AI as a service: how the mid-sized company can use artificial intelligence without building it

Everyone uses it, few scale it

McKinsey’s 2025 global survey on the state of AI showed that 79% of organizations use generative AI, up from 33% in 2023, and that 88% use some type of AI in at least one function. However, only 7% consider that they have fully scaled it (McKinsey). In procurement, for example, a McKinsey analysis reviewed by Industry Today estimates that AI adoption is about six times lower than in marketing and sales, even though its productivity potential is 25% to 40% (Industry Today).

For a mid-sized company, these figures have a practical reading. The technology is available and affordable. What is scarce is the ability to turn it into a sustained change in the way people work.

From the tool to the workflow

The most common mistake is starting with the tool. The company buys licenses for an AI assistant, distributes them and expects productivity to improve. In our diagnostic work with marketing teams, we found that the use of free versions of AI assistants and basic design tools did not translate into better results, because there were no automated workflows for content generation, metrics analysis or task management.

The alternative is to start with the workflow. Which process consumes the most time, has the most errors or delays decisions the most? Which steps in that process are repetitive and information-based? Those steps are candidates for AI.

Cases that already work in the mid-sized company

In our work we have found AI use cases with a good effort-to-value ratio. One was placing outdoor advertising for a distributor, cross-referencing foot-traffic and point-of-sale data to recommend where to invest. Another was using AI assistants so that small business owners could carry out a guided organizational self-diagnosis, with structured strategic questions based on the 7S framework, in a training program for an industry association. A third was the design of quoting and demand-forecasting tools within a transformation portfolio.

What they have in common is that they solve a specific problem for a specific user, with data the company already has or can obtain.

Why contract it as a service

For most mid-sized companies, building their own AI capabilities (a team of data scientists, infrastructure, model maintenance) is neither viable nor necessary. AI as a service gives modular access to those capabilities: a partner designs, implements and operates the solution, and the company pays for the result or for usage.

This model has conditions. First, the use case must have a business owner within the company, not just at the vendor. Second, there must be a result indicator agreed from the outset. Third, the company must retain ownership of its data and the ability to switch vendors. Fourth, there must be a human who reviews and approves what the AI produces before it reaches a client or a critical decision.

The human in the loop

Human oversight is not a concession to caution; it is a design principle. Generative AI models can produce incorrect content with an appearance of certainty. In commercial applications, such as drafting proposals or responding to clients, the most effective model is for the AI to prepare drafts and a person to approve them. In our own work we apply that rule: AI speeds up the preparation, the partner approves what goes out.

Measure to scale

An AI pilot is scaled when it demonstrates a measurable result. We recommend defining, before starting, a baseline (how long the process takes today, how many errors it has, how much it costs) and measuring against it. Without a baseline, the pilot is judged by enthusiasm, and enthusiasm doesn’t convince a board.

Risks that must be managed

AI as a service carries risks that the company must manage explicitly. The first is confidentiality: information provided to an AI solution must be protected with clear agreements on its use, storage and deletion. The second is vendor dependence: if the solution is built so that only one vendor can operate it, the company loses negotiating power. The third is quality: models can degrade over time if the data changes, and someone must monitor their performance. The fourth is regulatory compliance, in particular personal data protection.

None of these risks is a reason not to use AI. They are reasons to use it with clear governance: an owner, service agreements, periodic review of results and rules on which decisions can be automated and which require human approval.

Training the users

AI changes the way the people who use it work, and that change requires training. In the webinars we give for business owners, one lesson was that the quality of what an AI assistant produces depends heavily on the quality of the questions. That is why we include structured question guides for specific tasks. AI training should not focus on the tool, but on how to frame the problem, how to verify the answer and when not to trust it.