Working with an AI partner: considerations for a strong R&D proposal
Based on lessons learned across multiple AI innovation projects, this article describes practical characteristics of a strong AI partner and a well-structured collaboration. They can be used as a checklist when developing an R&D funding proposal and help increase the likelihood of your proposal being accepted.
Many organizations turn to external AI partners when they want to start an innovation project but lack deep in-house expertise. This is a sensible approach — AI projects are complex, multidisciplinary, and often experimental. At the same time, the number of companies positioning themselves as AI experts has grown rapidly, with many making similar promises of strong expertise and excellent results.
In this crowded landscape, success depends less on bold marketing claims or advanced technology and more on how well the partnership is structured and how well it fits the project. Selecting the right AI partner is therefore not just about technical capability, but about alignment in expectations, ways of working, and long-term goals.
To understand how these considerations fit into the broader context of how a proposal is evaluated, read the evaluation guide
1. Ensure a realistic division of effort
A strong proposal does not assign most of the effort solely to the AI partner. Even when AI is essential, the project is about more than building models and implementing algorithms. Success depends on substantial input from you, the non-AI partner, who provides domain knowledge, business context, and ownership of the results.
This starts with defining clear use cases and requirements. You must explain what problems need to be solved, how the solution will be used, and what acceptable performance looks like in practice. Without this input, AI development risks optimizing for the wrong objectives.
Providing relevant data is another major contribution. It typically involves extracting it from multiple sources, integrating and cleaning it, and explaining what it represents. Even without deep technical expertise, you must clarify how the data should be interpreted and what the model is expected to predict.
Evaluating results is also essential. Beyond technical metrics, you must assess whether the outputs meet real needs and support decision-making. Integration into existing systems, workflows, and user interfaces likewise requires close involvement from you.
Finally and ultimately, you are responsible for turning project results into business value — validating the solution with end users, building trust, and growing the business based on the outcomes. It must be clear that AI is a means to that end, not the end itself.
2. Ensure the right technical expertise on both sides.
A strong AI partner involves people with solid technical backgrounds, such as formal education in computer science or applied AI. At the same time, you should have concrete plans to hire or develop similar expertise internally. This signals a long-term commitment to using AI and reduces the risk of becoming permanently dependent on the external partner.
This level of expertise is important because AI projects are inherently exploratory. It is rarely clear upfront which models, approaches, or data representations will work best. A team needs to experiment, compare alternatives, and adapt as new insights emerge. Doing this effectively requires people who understand the underlying methods, not just how to use tools.
That understanding also becomes critical when things do not work as expected — which is common in innovative projects. Being able to use AI tools is not the same as understanding how they work. Diagnosing unexpected behavior, identifying root causes, and making meaningful improvements requires insight into how the models and methods behind the tools work.
Moreover, the AI field evolves rapidly. New models, techniques, and best practices are released at a high pace, often first described in scientific literature. Being able to follow these developments, assess their relevance, and apply them quickly and responsibly requires training and experience in reading research, implementing methods reliably, and judging trade-offs.
Finally, AI systems are software systems. They must be built according to sound software engineering practices, tested, monitored, and maintained over time. Having strong software engineering expertise ensures that solutions are not only innovative, but also robust, maintainable, and fit for long-term use.
3. Focus on relevant domain expertise
When selecting an AI partner, it is important to look for expertise that matches your specific problem domain, rather than broad, generic AI capabilities. Many AI partners position themselves as being able to solve almost any AI problem, using similar marketing language and success claims. In practice, however, AI expertise is often highly specialized.
Today, there are many partners operating in clearly defined niches, such as (3D) vision, recommendation systems, traffic and mobility analytics or planning. These specializations exist because applying AI effectively in each of these areas requires dedicated experience and accumulated know-how.
Partners who have already worked extensively in your domain bring more than general technical skills. They understand typical data characteristics, common challenges, and realistic performance expectations, and they stay up to date with relevant models, methods, and research developments. This allows them to move faster, make better design choices, and avoid pitfalls that generalist approaches often encounter.
