Applied AI
Production is day one
AI models have been good enough for a year to decide better than people on a loan or a health claim. Thomas Piani said so at NVIDIA Startup Day, the day before the keynote of NVIDIA GTC in Berlin. Piani is chief product officer of Taktile, a company whose platform banks and insurers use for such decisions and which tests the models on tasks from that work.
Even so, very few banks and insurers have AI in production for those decisions, in daily use on real cases. What makes that hard, according to Piani, is everything around the model. Tom Becker of the French model maker Mistral said the same at GTC, in a session about ABN AMRO, and cited a figure by which 88 percent of AI projects do not reach production.
This blog follows the route from a pilot, a test on a small scale, to a system in daily use. Most steps on that route start with a decision that only the client can take. We draw on the panel with Piani and on three sessions at GTC, which we went through afterward.
The result is named before the pilot starts
The first decision is which result counts. Gautier Cloix, who sat on the panel with Piani, is CEO of H Company, a Paris company that builds AI agents which operate software the way a person does. His advice to those who make the plans and run governments was to be very demanding of suppliers such as his own company.
H Company is changing its prices so that a client pays only for a result that can be measured in the real world. In Cloix's example, a client that spends 5 percent less on its purchases pays H Company 1 of those 5 percent. If no such saving follows, the client owes H Company nothing, even when the system runs well and everyone uses it.
Many of the things clients ask H Company to build have little real value, Cloix said. They make a great demonstration and save one person ten minutes. H Company says no to those requests.
Jaimie Weber starts at the same point. She is a physician at Tampa General Hospital in Florida, where she leads the work to move AI from idea to production. In a session on enterprise AI, she said that her teams first define which problem they are solving. A tool that solves no problem of the people who care for patients finds no users among them.
Does the process have to be fixed first?
An organization also has to decide how much of a process it redesigns before AI goes into it.
Weber improves the process first. Her team sits down with the people who do the work and asks how they do it and why. In about half of the projects, a change in the workflow turns out to be the whole solution. In the other projects, the technology goes on top once the workflow is efficient and the data is clean.
Cloix starts with the process as it is. A large company or a government runs about a thousand applications, he said, and 80 percent of them have no interface through which other software can use them. His agents work through the screen, as an employee does, so they can start on those applications as they are.
Rewriting all processes first costs a year in his view, and nobody knows whether the process designed today is still the right one by then. He starts with the existing process and improves in short rounds. Every round delivers a new version of the agent. In his experience, it takes about twenty rounds before the way a business works has really changed.
The length of a round decides how long that takes. If every round takes six weeks, twenty rounds take more than two years. With his agents, a round takes an hour or two. The software captures the screen, the voice, the keyboard, and the mouse of an employee who does the task, and a first version of the agent is there when the employee is done. That first version has flaws, and every next round improves it.
Becker described a third route, in which the process is rebuilt around the model. At ABN AMRO, every security alert used to be assessed by a person, who gathered the context, weighed the risk, and decided whether to act. The response always came a little late. In the new process, the information on each alert is gathered and ordered automatically, and a Mistral model makes the first assessment, which a person used to make. Part of the incidents is then handled automatically, and the people stay in control without having to assess every case.
According to Becker, a process like this leaves the pilot stage when it performs better than the old process, scales to the whole bank, and runs in the bank's own environment, under its control.
The three routes differ in how much of the process is redesigned before the first version runs.
The rules live in people's heads
Once the process is chosen, the system needs the rules that the people in it follow.
Taktile has worked for more than a year with Allianz Partners, the assistance and insurance arm of Allianz, on the health claims of travelers. Someone falls ill on holiday in Morocco, sees a doctor there, and submits the invoice online. The insurer has to decide whether the invoice falls within the policy, for every country and in every language.
The first obstacle, Piani said, was that part of the rules existed only in the heads of the reviewers, in teams spread over several offices. They know, for example, that a claim from a certain region calls for extra care with a certain currency. Getting such rules on paper was hard work.
For credit decisions, Piani said, the people who know how to decide well are the ones who write the credit policy and review the cases. They stay involved every time the system is improved, next to the engineers.
Becker said something similar about security. A general model understands cybersecurity in general. The IT environment of one bank, its security policies, and the way it escalates an incident are new to that model and have to come from the bank itself.
Trust is built in steps
A decision on a claim is sensitive and regulated, so the certainty has to be high before a system takes it. At Allianz Partners, that certainty was built up in steps.
The system first ran on claims from the past, whose outcome was known. Then it ran in the shadow of the reviewers for a period, on new claims and without its outcome counting. After that, it received a share of the real claims.
