This week in AI and product
AI has made it cheaper and faster to build. It hasn’t made it any easier to build the right thing. This week’s digest brings together ten reads on what that shift means for teams bringing AI products to market: earning user trust, measuring quality, and understanding why faster output doesn’t automatically translate into better outcomes.
These articles explore the challenges product managers face when building and scaling AI products: earning user trust, measuring product quality, and turning faster development into meaningful customer and business outcomes.
The top 10
Startups Magazine · Maksim Borisik, Senior PM at Adapty · 25 Sep 2025
Adapty had redesigned its no-code paywall builder, but customers were still hesitant to use it. Their concern was understandable: one wrong edit could damage a paywall responsible for generating revenue. The solution was an AI-generated first draft covering layout, copy, design and pricing. Instead of starting from a blank canvas, users could review and refine a ready-made version. The builder now serves more than 5,000 apps and 30,000 live paywalls, which account for roughly a quarter of clients’ aggregate revenue. The AI feature alone has generated more than 1,000 live paywalls.
Why it matters: This is a launch story backed by concrete numbers. The wider lesson is to treat user hesitation as a signal worth measuring and use AI to make the first step easier. That principle applies well beyond paywall creation.
Mind the Product · Anupama Garani, VP at PIMCO · 11 Jun 2026
Garani frames the first 90 days after launch as a diagnostic sprint rather than a victory lap. Her framework divides the period into three 30-day phases: validate the problem, check the foundations, and assess whether cost and trust are sustainable. Drawing on deployments across government, healthcare and consumer technology, she explains how bypass, rephrasing and editing rates can reveal problems users may not report directly. In one healthcare example, failed queries were attributed to missing data (44%), retrieval issues (31%) and mismatched intent (25%)—each requiring a different response.
Why it matters: This is one of the digest’s most actionable launch guides, particularly for its framework for identifying “silent failures” that users never explicitly report.
SVPG · Marty Cagan · 23 Jul 2026
Teams are using AI to ship faster, yet business outcomes are not necessarily improving. Cagan cites Atlassian’s 2026 data: 89% of executives say AI has accelerated work, but only 6% can point to organisation-wide ROI. He argues that the apparent paradox has a straightforward explanation: many companies are using AI to speed up a project model built around output rather than outcomes. Strong product teams use AI to accelerate discovery first—“build to learn”—and delivery second—“build to earn.”
Why it matters: Cagan offers a compelling explanation for why AI may be widening the gap between high-performing product organisations and those that simply produce more, faster.
-
Experimentation, AI and the Product Model
SVPG · Marty Cagan · 21 Oct 2026
Following on from his productivity-paradox argument, Cagan contends that when building becomes inexpensive, deciding what deserves to be built becomes the harder challenge. Experimentation, he argues, should be a core product competency—not a function limited to A/B testing specialists. He explores how AI can automate optimisation experiments, why AI products need evaluations before A/B tests, and how a hybrid model can work: product teams own experimentation, supported by expert coaching.
Why it matters: This recent follow-up turns the productivity-paradox diagnosis into a practical recommendation for how product teams should operate.
Forbes · Sergei Revzin and Vadim Revzin · 6 Jan 2026
AI is allowing a single AI-fluent generalist to take on work that once required designers, front-end engineers and analysts. But, the authors argue, the fundamentals of product success remain unchanged: form a hypothesis, reach the right customers, learn from their feedback and convert that learning into action. Founders should use tools such as Lovable, n8n and Cursor to build just enough to test an idea—not mistake a prototype for a finished product.
Why it matters: This founder-focused perspective challenges the “just ship it” mindset and reinforces Cagan’s argument from a startup vantage point.
Amplitude · Darshil Gandhi · 5 Jun 2026
Gandhi offers an accessible introduction to traces and trace analysis, code-based evaluations, LLM-as-a-judge approaches, and the difference between offline and online evaluation. Worked examples make the concepts concrete. The guide recommends applying low-cost code checks broadly and using LLM judges on a sample of roughly 5–15% of live traffic. It also stresses the importance of connecting evaluation scores to product outcomes such as retention and conversion.
Why it matters: A useful starting point for product managers who understand the importance of evaluations but haven’t yet built one. It is a vendor blog, however, and concludes with a product pitch.
-
Demystifying evals for AI agents
Anthropic Engineering · Mikaela Grace, Jeremy Hadfield, Rodrigo Olivares, Jiri De Jonghe · 9 Jan 2026
This deeper companion to the Amplitude guide comes from a model developer. It distinguishes capability evaluations, where a low initial pass rate can help establish a baseline, from regression evaluations, which should approach 100%. It also explains why pass^k matters for customer-facing agents and recommends starting with 20–50 tasks drawn from real failures. Examples include Descript, Bolt and Claude Code.
Why it matters: A robust evaluation process can help teams assess and adopt new models in days rather than weeks, turning product quality into a source of launch speed.
-
Why AI adoption is a people problem, not a technology problem
ITPro · Cathie Hall, Chief Product and Customer Officer at IFS · 25 Sep 2026
Hall describes a recurring pattern in industrial AI deployments: around 80% of licensed users log in once and never return. She attributes the problem partly to trust and positioning. Tools presented as performance-monitoring systems can make employees wary; tools positioned as partners that preserve and share experienced engineers’ knowledge are more likely to be used—and improved through feedback.
Why it matters: This is the enterprise counterpart to article #1: adoption can hinge as much on employee perception and trust as on the technology itself.
Productboard · 20 Jan 2026
This recap of Productboard’s AI Product Summit features perspectives from GitHub, S&P Global and Product Team Success. Among the examples, GitHub’s product operations team is building agents to triage incoming requests and critique product specifications. One speaker reports that 87% of product operations teams have AI initiatives, but only 12% see meaningful impact. The recommendation is to connect tools into coherent workflows rather than accumulate isolated point solutions.
Why it matters: The piece highlights product operations’ emerging role in turning scattered AI pilots into integrated systems and measurable outcomes.
ITPro · Angus Ward, CEO of Beyond Now · 20 Oct 2026
Ward examines how AI could reshape go-to-market strategy as buyer and seller agents begin interacting and transacting. In this environment, marketplaces may evolve from catalogues designed for discovery into systems that help agents make decisions. Partners will need to present their expertise in ways AI agents can interpret, while pricing models may increasingly reflect usage and outcomes.
Why it matters: A useful forward-looking perspective on how AI products may be discovered, selected and purchased—and what that could mean for launch and distribution strategies.
The editor’s takeaway
Taken together, these ten reads point to a clear conclusion: building is no longer the main bottleneck for AI product teams. The harder work is earning user trust, measuring quality and identifying the problems worth solving. Adapty and IFS show how fear and distrust can undermine adoption, while the first 90 days after launch can reveal where confidence is won or lost. The practical message is to make trust and user behaviour part of product measurement—not treat adoption as an automatic consequence of shipping a feature.
Evaluation is becoming an essential product-management discipline, too. The Amplitude and Anthropic guides recommend starting with real failure cases, combining code-based checks with model-based evaluation where appropriate, and building evaluations into the product process. Meanwhile, Marty Cagan’s pieces and the Forbes article make the case that faster delivery only creates value when teams first test their assumptions and learn what customers actually need.
The implications extend beyond individual product teams. Product operations is emerging as a way to connect AI tools into systems that deliver measurable impact, while the rise of agent-to-agent transactions could change how products are discovered and bought. The competitive advantage will belong not simply to teams that build fastest, but to those that learn quickly, earn trust and turn AI capability into outcomes customers value.
