon is an AI-powered continuous discovery platform. The conversations it runs are the most sensitive content an organization can produce about itself, which makes the handling of that content a product property rather than a policy appendix.
Workspaces provide separated areas with anonymization and spend limits, so the boundary between what a transformation team sees and what a function's management sees is enforced by the structure rather than by a reporting convention. The Insights Dashboard holds findings with traceability to the inputs behind them, which is also the mechanism that makes a re-identification floor auditable: it is possible to check what a given report would reveal about a small team before it is circulated.
Discovery Cycles cover the full population in scope rather than a sample, which has a confidentiality consequence worth naming. In a sampled exercise, the people interviewed are identifiable as the ones who were interviewed. Census coverage removes that signal.
At La Segunda, a specialty insurer, Horizon worked with a nine-person team and covered 100% of it, producing 10 findings against roughly 700 active cases and 48 hours of process mapping. Participants rated the experience 8.2 out of 10 and 90% said they would have the conversation again. A nine-person team is the hardest confidentiality case there is, since any role-level description identifies a person, and the willingness to participate again is the measure that reflects whether the handling was trusted. These figures describe one engagement and they are evidence rather than a guarantee.
The question worth asking any provider, in writing: what may you do with insight from our interviews after this engagement ends, and will you restrict it by named competitor and by sector?
Checklist before signing
- Does the agreement restrict the use of derived insight, including aggregated and anonymized forms?
- Is there a named list of competitors and a defined restriction period?
- Is model training explicitly prohibited without separate written consent?
- Is there a stated minimum group size for any reported figure?
- Are subprocessors disclosed, with notice before the list changes?
- Is there a deletion deadline for raw transcripts, with confirmation?
- Is reference and case study use addressed, including unnamed descriptions?
- Has the vendor stated their current clients in your sector, or their policy on disclosing that?
- Internally, who will see identifiable detail, and has that been stated to participants?
- Is there a written commitment that operational input will not be used in performance assessment?
Items 9 and 10 cost nothing and prevent the most common version of the failure, which is internal.
Common mistakes
Treating this as a security review. Encryption and access control do not address what a vendor is permitted to do with what they legitimately learned.
Relying on a standard NDA. Confidentiality clauses cover disclosure of information. They rarely cover the vendor developing sector expertise from your operation.
Accepting anonymized without a number. In a small team, a role-level description identifies a person. A stated minimum group size is the only version of anonymized that means anything.
Leaving model training unaddressed. Silence defaults to permitted in many standard terms, and interview transcripts are exactly the kind of content that gets used.
Covering deliverables and forgetting transcripts. The raw interview content is the sensitive asset, and many agreements only address the reports.
Promising confidentiality internally and then presenting identifiable detail. This ends participation for a whole function, and the next cycle inherits it.
Not asking who else in your sector they work with. The answer, or the refusal, tells you how the vendor thinks about the question.
FAQ
What is the main confidentiality risk in employee interview programmes?
Not a breach. The common failure is a vendor using insight legitimately gathered from your interviews to develop or sell services elsewhere in your sector, which a standard confidentiality clause does not prohibit because no information was disclosed.
Does an NDA cover the use of interview data?
Partly. An NDA generally addresses disclosure of confidential information to third parties. It does not usually restrict the vendor from building sector expertise, benchmarks or product improvements from what they learned, which is where the commercial exposure sits.
Is anonymization enough to protect participants?
Not on its own. In a team of nine, a description of one role's daily work identifies the individual. Anonymization is only meaningful with a stated minimum group size for any reported figure and a prohibition on reporting that would identify someone in a small team.
Should interview content be used to train AI models?
Only with explicit separate written consent. Many standard terms leave this unaddressed, which defaults to permitted. Interview transcripts are high-value training content, so the absence of a clause is not a neutral position.
What should happen to transcripts when an engagement ends?
Deletion or return within a stated deadline, with written confirmation. The agreement should cover raw transcripts and recordings specifically, since provisions that reference deliverables leave the most sensitive material out of scope.
How do you know whether a vendor's practice matches their contract?
Ask whether a client has previously required a sector restriction and what happened. A vendor with a position has been through it. A vendor forming the position during your call is telling you that no client has asked, which is itself the answer.
The agreement protects the people who spoke
Employees describing how their work actually functions are doing something that depends on trust, and the trust is not in the platform. It is in the person inside the company who asked them to participate.
Getting the commercial terms right is how that person keeps their word. The clause that matters most is also the one most often absent: what the vendor may do with what they learned, after the engagement is over.
See it. Fix it. Win it.