Teresa Torres's Continuous Discovery Framework Just Got an AI Upgrade in 2026

I remember sitting in that endless, slide-filled meeting where our team announced a new “continuous discovery” initiative. We had committed to weekly customer interviews per product trio, a promise lifted straight from Teresa Torres’s popular framework. The excitement quickly faded. Within a couple of months, our calendar was packed with sprints and releases, and the weekly interviews quietly petered out. We were not alone: Perspective AI, an AI-interview vendor, claims fewer than one in five teams sustain the weekly interview cadence Torres advocates. In the same vendor post, each interview is estimated to consume 4-6 hours of work for recruiting, scheduling, conducting, and synthesis. That burden was more than even three motivated people could protect against release pressure. The labor of continuous discovery kept us from actually doing continuous discovery.

By 2026, however, that pattern had started to change. AI-moderated interviews, where an AI agent conducts dozens of open-ended customer conversations in parallel, are becoming a practical upgrade for teams trying to make weekly discovery real. Perspective AI’s April 29, 2026 vendor post claims the human workload can fall from about six hours per interview to a 30-minute trio review, with roughly 100 minutes of weekly work per team member. More teams are also experimenting with always-on customer panels rather than quarterly surveys. Perspective AI’s June 8, 2026 vendor report and CleverX’s own marketing describe thematic synthesis from 20-40 interviews within days, panels of 80-300 customers, and entry-level costs under $500 per month. Those are commercial claims, not independently audited benchmarks.

In this article, I’ll explain how AI changes the continuous discovery operating model, naming Teresa Torres and the Opportunity Solution Tree directly, and show what the 2026 data says and where skepticism is warranted. I’ll approach the subject as a longtime discovery practitioner rather than a theorist. You’ll find structured comparisons of vendor claims and operational reality, plus concrete guidance for running a first AI-enabled discovery cycle. The Refonte Learning Product Owner Program covers backlog management and stakeholder communication, but its published curriculum does not name Torres’s framework or any AI interview tool. Treat this as a deep dive into continuous discovery habits in 2026: how the practice is supposed to work, what teams are actually doing, and why an AI interviewer may be the missing operational layer.

Why Most Teams Still Can't Keep Up With Weekly Discovery

Almost every team I’ve worked on wants to be more customer-centric, and many promise weekly discovery interviews. But within a few sprints that promise collapses. Why? The hidden workload is just too high. Running one traditional customer interview each week requires a string of tasks that add up to half a day (or more) of work per interview. Here’s a rough task list for one weekly interview with a 30-minute customer call:

• Identify and recruit participants: Define the segment, pull contacts from CRM or user database.

• Craft and send the invitation: Write copy, schedule the call, handle back-and-forth.

• Manage scheduling and no-shows: Confirm appointments, reschedule dropouts, and chase late replies. Perspective AI attributes a 20-35% no-show estimate for non-incentivized B2B studies to Nielsen Norman Group, although the figure is repeated in a vendor post rather than independently audited there.

• Conduct the interview: A 30-45 minute live call with the customer.

• Process the conversation: Transcribe or take notes, tag key quotes.

• Synthesize findings: Combine insights from multiple interviews into themes or updates to your Opportunity Solution Tree (if you use one).

Each of those steps seems small, but together they require roughly 4-6 hours of work for each 30-minute customer interview. Even a fast interview project can consume a full day: to complete eight interviews with a 35% no-show rate, you may need to schedule about 12. With three people in a product trio, the weekly burden can reach 15-18 person-hours, or roughly 5% of the team’s combined capacity.

When a real crisis hits, whether a launch deadline, a critical bug, or a major meeting, discovery is usually the first activity dropped. Perspective AI disclosed this problem in its April 29, 2026 vendor post: “Most teams that try the cadence fall off within 8-12 weeks because of recruiting and scheduling friction, not because they disagree with the cadence.” In practice, even teams that want weekly interviews often find the operating burden impossible without additional support.

What’s worse, teams often end up re-interviewing the same eager power users (burning out their best customers) or recruiting from incentivized panels (which biases the insights). The vendor narrative is that “talking to 200 customers” used to be prohibitively expensive, but now AI tools claim they can make it feasible. Meanwhile, traditional methods have teams giving up by the third month, leaving Torres’s continuous discovery concept sadly aspirational.

“We tried the weekly trio interview thing for a month, and then disaster recovery took all our time. The research ‘habits’ died because we literally had no hours left,” is a refrain I’ve heard from many Product Owner colleagues.

The underlying reason is simple: most teams can’t keep up because continuous discovery, as originally prescribed, demands an unsustainable fraction of their time. The friction lies in the human labor needed at every step: recruiting, scheduling, moderating, transcribing, and synthesizing. AI offers a chance to change the game by automating or collapsing these steps. Before we see how, let’s recap exactly what the original Continuous Discovery Habits framework asks teams to do.

What Continuous Discovery Habits Actually Asks Teams to Do

Teresa Torres’s Continuous Discovery Habits (2021) lays out a clear-but-ambitious recipe for discovery:

• Weekly customer interviews by the product trio. (Torres’s benchmark is at least one interview per week per product trio, with product manager, designer, and engineer all listening together.)

