Public AI or private AI? Why the best answer is the mix

There is a conversation we have almost every week. A company has run its AI pilot, it worked, everyone is happy… and then the monthly bill arrives. The question stops being «does this work?» and becomes «can I pay for this every month, forever?».

That second question is the one almost nobody asks before starting. And it is the one that decides whether an artificial intelligence project ends up in production or in a drawer. At Alterego we have been turning this over for a while, and the conclusion is unspectacular but highly profitable: the best AI solution for a company is almost never a single AI. It is a mix.

The problem is not intelligence, it is the bill

Frontier models —Claude, GPT and the rest— are extraordinary. They do things that were science fiction two years ago. But they are paid for by usage: while AI is a three-person experiment, you pay little because you use little. The problem appears exactly when the project succeeds. If people like it, they use it. And if they use it, the bill grows.

Add a second factor: some information simply cannot leave the building. Histories, case files, contracts, customer data. It is not only GDPR —though that too—, it is sleeping well at night. Faced with this, the market usually offers two bad answers: everything in the cloud «and you'll pay later», or everything on-prem with a mediocre model. We work on the third.

Line 1: public AI, when you need the maximum

We integrate frontier models into real client processes: document analysis, support, content generation and office automation. When a task needs complex reasoning, nuance, creativity or very long context, a frontier model wins. Trying to replace it with a small model just because it is cheaper is the classic saving that ends up costing dearly.

Our role is not to sell tokens: it is to design the workflow, connect it to the systems the company already has, and measure what is really happening. An AI that nobody understands is an AI nobody can optimise.

Line 2: private AI, from variable to fixed cost

The second line is to set up your own inference infrastructure and run open-weight models. For a large share of business tasks —classify, extract, summarise, answer with context, search internal documentation— they already perform perfectly well.

The economic shift is what matters: with public AI the cost is variable and rises with usage; with private AI the cost is fixed and predictable, and scaling usage lowers the unit cost. Sensitive data never leaves the organisation. When you move from paying per query to an infrastructure fee, the internal conversation changes: it stops being «how do we limit AI?» and becomes «where else can we put it?».

The thesis: do not choose, combine

We do not believe in the either/or. We believe in hybrid architecture: a routing layer decides, for each task, which model should solve it. The bulk of the volume —classifications, extractions, summaries, internal search— goes to the private model. The demanding minority —the hard case, the ambiguous document, the delicate decision— escalates to the public frontier model.

The result keeps intelligence where it is needed and a cost the company can budget for. It is not the cheapest or the most powerful possible: it is the best balance. Which, in the real life of a company, is exactly what you need.

MCP connectors: AI that touches real systems

A brilliant model cut off from your ERP is an expensive intern locked in a windowless office. That is why we build MCP (Model Context Protocol) solutions: connectors so a model —public or private— can query the CRM, read internal docs, write to the database or trigger a process. It is the piece that turns a chat into a work tool.

Building the integration once so it works with any model and client protects the investment: if you switch models in six months, the integration work is not thrown away.

Applied AI, not fairground AI

Our way of working does not change: first understand which company process hurts, then decide which technology fits, and only at the end talk about models. The other way round —starting with technology and hunting for a problem— is how entire budgets burn without anything changing.

If your company is at the «we like it, but the numbers don't work» stage, or «we cannot put this in the cloud», the answer is probably not a better vendor. It is combining them well.