Customer discovery for AI products in HR: what the framework looks like in practice

Why HR is a specific challenge for AI product discovery

The HR technology market is crowded with AI claims. Every ATS vendor has added "AI-powered" to its feature list. Every performance management tool now offers some form of automated insight. The noise level is high, which means buyers are both more aware of what AI might do and more sceptical about whether any specific product will actually deliver it.

This creates a paradox for founders doing discovery in this space. The people you interview have often already seen an AI pitch. They have participated in a pilot that went nowhere. They have strong opinions about what AI cannot do, shaped by tools that did not work as advertised. You are walking into a conversation where the prior experience is often negative, and where "AI" as a word carries more baggage than signal.

The second challenge is structural. The general framework for B2B AI discovery assumes that the person who feels the pain is, at minimum, adjacent to the person who makes the purchase decision. In HR, this is rarely true. The CHRO who signs the contract cares about compliance risk, cost-per-hire, retention metrics and board-level reporting. The Talent Acquisition Manager who uses the tool every day cares about whether it makes the next 200 CVs faster to review. The hiring manager who is actually frustrated with the process often has no seat at the table when the software decision is made. These are three different conversations, and confusing them produces the most common mistake in HR AI discovery: building a product the buyer likes but the user never adopts.

The three discovery conversations you need to have

Mapping the three stakeholder layers in HR before you design anything is not a nice-to-have. It is the minimum viable discovery. Each layer surfaces different information, and each requires a different conversation structure.

Where AI is actually being adopted in HR right now

Not all HR workflows are equally ready for AI intervention, and the ones with the most hype are often not the ones with the strongest adoption signal. After running discovery in this space across different organisation sizes and sectors, the pattern is fairly consistent.

High-signal workflows

Talent acquisition screening. The volume problem in recruitment is real: a mid-size company posting a role in a competitive market can receive 300 to 600 applications for a single position. The current process involves a recruiter spending the first two to three hours doing a pass that eliminates candidates who obviously do not meet the baseline criteria. This is high-volume, low-judgment work by design. It is the clearest case in HR for AI augmentation, and it is where the most mature products exist. The discovery signal here is strong because the pain is quantifiable: hours per role, applications per recruiter, time-to-first-interview.

Onboarding documentation. New hire onboarding generates a predictable volume of repetitive communication: welcome packs, role-specific compliance documents, policy summaries, first-week schedules. HR teams in companies with high hiring volume rebuild this from scratch for every cohort, often because the previous version is out of date by the time someone looks for it. AI document generation and updating is a genuine workflow improvement here, with low adoption risk because the output is reviewed before it reaches the employee.

Performance review synthesis. Mid-year and annual review cycles produce large volumes of manager feedback that HR needs to process, flag and act on. Summarising themes across a team, identifying outliers, or generating structured feedback from unstructured manager notes are all tasks where AI produces usable output faster than a manual process. The adoption dynamic here is good because the HR team, not the employee, is the first reviewer of the output.

L&D content personalisation. Learning and development teams in larger organisations spend significant time curating content for different roles and seniority levels. AI-assisted content tagging, curation and basic module generation reduces this time materially. The signal in discovery is that L&D managers almost universally describe their content library as "a mess that nobody navigates well."

Low-signal workflows

Compensation and benefits planning sits at the intersection of legal constraint, market sensitivity and internal politics. The decisions are high-stakes, the data is confidential, and the people who make them are not looking for an AI recommendation they cannot explain to an employee. Culture measurement and "AI-powered engagement scores" consistently produce interest in demos and resistance in adoption because the output is either so aggregated it is useless or so specific it creates liability. Succession planning involves judgment calls that senior leaders are not ready to delegate to an algorithm, and for good reason. These are not areas where AI adoption fails because the technology is not good enough. They are areas where the workflow contains dynamics that AI cannot safely navigate at this stage.

The discovery conversation with an HR team: what to ask

The framing problem in HR discovery is that any question that sounds like "how could AI help you" will produce one of two unhelpful answers: either a list of features copied from the last product demo the person attended, or a reflexive concern about job displacement that closes the conversation before it has started.

The conversation structure that works is built around reconstructing what actually happened the last time the team ran the process you are targeting. Not what they think about it in the abstract. What actually occurred, step by step, the last time.

Questions that produce workflow signal in HR

"Walk me through the last time you filled a role from the moment the JD was approved to the moment the offer was signed. Start at the beginning and tell me exactly what you did." This produces a workflow map, which is the most valuable output of any discovery session. The goal is to identify where the process slows down, where information is lost, where decisions are made without the data the person would actually want, and where the workarounds are.

"What is the thing in that process that you have been meaning to fix for two years but have not touched?" This surfaces the problems that are real enough to be named but not yet painful enough to have forced a solution. They are usually the highest-signal areas for a product intervention, because the person has already validated the problem by living with it for years.

"When the hiring manager came back with feedback on the last shortlist you sent, what did they say?" The gap between what HR thinks the hiring manager wants and what the hiring manager actually says is often where the most painful friction lives. Surfacing this in discovery points directly to the information architecture problem your product needs to solve.

"If this problem disappeared tomorrow, what would you do with the time?" This is the value anchor question from the general framework, applied to HR specifically. The answer reveals whether the problem is genuinely painful (the time recovered would go to higher-value work the person actually wants to do) or merely inconvenient (the time recovered is not clearly allocated to anything better).

Questions that close down HR conversations

Avoid anything that leads with AI explicitly. "How do you see AI fitting into your recruitment process?" is a question about technology, not about work. It produces technology opinions, not workflow signal. Avoid any question about efficiency that could be heard as "how would you do this with fewer people?" In HR, that framing activates a defensive posture immediately. And avoid feature prioritisation exercises with a pre-defined list. The features you have pre-defined are already the answer to a different question than the one you need to be asking.

