Working on the AI Strategy Lab enabled me to really broaden my understanding of what it means to develop an AI strategy. Prior to starting the project, I viewed AI adoption primarily through the lens of technological capability. However, throughout the work, I learned that the more difficult matters to solve are often not technical, but rather strategic: involving accounting for various stakeholders (e.g. investors, beneficiaries, etc.), the costs of implementation, accounting for system faults, and the necessary safeguards that need to be implemented before an organization can responsibly use it.
With the common aim of contributing to a positive change in the world, our team came up with an application idea (CareClear), designed as an AI-assisted post-consultation tool for non-Dutch-speaking patients. Namely, after receiving/extracting the patient’s post-consultation letter, the model turns it into a simple action plan in the patient’s preferred language, covering medication intake, follow-up appointments and guidance for further care. This aims to help patients better understand their treatment while reducing reliance on relatives/external assistants, repeated calls to healthcare staff, or informal translation.
Within the team, my main responsibilities revolved around the economics and governance/failure analysis of the product. With hospitals as the paying customers, the economic challenge was to show that the value created for hospitals could justify the cost of the service. Our team proposal was that hospitals would pay an annual license based on the number of patients supported, with pricing logic reflecting that the scarce value was not simply translation, but the combination of clinical verification, workflow integration and human oversight. When working on the cost estimation, one of the most important things I learned was that AI does not incrementally make a business cheap to run. As taught in the lectures, although the marginal AI inference cost can be very low, our product had significant fixed costs related to development, integration, security, and compliance. What I had not expected was how long it would take to recoup the costs – at the high estimated fixed expenses, a single hospital contributes only a small share, and willingness to pay remains a hypothesis rather than a result. Building on to that, despite having a rather simple model, I realized that industry regulations can become real cost sources. As health data falls in the category of special-category personal data, privacy and security was a significant part of the costs and risks of the product. I therefore researched the necessary safeguards, which led us to include control measures, such as a DPIA, NEN 7510/ISO 27001 certification, an EU AI Act review, consent management, and a clear notice to patients of the integration of AI-generated content – all of which substantially increased the costs of our product.
My second major responsibility regarding governance and failure analysis was particularly important because our product deals with medical information, where a small translation or simplification error could have serious consequences. Therefore, I made sure to identify all possible significant failure modes – including meaning-altering simplification of medication instructions, omission of warning symptoms, or hallucinated information – and classify them with an appropriate severity level. One thing that I noticed was that “accuracy” is not a single concept in such context. A translation can be linguistically fluent while still being unsafe. Similarly, I discovered the importance of accounting for the difference between reported and reliable data. As it would be rather difficult for patients to verify the translation on their own, we decided that CareClear should include a comprehension check in the form of a short test through which their understanding of the information is assessed. However, false positive occurrences were a significant gap – a positive confirmation does not guarantee real understanding. It made me aware that human oversight within content review, monitoring of error and verification rates is not just a useful addition, but a rather strict requirement – especially in high-stake fields like the medical one. To address this, we integrated a human-in-the-loop, responsible for high-risk medicine, who would also be auditing a set of randomly selected cases (calculated at approximately 5% of all summaries).
Beyond the above-described sections, I focused on connecting the team’s different pieces and keeping them consistent with each other, which taught me valuable lessons on team dynamics and collaboration. The group’s initial strategy of dividing the tasks and working on them individually resulted in lack of consistency within the final project, as the tasks were not independent from one another. For example, the economics depended on the value proposition, while the governance analysis affected the cost structure – these internal links was something that we initially failed to properly address. Within our governance analysis, once we identified that human review was a necessity for our product, our variable cost estimate completely changed (€0.25-0.81 instead of the initial €0.05-0.07 frame that excluded specialist review costs), changing both the economics and pricing logic. Nevertheless, after noticing these inconsistencies, we moved from working in isolation to regular review sessions in which each member challenged the assumptions in the others’ sections. Discussing our different perspectives helped us refine the business model, pricing, risk analysis, and future features.
Looking back, the project reinforced the lectures’ emphasis that “We implement AI” is not a strategy on its own. In CareClear, the strategy lies in the design choices around the model – restricting it to hospital-approved content, verifying its output against the connected source or escalating to human review when necessary. “Full automation” is not necessarily the right objective – human escalation, monitoring, and carefully defined boundaries become part of the product itself. Furthermore, I concluded that strong AI capabilities do not automatically make the business idea strong – in our case, the low marginal cost of AI was attractive, but the real strategic challenge was creating enough trust and measurable value to justify adoption. Finally, I figured out that sequence in strategy is equally important. In CareClear this translated to a controlled pilot against usual care, with explicit milestones (e.g. better next-step accuracy, no uncaught critical medication errors, per-language safety) before anything scales. Since the model itself is the part competitors can copy, the value and risks of it lie in how it is built into the organisation.
I now see a successful AI strategy as one that balances value, economics, feasibility and responsibility at once, with the central questions addressing where AI is allowed to act, who checks the output of each feature of the model, who is accountable for encountered failures, and what evidence would determine whether the project should scale or stop.
