Product at Heart 2025 offered diverse perspectives on the AI era we’re experiencing – from practical frameworks for prioritisation to transformational strategy. Once the videos come out, I’ll pop back and link those.
Henrik Kniberg positioned AI agents as “permanent interns”, capable but fallible tools that need structured environments, human feedback, and shared spaces to operate effectively. He emphasized tight scoping, audit logs, and human review to manage agent risk and enable safe, useful delegation.
Dominik Faber described building AI agents for recruitment, emphasizing the importance of CAIR (Confidence in AI Results) to balance value, error impact, and correction effort. His company uses agents in a transparent, efficient, and practical way to address high-volume recruitment.
Zamina Ahmad challenged companies to move beyond using AI as a tool and instead reimagine workflows, roles, and organizational learning to become truly AI-native. She emphasized the importance of hybrid human-AI systems and warned against the “tool trap” of shallow automation.
Jonathan Evens contrasted the value-focused ML era of 2012 with the GenAI bubble we’re in right now, advocating for product-led AI integration grounded in core user needs and gradual experimentation. He urged companies to build trust through transparency, thoughtful UI changes, and user model elasticity.
Transform
Zamina Ahmed argued that gaining efficiency through AI is not a winning strategy. Competitive advantage will come from transforming workflows and roles, in order to achieve greater impact. Ahmad emphasised that tinkering was over — the goal must now be to evolve and transform.
Jonathan Evens harked back to the ML era, 2012-2023, when ML (AI) products needed to provide business value, typically by providing unique routes to solve high scale problems such as analysis of satellite imagery – the current AI bubble is not always iterating towards value or even focussing on outcomes.
(I was reminded of Gibson Biddle’s DHM model: product advantage comes from delighting users in hard-to-copy margin-enhancing ways. Efficiency is table stakes. Simple use of an LLM via an API is not hard to copy. You will win in your market if and when you find the route to transform the way you drive outcomes by combining AI capabilities with unique elements across your product or organisation.)
Designing for agents
Both Dominik Faber and Henrik Kniberg talked about agentic workflows and how to control and collaborate with them. You need a surface to instruct them and refine their instructions. You need to control tool access.
Agents also demand governance controls. Teams must tightly scope tool access and privileges as well as monitor agent actions. (I was reminded of Simon Willison’s Lethal Trifecta, it’s very easy to land in a state where you’ve compromised security and opened up your organisation to data theft.)
Kniberg likened agents to enthusiastic interns, promising, but in need of supervision and fine-tuning. Don’t expect to set and forget, you will need to iterate on their instructions and privilege.
Faber’s organisation has embedded Forward Deployed Engineers (for more on the FDE role, listen to this podcast) with their customers. The FDE role allows for really fast iteration on a product solution for a customer and fits the extraordinary pace of change in the AI world at the moment as well as the uncertainty for what works.
What and where to AI?
Both Dominik Faber and Henrik Kniberg offered different complimentary models to consider in prioritisation: Faber focussed on risk and reward, and Kniberg focussed on time and fit.
Faber outlined a Confidence In A I Results (CAIR) score, calculated by Value ÷ (Error Impact × Correction Effort) or “how much do I potentially get out of this vs what could possibly go wrong”. The model allows stack ranking your possibilities giving one perspective for prioritisation. A more product market oriented approach.
Kniberg’s model took into account “quantity of time required” and “is a good use of my time”, then blended the amount of intelligence required. Tasks that consume time, offer low value, but require moderate intelligence are ideal candidates for agents. The model allowed his tool, focussed on helping organisations leverage AI, to judge areas for greatest return.
Nascent pricing strategies
As a sidenote to all this: I’m observing most products are charging by token consumption plus a margin. I suspect this is natural caution on behalf of product companies, wanting to pass on the liability and risk of running up accidental bills. Will we see pricing shift toward value, aligning cost with customer outcomes? Perhaps this is a natural subsequent step that we’ll see after the product industry matures out of this initial gold rush period.
This being so, so what?
AI is now table stakes in terms of efficiency, in both the organisational and the personal realms. The challenge to product teams is to rethink what can be done to provide customer value and delight, and business impact using AI, above and beyond time savings and productivity.
Value, delight, and impact, ’twas ever thus.
Leave a Reply