AI Strategy

AI strategy: value first, then architecture, then tool

AI strategy settles, before the tools, where AI creates value, whether the organisation builds, buys or hosts, how much risk it carries and in what order. It is the investment decision, not the control.

Most AI initiatives fail not on the technology but on the decision that should have come first. A tool gets adopted because everyone is adopting it, with no one naming the business problem it solves. AI strategy inverts the order: value first, then architecture, then tool. This page describes where AI creates value in a business, how to cut the build, buy and host decision, how much risk the organisation wants to carry, and what a roadmap looks like that delivers something within the first ninety days. It deliberately marks itself off from two neighbours: it is the AI-specific cut of the general Business Strategy, and it is separate from AI Governance, which settles not the where but the how of control. Both are needed.

Where AI creates value

Before any talk of models or providers, every use case belongs in a simple grid with two axes: the value contribution (what it is worth if it works) and the feasibility (how certain and how costly the path there is). The grid separates the four fields that order any AI roadmap:

High value, high feasibility. The first initiatives. A clearly bounded problem, available data, measurable benefit. This is where the roadmap starts, not at the most spectacular use case.
High value, low feasibility. The strategic bets. They belong on the roadmap, but as initiatives with groundwork first (data, capability, architecture), not as an immediate start.
Low value, high feasibility. The temptation. Easy to build, but with no effect on the business. This is where most budget burns in demos no one uses in production.
Low value, low feasibility. Not now. Document and revisit, do not start.

Value rarely sits where the headline is. High-volume routine work, knowledge-intensive research and the structuring of unstructured data carry more in practice than the one generative showcase. That is exactly why the grid starts with the business problem and not with the model's capability. This matches the sober finding from practice: AI fails where a flawed process is automated instead of being rethought first.

Build vs Buy vs Host

Once the use case is fixed, the architecture decision follows, and here lies the AI-specific difference from the classic Make or Buy question. With AI there are not two options but three, because the third decides data sovereignty:

Buy. A ready-made AI feature or a provider model as a service (commodity). Fast, low upfront cost, but the data leaves the house and dependence on the provider grows.
Build. An in-house solution on top of external models or proprietary logic. More control over the product, more effort, the question of model operation still open.
Host. An open-weights model on in-house or Swiss infrastructure. No data outflow into a US cloud, no provider lock-in, but hardware and operating cost and a possible quality gap to the largest closed models.

The third option is the very reason AI needs a strategy of its own. As soon as particularly sensitive data is in play, host is not the expensive special case but often the only one that satisfies data sovereignty and regulatory compliance in one. Once this architecture is built, it runs not for one use case but for many. The sovereign RAG architecture service is what makes exactly this path concrete. The choice is neither permanent nor the same for everything: an uncritical internal assistant may be bought, while research on client data belongs hosted.

flowchart TD
    A["Use case with<br/>clear value contribution"] --> B{"Particularly<br/>sensitive data?"}
    B -->|"Yes"| C["Host<br/>open-weights on own<br/>or Swiss infrastructure"]
    B -->|"No"| D{"Does it differentiate<br/>the business?"}
    D -->|"No"| E["Buy<br/>ready-made service,<br/>commodity"]
    D -->|"Yes"| F["Build<br/>own solution on<br/>external models"]
    F --> G{"Must operation<br/>stay in-house?"}
    G -->|"Yes"| C
    G -->|"No"| H["Build on<br/>provider model"]

The diagram reads top to bottom as a sequence of a few questions: first the data class, then the question of business differentiation, finally the operating question. It does not replace a detailed assessment, but it prevents the most common mix-up, a buy-convenience where the data class has long demanded host.

Risk appetite and roadmap

Strategy also means deciding upfront how much uncertainty the organisation carries. AI initiatives are not deterministic: the model does not answer the same way every time, and the benefit can rarely be quantified exactly before the pilot. The risk appetite sets where a human stays in the loop, what error rate a use case tolerates, and from when an initiative is stopped. These thresholds belong in the strategy, not in the later discussion with the first failure.

Grid and risk appetite become the roadmap. It orders the initiatives not by enthusiasm but by quadrant: the feasible value drivers first, the strategic bets with their groundwork behind them, the temptations deliberately set aside. A robust AI roadmap is part of the Innovation Management practice and not a one-off document; the ongoing scan of the model and tool landscape it rests on belongs to the tech radar. The honest part of the roadmap is the stop criterion: a pilot that does not show the expected value is ended, not rescued.

The first 90 days

An AI strategy proves itself not in the strategy paper but in the first quarter. A proven cut:

Weeks 1 to 4: scan and cut. Collect use cases, place them in the grid, select a single feasible value driver. In parallel, record the risk appetite and the data classes, because they determine the architecture.
Weeks 5 to 8: prove. Build the chosen case as a lean, measurable prototype, with real data and a success measure defined upfront. A fixed-scope Enterprise RAG Proof-of-Concept delivers exactly this proof in a set frame, instead of stretching the learning over months.
Weeks 9 to 12: decide. Measure the prototype against the success measure and make the build vs buy vs host decision for production. If the value holds, scaling follows with the fitting architecture; if it does not, the stop criterion is pulled and the next quadrant chosen.

This cadence keeps the strategy honest. It forces a proven benefit before money flows into scaling, and it makes the architecture decision on a real case rather than on a slide. What the strategy fixes in value and order is then kept steerable in operation by AI Governance, from model approval to evidence.

Data sovereignty as an entry condition

For data-intensive use cases, data sovereignty is not a downstream control point but an entry condition of the investment decision: whoever decides where a model runs also decides who gets access to the data it processes. For a Swiss company, the build vs buy vs host question is therefore also a location question. As soon as personal data flows into a US cloud, the revised Data Protection Act and the US Cloud Act can both become relevant, especially with US providers or US control over the data, and the buy-convenience turns into a compliance risk. For such use cases the strategy's centre of gravity therefore shifts towards the host option, and data sovereignty carries the decision from the outset. The classification of the model and tool landscape that the roadmap rests on is delivered by the Tech Radar and AI Governance service. The neighbouring Neuland pages go deeper into the individual layers: Language Models, AI agents, LLMOps and MLOps, and sovereign AI and the decision between fine-tuning, RAG and prompting.

References

Deloitte Tech Trends 2026. Annual report on the defining technology trends; only a small share of organisations run AI agents in production, and initiatives often fail on automating flawed rather than redesigned processes. (10.12.2025). www.deloitte.com/us/en/insights/topics/technology-management/tech-trends.html
OECD AI Principles, 2024 update. Five values-based principles and five recommendations for trustworthy AI, adhered to by 47 states and organisations. (03.05.2024). oecd.ai/en/ai-principles
NIST AI Risk Management Framework (AI RMF 1.0). Voluntary US framework for steering AI risk across the four functions govern, map, measure and manage. (26.01.2023). www.nist.gov/itl/ai-risk-management-framework
MIT Sloan Management Review and BCG Winning With AI. Study on the value contribution of AI; seven out of ten companies report little or no impact at first, with success lying in the interplay of strategy, organisation and technology. (2019). sloanreview.mit.edu/projects/winning-with-ai/

Ask AI

These links open external AI services, the conversation and its content are sent to their providers.