• An Opportunity Solution Tree (OST). A living visual that links a clear outcome (the measurable customer behavior you want to change) through opportunities (customer problems or needs) down to candidate solutions and assumption tests. The team updates the tree continuously based on interview findings.

• Outcome over output mindset. Focusing on customer outcomes (e.g., “increase feature adoption by 20%”) rather than just shipping features.

• Assumption testing, not debating. Move ideas forward by running experiments or tests on the riskiest assumptions rather than relying on opinions.

These four pillars, a clear outcome, an Opportunity Solution Tree, continuous interviewing, and assumption testing, recur throughout Torres’s book and talks. The idea is to make discovery a habit, like brushing your teeth, so it remains part of how the team works.

The Opportunity Solution Tree, Briefly

The Opportunity Solution Tree (OST) is a decision tree for discovery. It starts with the desired outcome at the root, then branches into customer opportunities, candidate solutions, and experiments or assumption tests. The structure helps teams connect what they are building to the customer need and business outcome it is meant to address:

• Outcome: A specific change in customer behavior.

• Opportunity: A customer need or pain point (discovered via research).

• Solution: A possible feature or change to address that opportunity.

• Experiment: A test (e.g., prototype, MVP, mini-experiment) to validate the solution and its assumptions.

When done well, every interview should feed into the tree by adding new opportunities and assumptions, and helping rank or drop potential solutions. But as noted, the reality is that putting this into action every week proved too much effort with manual methods.

For a broader strategic view, see Refonte Learning’s Product Owner Management in 2026. That article treats discovery conceptually; this article goes deeper by naming Torres, the Opportunity Solution Tree, and the 2026 AI-interview data.

Five Years Later: Where the Framework Actually Stands in 2026

Teresa Torres’s Continuous Discovery Habits is marking its fifth anniversary in 2026. Torres launched a year-long 2026 book club, with confirmed chapter-by-chapter sessions beginning in January and continuing through at least April, to emphasize that the habits are meant to be practiced rather than merely read about. Product Talk also reports that more than 135,000 people have bought the book. That figure is author- or publisher-reported and widely repeated, but it is not an independently audited sales total.

Despite the community interest, adoption remains patchy. Vendor-published summaries and industry anecdotes suggest weekly discovery is still rare. Perspective AI states that fewer than one in five product teams maintain the cadence. In my experience, teams that try it often stop after two or three months because time and recruiting friction overwhelm the habit.

The following adoption and persistence figures are vendor-published, so they should be treated as directional rather than independently audited:

• Fewer than 20% of teams meet the weekly-interview goal. Perspective AI’s April 2026 analysis says fewer than one in five teams sustain weekly interviews.

• The cadence often collapses after about 8-12 weeks. The same Perspective AI post says teams commonly fall off because of recruiting and scheduling friction.

• Interest is substantial. Product Talk reports more than 135,000 copies sold, and thousands of teams have experimented with the habits, but the sales figure is not independently audited and experimentation does not prove sustained adoption.

So by 2026 the continuous-discovery model is known and respected, but for many it remains aspirational. The good news is that, according to these vendor studies, technology is finally catching up to help make it operational. Let’s see how AI changes the math and whether it really solves the burnout problem.

What AI-Moderated Interviews Change About the Math

AI-moderated customer interview platforms aim to turn discovery interviews from a human-labor activity into a largely self-service process. The promise is that instead of spending hours manually recruiting and moderating each call, a single research outline can run dozens of interviews in parallel with an AI handling follow-ups, and automatically generate themes and summaries.

Here’s the high-level shift in unit economics:

• Traditional human-run interview: about 4-6 hours of human work per single 30-minute interview. That includes outreach, scheduling, conducting, transcription, and manual synthesis (imagine three people each spending about 6 hours every week on one interview).

• AI-moderated interview: A single outline or discussion guide is distributed by link or through a panel. The AI agent handles the conversation end to end and asks clarifying questions in real time, such as, “You said it’s slow. How slow: seconds or minutes?” After the sessions, the platform synthesizes themes and evidence for the team to review. The team can therefore review outcomes from many interviews without manually processing every session.

Perspective AI claims this compresses the workload sharply. Its vendor post says an AI-driven process can reduce 4-6 hours of interview work to a 30-minute weekly team review, or roughly 100 minutes of work per person in a three-person trio. That estimate is plausible only when recruiting, moderation, transcription, and first-pass synthesis are genuinely automated. Independent verification remains limited.

The weekly workload comparison is straightforward:

• Traditional Weekly Interview (per week, per trio of 3 people):

1.         About five hours of preparation and operations across three people, or roughly 15 person-hours.

2.         No automated synthesis; the team spends additional time collating notes and themes after each call.

3.         Few interviews per week (roughly 1-3 interviews total, as each demands so much manual work).

• AI-Moderated Weekly Interview (per week, per trio):

1.         Create one research plan and automate outreach through a panel, link, or in-product embed, with about one to two hours of setup by one person.

2.         Run 20-50 or more interviews asynchronously in parallel, with the AI following up on vague answers.

3.         Spend about 30 minutes reviewing a synthesized report, or about 100 minutes of distributed weekly work per team member according to Perspective AI’s vendor estimate.