Along the way, the team worked out when the system is too unsure of its outcome, so that the claim goes to a person.
In the session with Weber, Robbie Jerrom of Red Hat described the same approach at banks in the United Kingdom. An agent that helps with fraud detection went to a few users first and was made a little better in every round. Much of the work went into tests, demonstrations, and deliberate attacks on the system, to convince people without a technical background that it can be trusted. This system is not quite in production yet, Jerrom added.
For human approval, the rule there is that a person approves every action that cannot be undone. That matches how a bank already works, because at the end of a process one person signs for the final approval.
Security belongs to these steps as well. Before an agent goes live, the speaker of Oligo Security said on the panel, an organization has to know what the agent is allowed to do, be able to see what it does, and be able to stop it. In our earlier blog What did your agents do last month? we describe how that can be arranged, with one gateway that every request of an agent passes, a log that is kept outside the agent, and a way to stop the agent.
Going live is day one
Full automation was not the goal on the first day at Allianz Partners. What the reviewers decide is used to improve the system, so that the share of claims that pass without manual work grows over time.
Taktile has published the results of this case. In a six-month trial on real claims, that share rose by 30 percent, and the quality of the decisions improved by 4 to 6 percent on a selection of claims. The process is now being extended to the full portfolio of more than 30 million claims a year.
Piani's conclusion was that getting to production is just day one. How a system improves after that day has to be part of the plan from the beginning.
The model itself changes after that day as well. Organizations that went live a year ago now see the model their agent runs on being phased out by the company that makes it, Piani said. That calls for a fixed procedure that checks, at every change of model, whether the outcomes are still accurate and the costs still make sense.
Merck, the German science and technology company, works with EY on humanoid robots for its factories and plans for the same kind of change. The market for these robots moves fast, and the company wants to avoid starting from scratch with every new robot. That is why it invests in skills, such as recognizing an object or inspecting a product, which are kept in a library and are meant to move from one robot to the next. Jan Niklas Mülheims of EY summed it up in the session on this project. Skills stay, and robots can change.
Most of the work is in data and people
Becker, whose company makes the models, estimated that the models are 10 to 50 percent of what it takes to bring a system into production. The rest is the organization's own context, which the model is fitted to, and the infrastructure that runs it securely and reliably.
Immanuel Schweizer leads the work on AI agents at Merck, a company of 60,000 people. In the session with Weber, he passed on a rule of thumb that was told as a joke at an internal meeting. For every euro that goes into technology, two go into data and seven into change management.
Merck runs large programs to teach everyone the basics of AI. The people who review the work of an AI system have to know what is expected of them at that point, Schweizer said, and how their judgment is used to improve the system.
For Weber, the hard part of a project is always something other than the technology. When her team solves a problem that people run into every day, fits the solution into their way of working, and tells them how accurate it is and where it will fail, they start using it.
What the client decides, and when
Every stage in this blog comes with a decision that belongs to the client. The table puts those decisions in order.
| Moment | What the client decides | |---|---| | before the pilot | which result counts, and how it is measured | | at the start | which process comes first, and how much of it is redesigned | | during the build | who owns the rules, and who gets the time to write them down | | before going live | which steps lead to full use, who approves each step, and which actions stay with a person | | from day one | who owns the system, with which budget, and which test is repeated when the model changes | | throughout | how the budget is divided over technology, data, and people |
An AI strategy prepares these decisions before the first pilot starts, each with its moment and its owner. An AI strategy by Quilyx answers four questions.
- where AI earns its place in the organization, and in which order
- what that asks of governance, data, people, and suppliers
- which requirements the GDPR and the EU AI Act set for the applications
- which program takes the organization from the decision to working systems
The route in this blog belongs to the fourth question. Our principle for that route is to build small and run real. A system earns a bigger job by doing a small one well in daily operation first. That is why we recommend a sequence of applications that each prove their value.
Quilyx takes AI ambitions from the boardroom to working results. Those results start on the first day in production, and that is why the route to that day, and the period after it, belong in the strategy.
About this article: This article is our own reading of the panel From Agentic Breakthroughs to Product Adoption at NVIDIA Startup Day in Berlin on 20 October 2026 and of three sessions at NVIDIA GTC Berlin in October 2026, ABN AMRO and Mistral, Own Your AI and Exploring Physical AI. The ideas from those sessions belong to the speakers and are credited to them. The translation to clients, the positions and the choices in this article are ours. We wrote and edited it together with AI, which helped with the research, the working out and the wording. Quilyx carries the editorial responsibility.