Build Dynamic Products, and Stop the AI Sideshow

If every product leads with AI, what does that tell a customer about which product solves their problem? Eliza Cabrera approaches that question as a principal AI product manager at Workday, building its financial audit agent after working on policy-agent go-to-market, assistant early access, and early generative AI features. Jeremy Silva brings a data science and machine learning background, including language-model work before GPT-3, and leads product at Freeplay. Their shared concern is what happens when technological exploration becomes the product strategy.

Leading products, marketing, and strategy with AI initially signaled that a company was participating in a technological shift. Once everyone adopted the same positioning, that signal lost its distinction. Much of the resulting product work was valuable exploration: teams were building things to discover what the technology could do. But discovering a capability is not the same as discovering a differentiated customer solution.

The early sequence was understandable. After an initial encounter with LLMs, teams used chat interfaces, content editors, and existing applications to explore their boundaries. Multimodal inputs and outputs expanded what those experiments could accept and produce. Then vector databases and retrieval-augmented generation (RAG) offered ways to ground responses in relevant information, while larger context windows and more memory increased what a model could work with. The aim was greater accuracy, even where the underlying notion of truth remained contested.

Copilots reflected another constraint: people were more comfortable with an assistant beside them than with software taking over work or displacing jobs. As the appeal of automation became clearer, agents offered reasoning, tool use, and API calls to coordinate work across business problems. These approaches helped teams understand the technology. Their widespread adoption, however, meant that choosing chat, RAG, or agents could not by itself establish a differentiated strategy.

Silva describes a recurring pattern from Freeplay’s work with enterprise companies: leadership correctly prioritizes AI, then creates a centralized AI strategy that runs alongside the core product strategy. Separate initiatives, and sometimes separate teams, naturally produce features that feel disconnected from the rest of the product. The organizational separation becomes visible in the customer experience.

Three decisions reinforce that separation:

• Quarantine the risk. Teams confine AI to isolated product corners because reliability is uncertain. The underlying question is real: can this feature work consistently enough to create customer value?
• Start with the technology. A team builds a chatbot because it demonstrates AI, without establishing that customers struggle with support. Another adds summarization without evidence that users face information overload.
• Prescribe solutions from above. Leadership chooses the feature instead of setting direction and allowing product teams to discover the right solution.

AI reliability belongs inside product planning. Instead of isolating uncertain functionality, teams need evaluation systems, tests, and prototypes that make its risks assessable. Those practices help establish whether a proposed feature can deliver enough value, and what must improve before it can do so reliably.

Discovery still starts with the customer’s problem. Leadership supplies the strategic direction; product teams close to customers need room to experiment, prototype, and abandon unsuccessful approaches quickly. Aligning strategy, teams, and roadmaps makes that learning part of the main product effort. The resulting experience should feel cohesive: it can solve the problem better without needing to announce itself as AI.

Deep integration does not require rebuilding the entire product at once. Silva’s crawl, walk, run framework develops capability incrementally while establishing a foundation for later work. Consider his hypothetical customer-support SaaS company: it already has a mature shared inbox where support teams handle customer questions.

In the crawl stage, retrieved questions help a support representative ground a response. In the walk stage, a draft gives that person a starting point; new interface elements need not yet require a fundamental redesign. Running changes the scope of action: the agent operates across the product and feature set, so the core architecture and user experience must accommodate that work.

The stages accumulate. Search remains useful when the product starts drafting, and those capabilities become foundations for broader automation. Even the crawl stage is embedded in the customer’s workflow. It is a smaller integrated capability, not a disconnected feature waiting for a future redesign to make it useful.

Workday’s example begins with employee self-service. Employees need quick answers to workplace questions; when an answer is insufficient, they need a case or a live support person. Workday Help combines the two supporting systems: a knowledge base and case management. Its initial generative AI features improved the content that would supply those answers.

An author could upload a benefits policy more than twenty pages long and generate an employee FAQ or manager talking points in a consistent format. Cabrera describes enterprise content teams of roughly three to fifteen people, where maintaining a unified voice matters alongside reducing the work of writing each article. Generation was therefore part of an authoring workflow, rather than a separate destination for prompting a model.

Translation extended the same workflow. Cabrera reports that Workday Help supported translation into 34 languages in a couple of clicks. The author could keep a base article, generate English talking points, and manage French, Spanish, or Japanese translations as versions in the left-hand panel. The demonstration pairs policy content and manager talking points with translation controls, making multilingual content management part of the editor.

The interface was designed in 2023 for users who did not need to understand generative AI to benefit from it. It avoided making chatbots, prompt fields, and decorative sparkles the entire experience. Human review remained in the workflow, with an AI disclaimer and limited purple sparkles identifying AI functionality. Cabrera dates these Help features’ general availability to August 2024, the R2 release. This was the crawl stage: useful authoring and translation inside an existing product.

Once content exists, the next problem is helping employees use it while completing a task. In Cabrera’s walk-stage example, a manager named Elaine is changing a location to San Francisco. She understands many of the form fields, but not all of them. Workday Assistant provides a contextual copilot across Workday, with suggestions informed by the page she is using.

That context raises the stakes for data processing. A help article contains sensitive customer content, but assistance inside a task can encounter personally identifiable information, pay, or compensation inputs. The expansion also crosses organizational boundaries: a platform capability must work across human capital management and financials, including benefits, procurement, and core HCM. Shipping it requires stronger coordination between leadership’s direction and the teams implementing individual workflows.

For the run stage, Cabrera connects the same employee and manager use case to Workday’s Agent System of Record, announced a few months before the talk. She describes agentic capabilities operating behind Workday Assistant, so users can benefit without understanding agents. The direction is greater autonomy and proactivity: listen for policy changes, then notify people with suggestions. The February 2025 announcement described the Agent System of Record and new agents as still in development; this example presents the direction of the experience, not a confirmed general-availability milestone.

The coordination burden grows with that scope. A capability that began inside one product within a SKU now extends across the core platform, requiring top-down strategy and bottom-up execution to connect multiple product experiences. Cabrera describes Workday’s customer reach as roughly 60% of the S&P 500; the contemporaneous announcement instead reports more than 60% of the Fortune 500. These are different company groups, and the figures describe customer coverage, not adoption of these AI features.

Stopping the AI sideshow does not mean abandoning agents or the underlying technology. These capabilities are stepping stones through which an organization learns to build more useful experiences. Cabrera connects that progression to purposeful, vertical-specific products, echoing earlier talks by Sarah and Brian. The opportunity is to use the accumulated capability to address a broader problem space.

More powerful technology can leave a team solving the same old roadmap faster. Digitizing new data, including inputs from physical environments and spaces, can instead change which problems a product can address. Multimodal experiences make that especially compelling when they work together with little friction. Cabrera briefly identifies interoperability and reinforcement learning as relevant to agents; the broader destination is dynamic products that respond to their environment, extending what the product can do as its understanding of the surrounding world expands.