4. Potentially dozens of interviews’ worth of data each week, not just 2-3.

Traditional discovery often behaves like a sprint: a small batch of interviews consumes weeks of effort. AI-driven discovery promises an always-on feedback loop that continuously supplies evidence.

Technically, the AI agent does the work of the recruiter, moderator, transcriber, and part of the synthesizer. The product trio still interprets and decides, but it no longer loses half the week to administration. As Perspective AI puts it: “AI conversational interviews change the unit economics of an interview from ‘human-hour’ to ‘compute-second.’ A research outline written once runs against thousands of participants in parallel.”

From a 4-6 Hour Interview to a 30-Minute Trio Review

To make the vendor estimate concrete, Perspective AI says the workflow can move from about six human-hours per interview to roughly 100 minutes per team member per week, or about five total hours for a three-person trio. In that comparison, the trio moves from roughly 18 person-hours to about five. Recruiting and synthesis account for most of the claimed reduction.

To illustrate the shift, consider this simplified before/after scenario for one week:

• Before, with a manual process: We schedule two interviews and invite three or four people to allow for no-shows. The product manager spends about one hour on invitations, two hours on scheduling and rescheduling, one hour conducting two 30-minute calls, one hour on transcription, and one hour organizing notes and themes. The engineer and designer each spend an additional hour in the calls. Total effort is about eight hours for the product manager and one hour for each teammate.

• After, with an AI-driven process: We send a link or panel invitation, which takes about 30 minutes to configure, and the tool runs 20 interviews over the week. The trio attends a 30-minute Friday review of the generated themes and implications. Total team time is about two hours, while the platform handles recruiting, moderation, transcription, and first-pass synthesis.

This is admittedly vendor-optimistic. The actual gain depends on tool quality and team work habits. But the key idea is: we shift most of the accumulated 6 hours per interview into a one-off upfront cost plus a short review, enabling much higher throughput. Vendors emphasize this a lot. We’ll unpack the claims and whether they hold up in the next section.

The Vendor Data Behind the Shift: Why It Needs a Skeptical Read

Before we jump to practice, it’s crucial to treat all the time-saving statistics as vendor claims. Perspective AI and others (including CleverX) have a clear incentive to paint AI interviews as a magic bullet. None of these numbers are from neutral academic research; they’re self-reported stats or marketing research. That’s not to say they’re outright false, but take them as directional with potential bias.

Here are some headline “numbers” many have quoted, and the context around them:

• 4-6 hours to 30 minutes: Perspective AI’s April 29, 2026 post says a traditional interview requires roughly 4-6 hours of aggregate work and that AI can reduce the recurring burden to a 30-minute trio review. The same post frames the weekly workload as roughly 100 minutes per team member, implying a reduction of more than 75%. This is internal vendor data, not an independent study. It also excludes any additional time the team spends validating themes or conducting human follow-ups.

• Fewer than 20% of teams maintain a weekly cadence: Perspective AI cites surveys saying fewer than one in five product trios reach weekly interviews and says teams commonly stop after 8-12 weeks. The vendor does not identify a neutral, audited dataset for the headline figure, so treat it as directional.

• No-show rates of 20-35%: Perspective AI attributes this range to Nielsen Norman Group for non-incentivized B2B calls. The linked vendor post does not independently audit the figure. AI moderation may reduce calendar friction when participants respond asynchronously, but direct outreach still requires a wider recruitment pool and replacement participants.

• Cycle time from three weeks to three days: The June 8, 2026 Perspective AI report claims a typical interview cycle compressed from about three weeks to roughly three days. It also says 80% of researchers use AI somewhere in their workflow and that AI-assisted analysis is a leading trend. Those results come from a vendor summary of roughly 300 teams, not a third-party audit. A three-day cycle is best treated as an achievable vendor benchmark under favorable recruiting and panel conditions, not a guarantee.

• Standing panels and cost: The June 8, 2026 Perspective AI report describes panels of 80-300 customers receiving biweekly AI interviews and claims an always-on panel can become a line item under $500 per month. CleverX, another vendor, advertises a large verified respondent marketplace and high parallel throughput. At face value, those claims suggest major economies of scale. In practice, software fees, incentives, specialized screening, and enterprise requirements can push the real cost higher.

• Scale of interviews: Perspective AI says some brands can run thousands of interviews in parallel, while CleverX claims 50-100 parallel interviews compared with 5-10 per week in a traditional model. These figures are designed to illustrate an order-of-magnitude increase in throughput. They remain vendor claims and should be tested against your own completion rate, participant quality, and review workload.

These are commercial pitches from Perspective AI and CleverX, not neutral research. They still indicate what vendors say is operationally possible in 2026, but results will vary by segment, recruitment method, tool quality, and team discipline.

Takeaway: Expect big claims from tool vendors about 75-90% time savings, huge upticks in interview volume, and cheap always-on panels. Those claims often originate from internal benchmarks and select studies. Use them to guide your expectations, but double-check with smaller pilots.

