Issue 010 — Strategy, Infrastructure and the AI Stack

Most AI strategy documents are written long before the team has built anything in production. Most production systems are built without much reference to the strategy document. The gap between the two is where most of the wasted spend sits, and it is wider in most organisations than anyone is comfortable admitting.

The AI stack is where strategy and production are supposed to meet. A coherent stack has three layers. The model layer covers which models you run, where they run, and how you route between them. The data layer covers what knowledge you ground those models in, where it lives, and how you keep it current. The orchestration layer covers how the pieces communicate with each other, where human review is inserted, and how outputs flow back into systems of record.

Most teams optimise one of these layers and leave the others underspecified. The model layer gets attention because it is visible and model benchmarks are easy to compare. The data layer gets partial attention when something breaks. The orchestration layer, which is the layer that determines whether AI actually changes how work gets done, is frequently an afterthought. McKinsey's March 2026 research notes that nearly two-thirds of enterprise AI efforts remain stuck in pilots, and attributes the stall to leadership adoption, people readiness, and workflow integration. These are orchestration problems wearing an organisational costume.

The infrastructure is often in place. The tool contracts are signed. The organisation, the skills, and the operating model sitting alongside the technology are frequently not ready. That is what the numbers are actually measuring.

If you are heading into a planning cycle with an AI initiative on the table, the Prompt Pocket this week walks through all three stack layers and surfaces the gaps that tend to emerge six months in rather than before the build starts. Use it before you commit the budget.

Most AI investment conversations are framed as build versus buy. That is the wrong axis for most organisations. The sharper choice is between infrastructure investment and capability investment.

Infrastructure investment is concrete. It appears on a budget line as model contracts, orchestration platforms, integration work, and security review. It earns back over time through reduced unit cost and faster product velocity once deployed. Capability investment is less visible. It lives in the team's skill stack: prompt engineering, evaluation discipline, AI-aware product design, and governance that gets used rather than filed.

Boards approve infrastructure more readily. It looks real. Capability looks soft. Six months later, the infrastructure exists, the dashboards exist, the contracts exist, and the team is not using any of it well, because the capability layer was underfunded from the start.

The corrective is simple to state and difficult to implement. No infrastructure investment should be approved without a matching capability investment alongside it. Two budget lines, not one. If the capability line cannot be funded, the infrastructure line should not proceed either, because the infrastructure is unlikely to deliver value without it.

In practice, this means training time, hiring for new skills, tolerance of slower ramps, and early inefficiency. None of these score well on a standard return on investment template. They earn back asymmetrically. The second AI initiative is dramatically cheaper and lower risk if the first one included the capability investment the stack actually needed. If it did not, the second initiative starts over.

The hardest version of this argument is the most practical one: if a board will not fund the capability line, that is a signal about organisational readiness, not about the technology. It is worth surfacing before the infrastructure spend is committed, not six months afterwards.

Vercel AI SDK is a free, open-source TypeScript toolkit that gives you a single unified interface for working with multiple model providers including OpenAI, Anthropic, Google, and custom or self-hosted endpoints. Rather than wiring each provider's API separately, the SDK handles the abstraction layer, so adding or swapping providers becomes a configuration change rather than a rebuild. Streaming responses and tool calling are handled cleanly across providers, which is where integration effort typically accumulates in manual builds.

The optional AI Gateway adds model routing so different workloads go to different models without changing application code. It operates at provider list price with no markup; a $5 monthly credit is included, then billing is usage-based. For TypeScript teams on React or Next.js, where documentation and examples are strongest, the time-to-integration advantage is real.

The first-class developer experience leans toward React and Next.js, though the SDK also supports Vue, Svelte, and Node.js. Teams outside the React ecosystem or building primarily backend or agentic architectures will find documentation thinner and need more independent exploration. This is an editorial observation from working with the toolkit, not a limitation in the SDK's capabilities.

Try Vercel AI SDK →

Some links are affiliate. Zymbos AI may earn a small commission at no extra cost to you.

There is a version of AI adoption where the model is the strategy. Pick the best one, integrate it, wait for results. Most organisations have discovered this does not work, or they are currently in the process of discovering it.

The stack is the strategy. Not because of the technology itself, but because a coherent stack forces you to answer questions that strategy documents tend to skip: what data are you grounding the model in, where is human judgment inserted, and how are outputs actually changing how decisions get made? These are organisational questions, not technical ones.

The capability investment argument in the Deep Intelligence this week is, at root, the same point. You cannot buy your way to an organisation that uses AI well. The technology is ready. The harder work is building the team and the operating model alongside it. That is still where most organisations are behind, and it is where the real returns are waiting.