Recommend's Trust Layer: Checking the Facts Before AI Acts
Recommend’s Agentic Trust Layer verifies claims before AI systems act, helping prevent bad information from becoming bad action. Built as open-source infrastructure and MIT licensed, it gives teams a transparent way to add verification before execution.
Imagine an AI agent that monitors the market and finds information that one of your main competitors has just cut prices by 20 percent. That could be important enough to change a campaign, adjust an offer, or alert the sales team. If the agent has enough autonomy, it could even trigger some of those actions itself.
The problem starts when that information is wrong or only partly true. Maybe the discount applies to one product rather than the full range. Maybe it is only available in one market. Maybe the source is outdated, or the model has misinterpreted what it found.
With traditional AI use, a mistake like that usually ends up in a piece of text that a person can still review and correct. With AI agents, the situation is different because information can turn directly into action. The more autonomy agents have, the more important it becomes to verify the facts behind their decisions.
The Recommend Agentic Trust Layer was built for exactly that step. It is not another agent. It is a layer that other AI systems can call when they want to verify a claim before acting on it.
Asking AI how confident it is is not enough
One problem with large language models is that their tone is not a reliable measure of whether an answer is correct. A true claim and a false one can sound equally convincing, and simply asking a model, "How sure are you?" does not solve the problem.
The Trust Layer therefore does not ask the model to judge its own knowledge. Instead, it takes a specific claim and checks it against sources available at that moment.
In our competitor example, the claim might be: "Competitor X has cut prices by 20 percent."
The system then looks for evidence that supports or contradicts that claim and returns two separate scores.
The first is a truth score, which measures how strongly the available evidence supports the claim itself.
The second is a confidence score, which tells us how much we should trust that assessment.
The distinction matters because "probably true" is not the same as "reliably verified." A claim may be supported by one source, but if that source is old or unclear, the system will take that into account and lower its confidence in the result.
The output also includes the sources used and a short explanation of how the system reached its conclusion.
What verification looks like in practice
There are several steps behind the final result, but the basic logic is straightforward.
The system first defines as precisely as possible what is actually being checked. If a claim contains several separate facts, it breaks them apart so they can be verified individually.
It then searches for evidence across multiple sources. Not every source is equally useful for every type of question. For a recent business event, current web sources may be most relevant. For a claim about cause and effect, scientific literature is a better authority. The system therefore chooses sources based on the type of claim instead of checking every question in the same way.
It also does not simply accept the first conclusion it gets. The system checks whether the evidence really matches the claim, whether it is current enough, and whether different sources agree with one another.
If the evidence is weak, incomplete, or contradictory, confidence goes down. The final calibration step can make the system less confident, but it cannot artificially make it more confident.
It is also designed not to give a definitive answer when the evidence cannot support one. A claim about what a company will do next year is not the same kind of question as a claim about results the company reported last week. The future can be estimated, but it cannot be verified as fact. In those cases, the Trust Layer limits the confidence of the conclusion rather than presenting a prediction as something proven.
Internal information can be verified too
Not all information used by an AI agent is available on the public web. In a business environment, agents often work with internal documents, policies, metric definitions, or other information that exists only inside the company.
The same principle still applies.
If an internal document says that a certain metric is calculated in a specific way, the Trust Layer can check whether the source cited by that document actually says the same thing. That makes it possible to distinguish between a claim that is properly supported and one that may sound reasonable but is not backed by the documentation.
This makes verification useful even for information that outside sources cannot confirm. The question is not only whether something is generally true, but whether there is appropriate evidence for the specific claim the agent is using as a fact.
Why this matters now
AI models are already very good at finding information, working with text, and drawing conclusions from it. At the same time, AI agents are increasingly able to use those conclusions to trigger real processes.
That shift creates a new requirement. If an agent is going to act without a person manually reviewing every step, the system needs a reliable way to tell the difference between well-supported information and something that merely sounds convincing.
The Recommend Agentic Trust Layer is designed to be that control point. Within Recommend, it sits in front of systems that turn data into commercial actions, but the verification layer itself has been released as open-source infrastructure that other teams building AI agents can also inspect, use, and extend.
As agents take on more responsibility, verifying the information behind their decisions will become an increasingly important part of the stack. The idea behind the Trust Layer can therefore be reduced to one simple rule: before AI acts on a claim, there should be enough evidence to support it.