In practical terms, test the claims with a bounded pilot. If a vendor says you can move from three weeks to three days, measure whether your first project does. If it says 80% of researchers now use AI, remember that adoption somewhere in a workflow does not mean every team has an effective always-on system. The broader shift is real: discovery is moving from a manual sprint toward a continuous, partially automated flow, but the model still requires disciplined deployment.

Standing Customer Panels: Always-On Research Instead of Sprints

One recurring theme is the shift from episodic research sprints (big study, analyze, ship, repeat next quarter) to always-on, panel-based discovery. Vendors describe founders and high-performing teams maintaining standing customer panels as their primary feedback loop. That means a rotating cohort of real users that you can interview (via AI or humans) regularly.

According to Perspective AI’s June 8, 2026 vendor report:

Teams maintain panels of 80-300 customers (often B2B or key users).

• They send biweekly AI-moderated interviews into this panel (perhaps half the group every two weeks, or rotating sub-groups), so the panel constantly yields fresh insights.

• Crucially, the Perspective AI report describes an always-on panel as an expense under $500 per month. That would be far less than many quarterly research engagements, but the claim remains vendor-published and may exclude incentives, specialized recruitment, or enterprise requirements.

Here’s how an always-on panel can work in practice:

• Recruitment: The company enrolls a pool of opt-in users or customers into a panel. This could be an email list, a user community, or customers offering feedback (sometimes incentivized minimally). The panel need not be huge; 100-300 people is typical (as per vendor reports), but they ideally represent your core segments.

• Cadence: Every 2-4 weeks, you trigger an AI-driven interview to some or all of them. The interview link might be sent via email, in-product notification, or some CRM automation. Because it’s AI, respondents complete it on their own time.

• Insights: The AI platform compiles responses from dozens of these panelists. Each cycle, you get a thematic report of what this group is saying about current opportunities or prototypes.

• Iterate: You update your Opportunity Solution Tree or backlog based on fresh data continuously, rather than waiting a quarter to gather 10 responses.

What “Always-On” Actually Costs in Practice

• Cost: Reusing a standing panel through a platform such as CleverX can lower the cost per cycle. Perspective AI claims large panels and recurring interviews can start at under $500 per month. Verify whether a quoted tier includes qualified sampling, incentives, multilingual support, integrations, and the volume you actually need.

• Effort: You do still need someone (the Product Owner or a product designer) to write a good discussion guide and review the AI report each cycle. But this is a relatively modest weekly job, compared to scheduling multiple interviews per person.

• Data Volume: Because interviews run continuously, sample sizes per week can be higher. For example, if your panel is 100 and response rate is 50%, 50 interviews in a cycle is routine. That means you can compare subgroups or drop questions flexibly, something rare in a 10-person study.

• Engagement: One risk is panel fatigue. If you send a survey or AI interview too frequently, even if it’s short, people may tune out. Vendor claims imply low dropout costs, but in practice you’ll still lose some to attrition. So you might rotate subsets or refresh the panel periodically.

Overall, standing panels can make continuous discovery feasible without an interview sprint that consumes the next quarter. Vendors position them as an alternative to research engagements costing $10,000 or more. Whether the economics hold depends on your customer base, screening requirements, incentives, security needs, and desired level of human moderation.

“In our startup, we built an email list of 200 signups and ran AI interviews whenever we shipped a new beta feature. By month 2 we had 80 responses and spotted a UX issue we never would have caught in a quarterly survey,” says one product leader I talked to.

It aligns with what CleverX and others say: scale and speed at low cost. But I’d add a note of caution: these panels still require careful sampling and incentives to avoid vanity metrics or skewed responses. Just because interviews are cheap doesn’t mean any outcome is valid without planning. We’ll discuss that under “reading no-show and sample numbers” below.

What AI-Moderated Interviews Still Get Wrong

AI interview tools are powerful, but they are not magic. As a seasoned practitioner, I always watch out for what still needs a human touch. Here are some limitations and pitfalls of AI-moderated interviewing:

• Lacks human empathy and nuance. An AI agent can ask clarifying follow-ups, but it might miss emotional cues or subtle shifts in tone. For example, if a participant hesitates or sighs, a human moderator might pause or comfort them; AI will just move on. CleverX itself notes that “human moderators still outperform AI on highly ambiguous topics, emotionally sensitive subjects, and sessions where hypothesis pivots mid-interview.” In plain terms: if your discovery question is emotionally charged (e.g., about job loss, health issues) or you want to deeply explore a complex personal story, a human is better. Use AI for more transactional or straightforward topics.

• Dependence on question design. AI won’t improve on a poorly written guide. The tool executes your script intelligently, but if you ask bad questions, you’ll get bad answers faster. For instance, an AI agent can’t magically generate deep insights if your prompts are vague or leading. You still need to invest effort into a good discussion guide; albeit that effort scales across many interviews.

• Superficial consistency versus creative thinking. AI agents ensure every participant hears the same (structured) line of questioning, which is great for consistency. But humans can improvise; say, someone hints at an unrelated pain point, a skilled interviewer might follow that new thread. An AI will usually stick to its guided topics unless specifically instructed. So there’s less serendipity. Some AI tools allow conditional probing (e.g., if a keyword appears, ask something specific). Still, an AI may miss an unexpected but important tangent.

