In the first two articles I argued that any company with a large customer base can now put every customer conversation on one pile and let AI agents answer any business question about it. This article is about the part that gets glossed over: what it costs, and why the answer to that question determines whether this becomes a routine management tool or an occasional expensive experiment.
The paradox
An agent-driven analysis of your customer conversations is dramatically cheaper than the consulting projects it replaces. It is also, potentially, wildly unpredictable. Ask the same question and it might cost one euro in tokens, or a hundred, or half a million. Agents loop. They read more. They cross-check, follow leads, re-read. The deeper they go, the better the answer — and the higher the bill.
On its own that would be manageable. The trap is that you cannot know the value of an insight before you have it. Cap the spend and you get shallow answers that may walk right past the thing that mattered. Leave it uncapped and you may pay a great deal to learn that nothing interesting was there. With per-token pricing, every question is a wager against a meter you do not control and cannot forecast.
The old world was slow and sample-based, but at least a research project had a price. A manager who cannot predict what a question costs will, quite rationally, stop asking questions. And then we are back to guessing.
Change the unit of cost
The way out is not a cleverer budget. It is a different unit of cost. Run open-weight models on infrastructure you already own, and the dependency on tokens disappears. What remains is compute time.
That changes the economics completely. If you are willing to wait a couple of hours for a thorough answer, the marginal cost of thoroughness is close to zero. Run the analysis overnight, when the GPUs your engineers were using all day sit idle. Let the agents be as exhaustive as the question deserves, because nothing is metering them. In the morning, the report is there.
Thoroughness stops being a luxury and becomes the default.
The second reason
Cost is not the only argument for owning the infrastructure. The pile we are talking about is every conversation your company has ever had with a customer. There is no more sensitive dataset in a bank, an insurer or a telco. Sending all of it through a shared, third-party platform is a decision most legal and security teams would not take lightly — and under regimes such as DORA, many cannot take at all.
Keeping the data, the models and the agents inside your own perimeter is therefore not a nice-to-have. It is what makes this approach usable in exactly the companies that need it most.
“But aren’t open-weight models weaker?”
This is the question we hear most, and it deserves an honest answer. For the vast majority of business questions, which model you use matters far less than what context you give it. What wins here is not a marginally better model; it is the fact that the agents can read all of your conversations. A frontier model in someone else’s data centre will never know your customers. An open-weight model running on your infrastructure, with every conversation in reach, does. And once your context is under your control, both options remain open to you — while competitors renting a shared platform have neither.
Where we stand
At DataSentics, a Bull company, we are investing heavily in exactly this: a solution built on open-weight AI models that runs in your infrastructure. We can deploy it on your existing Kubernetes clusters, or — as a partner in Europe’s AI factories — provide the infrastructure with it. As a French-owned company with nearly a century of European heritage, we see this as a rare position from which to help European companies make better-informed decisions, listen to their customers faster, and evolve with their needs rather than behind them.
Better strategic decisions, made on complete evidence, at a cost you can predict. The business case for that does not have a ceiling. The only question is whether you own the tool that gives it to you, or rent the meter.