How to Align AI Product Strategy With Infrastructure Realities

You've probably been in a meeting like this before. You're there, taking it all in, as someone pulls up a slide with the latest AI model benchmarks, another person pitches a new agentic workflow, and everyone nods enthusiastically. Then, three months later, the pilot stalls because nobody planned for what happens when you need to run thousands of background agents at scale. The model worked great in testing. The infrastructure couldn't keep up.

That disconnect between AI ambition and infrastructure reality is exactly what Jim Fairweather, head of AI go-to-market at Google, tackled in his session on aligning AI product strategy with infrastructure realities in the agentic era at Domo BUILD 2026. Jim's core argument is refreshingly blunt: Your AI product outcomes are strictly bound by your underlying compute capacity. No amount of clever prompting or model selection will save you if your infrastructure can't deliver.

What follows is a practical framework for aligning AI product and infrastructure strategy, built from Jim's session insights. You'll walk away with a clear playbook for treating capacity as a feature, designing for multi-model resilience, and getting your leadership team aligned on the metrics that actually matter.

Treat capacity as a product feature

Infrastructure planning starts at the design phase, not after launch

The first shift in thinking is to stop treating infrastructure as something that gets figured out after you've designed the product. Jim made this point vividly: "If your product strategy doesn't include an explicit and co-designed infrastructure strategy, you're building a beautiful sports car with a garden hose for a fuel line. It will choke."

Every token your agent generates has a physical cost in watts, amps, and thermal dynamics. That's not an abstract concern for the ops team to handle later. It's a constraint that shapes what your product can actually do.

Here's how to put this into practice:

• Map compute requirements during product scoping: Before finalizing any AI feature, estimate token volumes, context window sizes, and expected concurrency. These numbers determine whether your infrastructure can support the experience you're designing.
• Include infrastructure leads in product reviews: If your product team and infrastructure team aren't in the same room during design, you're setting yourself up for surprises at scale.
• Budget for capacity alongside features: Treat compute costs as line items in your product budget, not as overhead that gets allocated after the fact.

Design for a multi-model portfolio

Route tasks dynamically and build in redundancy for resilience

The era of picking one model and running everything through it is over. Jim called this the "model-only illusion," noting that code is a level playing field, not a competitive advantage. Everyone has access to increasingly powerful models. What differentiates you is how you orchestrate them.

A portfolio approach means matching each task to the right model based on complexity, cost, and latency requirements. Simple, high-speed actions go to lightweight models. Complex multi-step reasoning gets routed to more capable (and more expensive) models. Multimodal tasks go to models built for that purpose.

Beyond routing, you need redundancy. Jim described building "a modular abstraction layer so that if a model has a bad Friday, your system doesn't even blink." This means your architecture shouldn't depend on any single model or provider staying available and performant.

To implement this approach:

• Audit your current workflows by complexity: Identify which tasks genuinely need your most powerful model and which could run on something faster and cheaper.
• Build abstraction layers: Design your system so you can swap models without rewriting application logic. This protects you from provider outages and gives you flexibility as the model landscape evolves.
• Test failover scenarios: Don't just assume redundancy works. Simulate model degradation and verify your system routes around it gracefully.

Ground agents in governed enterprise data

Data hygiene is a business imperative, not an IT task

Your models are only as good as the context you feed them. Jim emphasized grounding models natively into enterprise data and called data hygiene a foundational business imperative that creates sustainable competitive advantage.

This isn't about having a data warehouse. It's about ensuring your agents can access clean, governed, contextually relevant data in the formats they need, when they need it. Without that foundation, you're asking AI to operate blind.

The practical steps here are straightforward but often underinvested:

• Establish a unified data foundation: Consolidate your data sources so agents don't have to navigate fragmented systems to get context.
• Treat data quality as a product metric: Track and report on data freshness, completeness, and accuracy the same way you track product KPIs.
• Secure access with governance built in: Agents need data access, but that access needs to respect your existing security and compliance requirements from day one.

Measure ROI in stages

Track utilization, depth, and breadth as you scale toward autonomous operations

Jim outlined a three-stage framework for measuring AI ROI that maps directly to infrastructure planning. The stages are utilization, depth, and breadth.

Utilization is table stakes: are people using the AI tools? This is where most organizations start, measuring individual time savings and adoption rates.

Depth measures how much of an end-to-end workflow you've automated. When you automate entire workflows, you eliminate bottlenecks and start seeing organizational impact, not just individual productivity gains.

Breadth is where the game changes. This is about building autonomous operations where systems don't just monitor the business but act on it. Jim warned that reaching this stage hits physical reality hard: orchestrating thousands of background agents and running millions of tokens of context requires enormous capacity.

Plan your infrastructure investments against these stages:

• Stage 1 (Utilization): Infrastructure needs are modest. Focus on reliable access and acceptable latency for individual interactions.
• Stage 2 (Depth): Workflow automation increases token volumes and concurrency. Plan for sustained load, not just peak capacity.
• Stage 3 (Breadth): Autonomous operations demand significant scale. Budget for the power, cooling, and compute that continuous agent orchestration requires.

Align leadership on cost per outcome

Get product, engineering, and finance sharing the same metric

The final piece is organizational. Jim argued that the chief product officer, chief technology officer, and chief financial officer need to sit at the same table, sharing a single unifying metric: cost per outcome, not just cost per token.

Cost per token is easy to measure but misleading. It doesn't tell you whether your AI investments are delivering business results. Cost per outcome ties infrastructure spend directly to the value you're creating.

To make this work:

• Define outcomes that matter: What business results are you trying to drive? Customer resolutions? Processed transactions? Decisions made? Pick metrics that connect to value.
• Track infrastructure costs against those outcomes: Build reporting that shows not just what you're spending on compute, but what you're getting for it.
• Review together: Make cost per outcome a standing item in cross-functional leadership reviews. When product, engineering, and finance see the same numbers, they make better decisions together.

Put this framework into action

Before your next AI pilot moves into production, test it against the framework. Start with the business outcome, then estimate the capacity, token volume, concurrency, and latency required to deliver it reliably. Match each task to the right model, confirm that agents can access governed enterprise data, and identify where the initiative sits across utilization, depth, and breadth.

Bring product, technology, and finance leaders together before demand exposes gaps in capacity, cost, or governance. Aligning early can help you choose stronger investments, reduce avoidable rework, and build AI products that continue to perform as adoption and autonomy grow.

Jim’s full session goes further, covering how to prioritize AI initiatives, limit agent sprawl, and prepare organizations for institutional AI. Watch it at Domo BUILD 2026 to get the complete picture and hear Jim's Formula One analogies firsthand. The gold rush era for AI experimentation is winding down. Operational excellence is what wins now.