• Quality of synthetic analysis. After the interviews, AI generates themes, quotes, and summaries. These can be accurate or obviously off. You should always double-check a sample of transcripts manually. CleverX recommends checking 15-20% of transcripts to ensure theme accuracy. If the AI’s analysis is wrong, your decisions will be too.

• No substitute for human sense-making. AI cannot make the product decision. Perspective AI summarizes the division of labor this way: “The trio still owns judgment; the AI handles the labor.” Product owners, designers, and engineers still have to interpret the evidence, challenge themes, and decide what to do. Blind trust in automated synthesis can damage prioritization just as easily as poor manual research.

• Ethical and privacy issues. AI-driven interviews still require clear participant consent, whether verbal or captured through an explicit consent step. Recording or analyzing people without opt-in can create legal risk, especially under the GDPR and other applicable laws; model bias and cross-cultural interpretation also require review. CleverX advertises multilingual support for global studies, but teams should verify that feature against their actual languages and user populations.

• Sample bias if misused. Tools often come with integrated panels (like that eight-million-person panel claim). If you rely solely on these panels, remember who’s on them: are they professionals who signed up for general research? That may or may not match your target customers. For instance, a panel of eight million professionals might under-represent your everyday small-business users. You still need to specify screening criteria carefully. A common mistake is assuming “AI scales automatically” means we don’t have to think about who we’re talking to. We absolutely do.

AI moderators can accelerate research, but the craft remains human. Think of AI as a highly productive assistant: it handles routine probing and note-taking, while the team remains responsible for framing questions, empathizing, spotting anomalies, and making decisions. The best teams automate the tedious work, spot-check the output, and preserve room for human-led conversations. As one UX researcher put it, “AI is like a caffeinated junior researcher: it’s fast and enthusiastic, but I still need to guide it and correct it.”

What This Means for the Product Owner's Actual Job

All these changes have real implications for how you, the Product Owner, spend your time and what skills matter. The advent of AI-driven discovery doesn’t make Product Owners obsolete; it shifts our focus.

• More time on synthesis and decision-making. Traditionally, Product Owners spent significant effort setting up interviews and digesting notes. With AI handling the administrative work, you can focus on interpreting insights. The new bottleneck is analysis depth, not analysis throughput. Expect to hold shorter weekly triad meetings (about 30 minutes) to go over AI reports and update the Opportunity Solution Tree or roadmap accordingly. Your role becomes curating and validating the AI’s output: reading transcripts snippets, selecting quotes, and tying them to customer outcomes.

• Stronger collaboration with design/engineering. Continuous discovery has always been about the cross-functional trio. That doesn’t change; it actually intensifies. Each week, all three attend the AI interview review. You as Product Owner still tie it back to business outcomes, but the designer hears raw user language, and the engineer can ask follow-ups. The difference is they’re doing it together asynchronously (if AI interviews are chat-based) or in one recap meeting. So Product Owners should strengthen their facilitation skills: running those quick review meetings effectively becomes part of the routine. (See also Refonte Learning’s article “Product Owner in 2026: Future Trends” which notes the trend towards even more customer-centric and UX-integrated roles.)

• Greater emphasis on customer empathy and insight literacy. The bar is raised for product owners to understand and translate what customers say. The metrics that matter become qualitative: are we hitting customers’ “language of the problem”? Are solutions tracing back to opportunities? If AI produces a theme, the Product Owner needs to judge it (“does this match what our churning users really cared about?”). In other words, soft skills like empathy and systems thinking get an upgrade. Torres talks about mindset shifts: being outcome-oriented, customer-centric, and experimental; these are exactly the mindsets Product Owners will need.

• Staying current with tools. A modern Product Owner should understand AI research tools or at least be ready to learn them. You may need to configure an AI interview, filter responses, review transcripts, or shape the moderator’s probing behavior. This is less about coding than about mastering a new software category, much as Product Owners learned Jira and Confluence. The Refonte Learning Product Owner Program does not name a specific AI interview tool, but it does cover Jira, Confluence, and roadmap planning, which are foundational for integrating a new discovery practice.

• Balancing speed with rigor. Faster research creates more temptation to move too quickly. Good Product Owners set safeguards, including periodic live interviews, transcript spot-checks, and clear evidence thresholds. The pace of 2026 enables more experiments, but it does not remove the need for sound experimental design.

For a broader view, Refonte Learning’s Product Owner in 2026: Future Trends covers general shifts such as AI-driven analytics and stronger UX integration. This article focuses on a concrete change: Product Owners can now orchestrate recurring customer conversations with AI support. The role increasingly involves ensuring that customer evidence reaches the backlog, roadmap, and decision process. AI in research supplies the qualitative input that complements AI-assisted prioritization models and dashboards.

In short, expect your day-to-day to include something like: starting Monday by reviewing an AI-generated discovery report, adjusting the sprint backlog or roadmap based on a new opportunity identified, and ensuring the team understands why the customer needs a certain feature. Whereas before you might have only done that monthly or quarterly, now it could be weekly.