In addition, domain-focused partners often have existing platforms, tools, or reusable components tailored to their area of expertise. Being able to plug your use cases into such a foundation can significantly accelerate progress and reduce development effort, allowing the project to focus on creating value rather than rebuilding standard functionality.
4. Choose technology for the problem, not the other way around
A strong AI partner avoids technology push. While large language models, deep learning and reinforcement learning are currently very popular, not every problem requires such sophisticated techniques. In practice, simpler approaches are often sufficient and more appropriate.
Classic machine learning methods — such as linear or logistic regression, clustering, decision trees, or gradient boosting — can deliver excellent results with far less data, compute, and development effort. More complex models typically require large datasets, significant infrastructure, and long training times. Few organizations have, or even need, the required internet-scale data or the resources to train models for days or weeks.
Pragmatic partners select technology based on the problem and the available constraints, not on trends. They start with the simplest approach that could work and only introduce more advanced methods when there is a clear need. In some cases, this may include fine-tuning an existing foundation model, provided it is trained on data close to the target domain. This focus on fit-for-purpose solutions leads to faster progress, lower risk, and more sustainable outcomes.
5. Make knowledge transfer tangible
Many proposals claim to support “knowledge transfer” and “co-creation” between you and the AI partner, but these terms only have value if they are backed by concrete ways of working. Effective knowledge transfer does not happen through presentations or occasional status meetings alone.
What matters is close, day-to-day collaboration. Having a dedicated expert from your AI partner work alongside your expert(s) on a regular basis — sharing decisions, reviewing results, and solving problems together — is far more effective than loosely defined interactions. This kind of collaboration allows knowledge to be absorbed naturally as part of the work.
In proposals, this should be visible in the work program. Tasks should explicitly be executed jointly by you and the AI partner, rather than being fully outsourced. When collaboration is embedded in how the work is planned and executed, co-creation becomes real rather than just a buzz word.
6. Plan for deployment, maintenance and continuity
In the proposal, it’s important to describe how the project moves beyond experiments and prototypes, including deployment, integration, monitoring, and long-term maintenance. A strong AI partner is open to collaborating under a model where you retain ownership of the data and models generated during the project, and where vendor lock-in is avoided. They should be willing to support deployment on your premises if needed and help your team maintain and extend the solution over time. Partners unwilling to agree to these principles may not be a suitable fit for long-term collaboration.
Key takeaways
Successful AI innovation projects aren’t just about using the newest models and algorithms. They are about making practical choices, handling uncertainty, and building skills and systems that last beyond a single project. With that in mind, the following recommendations summarize what to look for in a strong AI partnership and how to create proposals that lead to real, lasting results, with a much larger chance of acceptance.
- Structure the work realistically: clearly define the division of effort between AI tasks and the contributions of your internal team. Ensure the project is not treated as AI-only.
- Ensure the right technical expertise: collaborate with partners who have deep technical backgrounds and make plans to build similar skills internally to reduce long-term dependency.
- Understand the AI tools: make sure the partner can go beyond simply using AI tools and has enough insight into how models work to debug, adapt, and improve them over time.
- Prioritize domain-specific experience: choose partners with proven expertise in your specific field, including familiarity with relevant data types, and state-of-the-art techniques.
- Be pragmatic about technology choices: avoid “technology push” — use methods that fit the problem, whether simple machine learning or fine-tuning a foundation model, rather than defaulting to the latest hype.
- Foster close collaboration: ensure knowledge transfer and co-creation happen through day-to-day collaboration rather than distant reporting.
- Plan for deployment and long-term use: describe in the proposal how models will be deployed, integrated, monitored, and maintained beyond experiments and prototypes.
- Clarify ownership and avoid lock-in: data and model ownership should remain with your organization, with clear agreements to support long-term maintenance and the ability to switch partners if needed.