Generative AI in the public sector: three keys to procuring wisely

Generative artificial intelligence has firmly entered the public sector’s digital agenda: almost half of Spanish administrations are already using it for chatbots, automation and workflow management, with public spending expected to exceed €30 million in 2024.

Yet reality often falls short of expectations. Over 70% of public-sector generative AI projects never move beyond the pilot stage, according to a study of 4,372 initiatives by the technology consultancy Nortal. In the private sector, MIT research shows that 95% of AI projects never reach full-scale implementation.

This makes it clear that a structured approach is essential if investment is to deliver real impact. At Gobe, we highlight three key principles for procuring generative AI successfully:

1. First the challenge, then the AI

Being clear about the challenge you want to solve and the outcomes you expect is essential. The goal is not to acquire AI for its own sake, but to understand how it can improve citizen services, optimise processes, or manage information more efficiently.

Looking at how similar challenges have been tackled elsewhere helps you assess whether generative AI is truly necessary. Often, the same results can be achieved by adjusting internal processes or making better use of existing digital tools.

Hypothetical example: a network of public libraries wants to use generative AI to recommend reading. After analysing the problem, they find it’s enough to improve subject tagging, add more precise metadata, and include intuitive filters in the catalogue - achieving useful recommendations without resorting to generative AI.

Even when generative AI is confirmed as the right solution, clearly defining the use case and objectives from the outset helps guide implementation and measure impact effectively.

2. New technologies, new procurement processes

Procuring generative AI is not the same as buying traditional software. Alongside the major vendors, there are specialised start-ups and SMEs offering more innovative solutions tailored to specific needs. Limiting yourself to the big players means missing opportunities in a market growing by nearly 40% a year.

In this context, piloting is crucial. Testing solutions before drafting tender documents or committing resources makes it possible to assess how the technology works in a real environment - using your own data and processes - and to detect potential risks before scaling.

In Spain, the Public Sector Contracts Law (LCSP) provides concrete tools for this. For example, Article 183 allows the use of a project competition to pilot different solutions within the same process and, subsequently, to move to a negotiated procedure without publicity to select the best fit. As we noted in our piece on public procurement, this mechanism enables legal, structured experimentation before committing major resources.

In short, looking beyond the big players and piloting first reduces risk and helps identify the solution that truly fits your needs.

3. Buy today with tomorrow in mind

As noted at the outset, many AI pilots do not progress beyond the test phase. This is not always a problem: sometimes the decision not to scale is because the solution does not meet the objectives. However, recent studies from MIT and the OECD show that, quite often, the cause lies in a lack of long-term planning. To ensure a successful solution can be implemented, we highlight two key aspects to consider:

• Technology integration. One of the public sector’s main challenges is legacy systems, which operate in silos and make it difficult for AI systems to access quality data. Maintaining these systems - obsolete yet essential - accounts for around 37% of public IT spending in Spain.
  
  Before acquiring any solution, it is vital to assess the existing technology infrastructure and data-sharing protocols, including the legal feasibility of sharing information between systems. It is also advisable to verify that new solutions include APIs (standard ways for systems to communicate) or middleware (a “bridge” that enables different systems to work together), ensuring interoperability and a cohesive, efficient technology ecosystem.
  ‍
• Costs. Generative AI is priced differently from traditional software. Organisations typically pay via subscriptions, pay-as-you-go, or a mix of the two, and costs can rise quickly as adoption scales. As a report from the Cotec Foundation explains, a key challenge is correctly identifying the billing units (tokens, API calls, users, etc.), as these directly impact the monthly cost and can create unexpected increases after a successful pilot. Pilots therefore make it possible to project real costs, negotiate usage caps, and anticipate additional expenses related to deployment, integration and maintenance - factors that are less predictable than with traditional, fixed-licence software.

In short, procuring generative AI in the public sector requires planning and a long-term view. Done well, this technology can significantly improve the efficiency and quality of public services. For concrete examples of how AI is being applied in the public sector, you can read our article.

We would love to hear about your experience with generative AI projects in the public sector and what has worked in your day-to-day practice. If you would like to share ideas, data or perspectives to help improve the implementation of these technologies, get in touch at leyre@gobe.studio.