Where This Fits (and Doesn't) With Your Certification

Certifications such as CSPO, PSPO, and SAFe Product Owner provide valuable foundations, but their standard curricula do not teach Teresa Torres’s framework, the Opportunity Solution Tree, or AI-moderated interviews in depth. They focus on backlog management, Scrum events, and stakeholder communication, all of which matter, but none substitutes for a weekly customer-discovery operating model.

For example, the Refonte Learning Product Owner Program (the 3-month course we referenced) emphasizes:

• Introduction to Product Ownership: Basics of the role in Agile teams.

• Backlog Management and Prioritization: How to manage requirements and priorities.

• Stakeholder Communication and Collaboration: Working with teams and executives.

• Tools covered include Jira, Confluence, and general roadmapping; methodologies like Scrum/Kanban.

• Mentored by an experienced Product Owner (Kevin Harris, 10+ years).

These modules do not explicitly teach Continuous Discovery Habits or the Opportunity Solution Tree by name. As of this writing, the live program page also does not name Teresa Torres or any AI interview software. The honest framing is that the program builds directly relevant foundations, while continuous discovery and AI-moderated interviewing remain specialized practices that learners must add through self-study and practical experimentation.

This distinction matters. Sending a Product Owner to a Scrum course will not automatically establish weekly discovery. Teams may need self-study, coaching, or an on-the-job pilot. The foundation remains useful: backlog prioritization helps track experiments from the Opportunity Solution Tree, while Jira and Confluence help document hypotheses, evidence, and decisions. The certification or training supplies Agile discipline; continuous discovery is a workflow layered on top.

Standard Product Owner certifications cover how to manage value, backlogs, and collaboration, but not the full discovery framework or AI tooling. Continuous discovery is an advanced practice area. Certification skills still help teams execute it, especially when converting findings into clear user stories, experiments, and stakeholder decisions.

CSPO, PSPO, and SAFe POPM Don’t Teach This

SAFe POPM remains high level, while Scrum Alliance’s CSPO and Scrum.org’s PSPO focus primarily on Product Owner responsibilities within Scrum. They may mention research or stakeholder understanding conceptually, but they do not prescribe weekly interviews or an AI-moderated workflow. A training session that teaches continuous discovery by name is therefore an additional specialization rather than standard exam preparation.

The practical upshot: you won’t “learn AI discovery” in your exam prep, so consider a few options:

• Seek out workshops or articles on the Opportunity Solution Tree (these are increasingly common, given Torres’s book club in 2026).

• Experiment with a team using an AI interview tool on a small question (even a free trial). Learn by doing.

• Use your program’s stakeholder and communication skills to bring teammates on board. For example, pitch one AI interview cycle around a specific feature hypothesis, then share the evidence and the resulting decision.

Getting Your Team's First AI-Moderated Interview Cycle Right

If you’re convinced to at least pilot an AI interview, how do you start? The process still needs discipline. Here’s a step-by-step approach (each bullet is roughly one week’s focus, though you could sometimes combine or adjust timing):

1. Clarify one outcome and question. Don’t try to map an entire OST yet. Pick a single product outcome you care about (for example, “increase trial-to-paid conversion by 15% in our CRM product”). Identify one target customer persona or segment for this question (for example, “IT managers at SMBs who use our trial”). Write down the specific research question tied to that outcome (for example, “Why do some trial users not convert to paid?”).

2. Craft a discussion guide. Write 5-8 open-ended questions that cover the initial context, the core decision, and a wrap-up while leaving room for probing. Decide whether to allow text, audio, or video responses; text is often sufficient for B2B, and the session should take about 15-30 minutes. CleverX’s playbook suggests organizing the guide around 3-5 themes with probing directions.

3. Choose your panel or outreach. If you have a list of actual customers/contacts matching the persona, prepare that. If not, set up a paid panel via a tool (like CleverX’s eight-million-person pool) or even a simple screener in Google Forms to recruit fresh people. For the first run, a hybrid is fine: for example, email out 50 invites to qualified past leads or signups, plus 20 from a panel service.

4. Pilot 3-5 sessions manually. Even though it’s AI, pilot a few interviews. Some platforms allow testing with colleagues or dummy accounts. Check that the AI is asking as you expect. Adjust any question that yields a one-word answer or confusion. For example, if the AI misinterprets a question, rewrite it. (CleverX suggests doing a mini-pilot as step 5 in their guide; it’s vital.)

5. Launch the full study. Deploy the interview link or panel. If using a panel, set your screener (for example, job title, company size). If inviting your own contacts, email the link and perhaps follow-up once. You might offer a small incentive (gift card, donation) to improve response, especially if people don’t know you.

6. Monitor response. Keep an eye on completion rates. If it stalls at 20%, check if people find it too long or boring. You might send a reminder after 2 days. The target is at least 20-30 completed interviews for good thematic saturation (add more if splitting subgroups).

7. Review AI output weekly. Decide on a regular time (for example, Friday afternoons) where the product trio meets for 30 minutes. Open the AI’s summary report: look at themes, quotes, and answer breakdowns. Have each team member note anything surprising. Use this meeting to update your Opportunity Solution Tree: add any new identified opportunity (customer need) and drop ones that get contradicted.

