An agent that improvises can't sign off an investment: the case for determinism

What it costs to hand a repetitive task to an agent

David Teruel, founder of GreenPowerFinder, put a question to the audience at RENMAD UsefulAI 2026 on 2 June. Imagine a task that has to be repeated tens, hundreds or thousands of times. Who should do it? The answer he usually gets, he said, is "I've got it, an artificial intelligence agent". That was what he was there for: "I've come to take that myth apart a bit".

He headed off the easy reading himself, that this was an anti-AI speech. His target was the panel's premise, autonomous agents, and his reason was money. "I think that delegating that kind of task, that kind of work to artificial intelligence agents isn't always the best decision", he said.

"I don't think artificial intelligences should make decisions, they should give us the information so that we can improve our decision-making", he added. He had already summed up why in a single sentence: "a badly made decision costs money".

Why the same question gets two different answers

He backed the technical argument with an example of his own, and a hard one. Give a generative model a scanned, twenty-five-year-old municipal zoning plan and it handles it beautifully, drawing conclusions and proposing alternatives. Start feeding it a large body of varied documents that contradict one another, with legal hierarchies and targeted amendments layered on top, and up come "the famous overlaps, the lost in the middle, the overflow windows, the context windows from cognitive overload". These are documents that overlap, and facts the model loses track of when they sit in the middle of a long text, inside a context window, which is everything it can hold in view at once.

Teruel named the consequence: "you get a non-deterministic system […] when we're going to make a decision it's going to give us fluctuating information, it won't always give us the same information or always draw the same conclusions for us".

Outside the sector, this has been documented and measured. In September 2025 Thinking Machines Lab ran a simple experiment: it sent a thousand identical requests to the same model with the temperature set to zero, which amounts to explicitly asking it not to improvise, and got back eighty different responses. The answers began to diverge at token 103; a token is the smallest unit of text a model writes with. The cause lies in the GPU kernels, the computation routines the chip runs. Their numerical result depends on the size of the batch your request is grouped into, and that size changes with how many people are using the service that second. When the kernels were rewritten so that the result no longer depended on the batch, all thousand responses came out identical, and the computation took roughly twice as long.

The fluctuation Teruel described lies in the system architecture. Switching models or turning down the temperature leaves it where it is, and genuinely removing it has to be paid for in compute.

Since 2024 there has been a name for measuring what that fluctuation costs you: `pass^k`, the metric the AI agent company Sierra introduced with τ-bench, its benchmark suite. `pass^k` asks whether the agent gets it right k times in a row, for example eight attempts and eight successes. On retail customer-service tasks, the reference agent in that study succeeded more than 60% of the time on the first attempt and fell below 25% when it had to get all eight right. That agent dates from 2024, and its score does not reflect today's state of the art. The metric still holds, because an investment process is a chain of right answers in a row.

What is at stake for a developer when the score moves

Álvaro Sotorrío, CEO and founder of ORKA Energía, described what has to be paid out before you have anything: "in project development you have to put a lot of money up front, you have to post guarantees, you have to get reports done, you have to pay for options". Each of those costs follows from an earlier decision about which opportunity is worth the outlay, and you are the one fronting the money.

That decision is where Teruel's product comes in: a pre-feasibility report that, in one click, cross-checks zoning, the grid connection point and available capacity, plus odour dispersion for biomethane, and sums it all up in a score. Hence the condition he himself sets for it: "it's very important that that information is predictable, that it's deterministic, that it doesn't fluctuate, that it isn't going to give me a different score every time I click".

A report that gives one score on Monday and another on Tuesday sends the developer off to post guarantees against a number that moves.

Break the problem down until it is deterministic

Teruel's way out starts with the size of the job: "if we have a very big problem, we're going to split it into very small problems". Many of those can be solved "with basic traditional algorithms", and "what we're going to delegate are always very specific, very concrete tasks that really will be deterministic". The systems he is after are repeatable, auditable, traceable "and above all deterministic".

That led to his answer to the question he is always asked. "Lots of times, which model do you use? It really doesn't matter." He prefaced it with a condition, "if the solutions are well designed", and stopped short of making it a rule: "the solutions tend to be model-agnostic".

Xavi Riba, founding partner at Temos.ai, sized up the gap between a proof of concept and a product that holds up, and he did so as someone who sells agents: "you really can knock together a little pilot in an afternoon, a POC where you say, hey, it more or less works. Now, building a reliable agent in production that works is extremely difficult. I mean, we're at the two extremes." His bar for "reliable" has two conditions: "that it doesn't hallucinate, that it doesn't give you a sub-optimal result two times out of three".

The two stumbles the vendors themselves described

Both came out when the moderator, Alicia Carrasco, asked them for a success story, or one that had taken longer than expected. Sotorrío switched on his platform's agents module convinced of the effect it would have: "I said, given what it's capable of, people are going to be blown away. Well no, people were using it to get at that same information" they were already getting from the filters. He fixed it by explaining to clients what an agent can be asked in natural language, and in his view they are now "getting the most out of those agents". His conclusion: "to begin with I think you have to educate people to be able to take advantage of that paradigm shift".

Riba automated replies to submissions lodged during the statutory public comment period and hit the same wall: "it's very hard to change the way people do things […] you have to trust that the response that gets produced is good". His lesson: handing over the tool is not the whole job, "rather it's a bit about co-creating, adapting, personalising, customising".

Both stumbles were about adoption, and both match what is being measured outside the room. Deloitte's Tech Trends 2026 finds that 38% of the organisations surveyed are piloting agents and only 11% have them in production. That is the gap between piloting and production, and the two cases go some way to explaining it.

Determinism as a selling point

The panel did not agree. Three of the four had come to sell what the first speaker had called into question, and the disagreement was left standing in full view of the audience. Sotorrío's condition does not settle it, but it does mark out the ground they share. In his words: "I think it's also very important that the AI isn't a black box", and "you always have to be able to verify where that data comes from", because the alternative is "I ask the AI, it gives me an answer and blind faith in what it gives you". It was the same thing Teruel had asked for, this time from the person selling the agent.

That shifts where the value sits. Reading regulations and drawing conclusions from a PDF, Teruel said, "I see it as a trivial problem and whoever doesn't do it is late and whoever is doing it is no longer adding value". The model is heading towards commodity status: swap it for another and it makes no difference. What still sets you apart is proprietary data, how the steps are orchestrated and the guarantee that the same click gives the same number.

A generative model improvises and surprises you every time, and that is precisely what rules it out of signing off an investment. Once you accept that, determinism is the feature you can sell. Break the problem down, delegate only what is deterministic and keep every number next to its source.

That is what lets an investment committee or an auditor accept a number that came out of a machine. It is what makes artificial intelligence bankable.

Panel at RENMAD UsefulAI 2026, Madrid, 2 June 2026. David Teruel, founder of GreenPowerFinder; Xavi Riba, founding partner at Temos.ai; Álvaro Sotorrío, CEO and founder of ORKA Energía; Christian Crespo de León, CEO of icloudCompliance. Moderated by Alicia Carrasco, CEO of OlivoEnergy.