Worked example: AI-assisted screening for talent acquisition

This is the use case with the most mature signal in HR AI discovery, so it is worth walking through concretely to show what the discovery process actually surfaces, versus what founders typically build in response to it.

The problem as described by the buyer

A VP of HR at a mid-size technology company explains that their cost-per-hire has increased 40% over two years. They want AI to "streamline the recruitment funnel." They have a budget. They want a pilot by Q3. This is a real problem described at the wrong level of resolution to build anything useful.

The problem as lived by the daily user

The Talent Acquisition Manager running the same hiring process describes something more specific. They receive an average of 380 applications per role. The first pass takes about two and a half hours: scanning for obvious disqualifiers like location, salary expectation, years of experience, language requirements. This produces a shortlist of 40 to 60 candidates who get a second read, which takes another hour and a half. The problem is not the time itself. The problem is that this work happens before 9am on a Monday because it needs to be done before the weekly sync with the hiring manager, and it is the first thing on their calendar, which means it crowds out the relationship work they actually care about.

That is a completely different problem than "streamlining the funnel." It is a scheduling and attention problem disguised as a volume problem. And it points to a different product intervention: not a ranking algorithm, but a pre-processed shortlist delivered on Friday afternoon so the TA Manager can review it before Monday rather than producing it on Monday morning.

What a well-scoped MVP addresses

The minimum product worth building and testing is not a full screening platform. It is a tool that takes the JD and the application batch and returns, by Friday afternoon, a structured shortlist with a one-paragraph rationale for each candidate against the stated criteria. The TA Manager reviews it, adjusts the weightings where the rationale is wrong, and sends the result to the hiring manager. The metric: does the TA Manager use this output without re-doing the work from scratch? If yes, you have a product. If they consistently override the shortlist, you need to understand why before building anything else.

The hiring manager problem, which the MVP above ignores

The friction that produces the 40% cost-per-hire increase is not actually in the screening step. It is in the feedback loop between the TA Manager and the hiring manager. The hiring manager reviews the shortlist asynchronously, comes back three days later with vague feedback ("not quite what we were looking for"), and the TA Manager has to reconstruct what criteria changed without a structured record of what the hiring manager was actually thinking. The next iteration of a screening AI product, once the first MVP is validated, is a tool that captures the hiring manager's feedback in a structured form at the point of review, feeds it back into the screening criteria for the next pass, and creates an audit trail of how the definition of the ideal candidate evolved over the course of the search. That is the product that solves the actual business problem. But you only find it by doing the discovery conversation with both the TA Manager and the hiring manager, which most teams skip.

The adoption barriers specific to HR in Europe

Building an AI HR product in Europe adds a regulatory dimension that is not optional and not lightweight. Most AI projects fail for organisational reasons, but in HR, there is also a specific legal architecture that shapes what you can build and how you can sell it.

The EU AI Act classifies AI systems used in employment, worker management and access to self-employment as high-risk. This includes automated candidate screening, worker performance monitoring and task allocation systems. High-risk classification means mandatory conformity assessment before deployment, human oversight requirements built into the product, technical documentation, and registration in the EU database for high-risk AI systems. For a founder building in this space, this is not a compliance checkbox to handle before launch. It is an architectural constraint that needs to be factored into the product from the first design decision. How you store candidate data, how you log screening decisions, what override mechanisms the human reviewer has, and how you document the model's criteria all need to be designed with high-risk compliance in mind.

GDPR adds a second layer. Automated decision-making that produces legal or similarly significant effects on individuals requires either explicit consent or a specific legal basis, and must include the right to human review. For a screening tool, "legal or similarly significant effects" almost certainly covers a rejection from a job application. This means your product cannot be a black-box ranker. It needs to support human review in a way that is documentable and explainable to the candidate if requested.

The practical implication of both frameworks is that the fastest path to market in European HR AI is not automated decision-making but decision support. A tool that presents structured information to a human who makes the final call is in a fundamentally different compliance category than one that makes or recommends the decision autonomously. This is also, as it happens, a better product for the daily user, who does not want an algorithm taking over the judgment calls they were hired to make.

Pricing signals in HR discovery

The value anchor in HR is more concrete than in most verticals because the cost metrics are well established. Cost-per-hire, time-to-fill, recruiter capacity (roles per recruiter per quarter), and turnover cost are all numbers that HR teams track and can usually quote from memory. This makes the pricing conversation easier to ground than in verticals where the value is diffuse.

The question that surfaces willingness-to-pay most reliably in HR is not "what would you pay for this tool?" It is "if this tool gave your team back ten hours per role, what would those ten hours go toward?" If the answer is high-value work the person genuinely wants to do (relationship building with passive candidates, improving offer acceptance rates, reducing time-to-productivity for new hires), the willingness-to-pay is real and you can price toward the value of that recovered capacity. If the answer is vague ("we could probably handle more roles"), the recovered capacity has no clear destination, which usually means the budget will not materialise when the procurement decision arrives.

A concrete anchor: a recruiter in a mid-size European company costs between 50,000 and 80,000 euros per year in total employment cost. At capacity of 4 to 6 roles per quarter, recovering two hours per role represents roughly 5% of their annual capacity. A tool that demonstrably does this is worth, to the organisation, somewhere between 2,500 and 4,000 euros per recruiter per year as a floor, before accounting for quality improvements or speed-to-hire effects. If your price point is below that floor, you have a pricing problem. If the prospect cannot see that value clearly, you have a discovery problem.