8. Decide next steps. If clear patterns emerge (for example, many cite "lack of a key feature" as the reason to drop), decide on a quick experiment (for example, a new demo of that feature, a landing page test). Or you might refine and run a follow-up AI interview in 1-2 weeks focusing on one insight. The point is to keep iterating in short cycles, guided by customer voice.

Key practices:

• Start small and iterate. Don’t launch a massive multifaceted discovery all at once. Fix one outcome and one interview topic. The Perspective 30-60-90 plan suggests exactly that: make one weekly interview “real” first by embedding it in product (for example, post-signup), then build the tree.

• Use the AI tool as intended. Some AIs work via chat interface, some via pre-recorded question-and-answer sessions. Train your team on how to read the transcripts or use the dashboard. If the tool writes a raw report, consider copying key quotes into your OST or Confluence page so the team doesn’t lose them.

• Rotate team roles. Even though AI does the talking, the trio should share responsibilities: one person can monitor volume and completion, another can inspect verbatim quotes, and another can compare segments or opportunity themes. Rotate those roles periodically so everyone remains engaged.

• Keep documentation. If you’re not already, use a shared doc or wiki to log insights. For example, after each AI cycle, write a short memo: “Interview #1 (AI) findings: [top themes]. We updated OST with new opportunities: X, Y, Z.” Over time this becomes a log of how discovery evolved.

The following decision table summarizes the first cycle:

By following a clear process, you avoid pitfalls like “launching randomly” or ignoring the AI data. Each cycle should feel like a micro-sprint in your discovery process.

Common Mistakes When Teams Add AI to Discovery

Even with AI, teams can stumble. Here are some common missteps I’ve seen (and how to avoid them):

• Treating interview volume as a substitute for insight. Just because you can run 100 interviews doesn’t mean you automatically have wisdom. Avoid the trap of thinking “more data = more truth.” What matters is what you ask and how you interpret. Always clarify what decision each question is meant to inform. (If your AI output is just a flood of themes, you could get analysis paralysis.)

• Sending a survey instead of a conversation. Some teams use AI wrongly by feeding it checkbox-like questions or yes/no queries. Remember, AI interviews should be open-ended dialogues. If you only ask structured questions, you lose the richness. Use the AI’s strength: allow it to probe with follow-ups (for example, “I see you said X, please tell me why.”).

• Ignoring the need for incentives or reminders. AI can remove scheduling work, but people still need a reason to participate. If the response rate is very low, consider a modest incentive and a clear reminder sequence. Some vendors manage incentives within the platform; others require a separate budget and workflow.

• Not iterating on the script. Early interviews (or pilots) should refine your guide. A classic error: deploying the same script week after week even if it’s not yielding depth. For example, if participants often answer “fine” to “how are you?”, maybe omit that in future. Treat your discussion guide as living: add or remove questions based on what was or wasn’t insightful in the previous cycle.

• Overlooking technical glitches. New tools have kinks. Maybe some participants’ recordings didn’t save, or an audio-to-text glitch omitted entire answers. Always spot-check a subset of transcripts. If you find issues (for example, poor transcription of technical jargon), adjust settings (maybe switch to audio or add specific terms to the vocabulary) or clarify instructions ("please speak clearly into your mic").

• Under-communicating results. Gathering insights is useless if they don’t inform decisions. A mistake is doing all this work and then not adjusting the backlog or roadmap accordingly. Make sure that each review meeting ends with clear actions: add a user story to the backlog? Scrap a feature? Note in the OST that a hypothesis was confirmed or not? This ties continuous discovery back to delivery.

• Failing to calibrate AI against human-led responses. Periodically compare an AI-moderated cycle with one or two human interviews from the same segment. If findings diverge, investigate whether the AI missed nuance, the human moderator introduced bias, or the samples differed. This calibration helps the team build justified trust rather than blind confidence.

Avoiding these mistakes makes AI-enabled discovery more likely to produce usable insight. The tool can automate labor, but it cannot define the decision, judge the evidence, or align the team. Treat it as one powerful method in a broader product-research toolkit.

Treating Interview Volume as a Substitute for Insight

A team I advised once ran an AI interview with 100 customers in a week and surfaced ten new themes, yet ended the cycle more confused than before. They had no clear analysis plan and had not updated the Opportunity Solution Tree, so the themes floated without a decision context. More interviews help only when a framework, outcome, or hypothesis organizes the evidence. Volume without structure produces information, not direction.

Reading the No-Show and Sample-Size Numbers Honestly

Customer research always involves dealing with reality: not everyone shows up, and we need enough interviews to feel confident. The vendor statistics help, but we should translate them into practical planning.

• No-show rate of about 20-35%: Perspective AI attributes this B2B estimate to Nielsen Norman Group. Treat 30% as a planning assumption rather than a universal benchmark. To secure 30 completed interviews, you may need 40-50 qualified invitations, unless a platform over-samples automatically.

• Completion and Engagement: AI interviews might see higher completion than open calendar invites, because participants can pick their time and aren’t in a video call. But they might also answer hastily. Check the response depth (for example, average session length). Some tools report an 80% completion rate as a good target.

• Sample Diversity: Do not let the ease of AI narrow the sample. For a B2B SaaS segment, include power users, churned customers, different company sizes, and distinct use cases. If one subgroup is underrepresented, such as only five SMB customers among 30 completions, target that segment specifically in the next round. In practice, you may run separate mini-studies for new users, long-term users, and churned accounts.

• Statistical vs Qualitative Saturation: In qualitative research, we talk about “saturation” (no new themes after a point). That often hits around 12-20 interviews per homogeneous segment. If you run 30-50, you should cover most obvious themes per segment. More volume then is mostly for confirming or for deeper subgroup splits (e.g., compare sentiment between segments). Don’t feel pressure to run endless interviews just for the sake of numbers; stop when new interviews only yield minor stuff.

• Quality of Data: Track metrics like completion rate, average length, and dropout rate within the AI tool (most provide these). If they fall below certain thresholds (completion <70%, or average length under 2 min), reconsider your approach (maybe questions are confusing).

• No incentives vs small rewards: Many B2B users will do interviews if they feel involved in shaping the product. But even $5-$10 gift cards per interview can lift response. Vendors often include or recommend incentives. It can be surprising: spend $200 on gift cards to reliably get 30 completes ($6 each) versus $0 incentive and chasing 80 emails for the same.

In summary, plan for imperfect turnout. Vendors like to advertise panel ease, but real customers may ignore unplanned surveys. A realistic strategy: double your target invites, monitor progress mid-week, and pump in more invites if needed (or convert to a paid panel if you hit your internal limit).

Keep reports honest. If a vendor says 100 interviews ran on schedule, ask how many people were invited, how screening worked, and how many responses were excluded. Record invitations, starts, completions, dropouts, session length, and segment mix. Over time, those figures become your own operating benchmarks.

Product Owner Salaries in 2026

In the United States, Product Owner salary estimates vary by source but generally sit in the six-figure range. Glassdoor and ZipRecruiter show a roughly $30,000-$40,000 difference in their headline figures:

• Glassdoor: Reports median total pay of about $142,000 per year for a Product Owner, with a typical range of roughly $110,000-$187,000. Some Glassdoor views show higher senior-level estimates, so readers should compare the exact role and experience level rather than treating every figure as interchangeable.

• ZipRecruiter (August 2026): Reports a national average of $112,891 per year. Its middle 50% runs from about $93,500 at the 25th percentile to $129,500 at the 75th percentile, with the 90th percentile around $150,000.

In practical terms, many sources place common Product Owner compensation between about $113,000 and $142,000, while senior or specialized roles may pay more. The source gap probably reflects different samples, definitions of total pay, employer mixes, and experience levels. Glassdoor’s headline figure is roughly $30,000 higher than ZipRecruiter’s average.

Product ownership remains a well-compensated career, but salary should not be read as a single market truth. Candidates should compare location, industry, seniority, base pay, bonus, and equity. Skills such as structured discovery, evidence-based prioritization, and stakeholder leadership can strengthen a candidate’s case for higher-responsibility roles.

Glassdoor’s higher figure may reflect more total-pay reporting or a different employer and experience mix, while ZipRecruiter may capture a broader range of postings. The divergence is more informative than either number alone: compensation datasets use different methodologies, so Product Owners should triangulate rather than choose the most flattering estimate.

Building This Skill Set: The Refonte Learning Product Owner Program

For professionals who need a stronger foundation in structured product work, the Refonte Learning Product Owner Program is one option. It does not claim to teach continuous discovery, Teresa Torres’s framework, the Opportunity Solution Tree, or a named AI interview tool. It does cover the core disciplines needed to operate those practices responsibly.

Here are the verified facts about the program, including its format, curriculum, tools, credentials, and fees:

• Format: Three months at about 8-10 hours per week, delivered through online training and an internship format.

• Curriculum: Three core modules:

1. Introduction to Product Ownership (understand the Product Owner role in agile teams).

2. Backlog Management and Prioritization (create, refine, and rank your backlog for max value).

3. Stakeholder Communication and Collaboration: Techniques for working effectively with developers, executives, and customers. These skills help a Product Owner turn research evidence into aligned decisions and communicate why priorities changed.

• Tools and methods: Jira, Confluence, general roadmapping tools, and Agile practices including Scrum and Kanban. These tools can support an Opportunity Solution Tree workflow by connecting opportunities, experiments, and backlog items.

• Mentor: Professor Kevin Harris, with more than 10 years of experience and a Fortune 500 background.

• Certification: A Training Certificate and a Certificate of Internship upon completion. The program itself does not issue CSPO, PSPO, or SAFe POPM certification.

• Fees: $300 as a one-time payment, or two installments of $204 and $98.

The program’s Backlog Management and Stakeholder Communication modules provide the organizational foundation for discovery. Backlog skills help translate research into experiments and delivery choices, while communication skills help teams share evidence and challenge assumptions. Continuous discovery and AI interviews are complementary practices layered on top of that core Product Owner toolkit.

The Refonte Learning Product Owner Program can help learners build the Agile, backlog, and stakeholder foundations needed for product ownership. Learners should supplement the curriculum with Teresa Torres’s book, Opportunity Solution Tree practice, and hands-on customer interviewing. The strongest path combines formal foundations with repeated discovery work on real product questions.