GUIDE · EVIDENCE

Evidence-centric AI

Evidence-centric AI, also called evidence-grounded or grounded AI, is AI whose outputs are derived from retrievable evidence and stay traceable to their sources, so the work is defensible rather than merely plausible. The distinction it draws is the one that matters to a business: not whether an answer sounds right, but whether you can show what it rests on. It is an approach to using AI, not a single feature, and its purpose is to make AI output something an organisation can stand behind.

IN SHORT

Evidence-centric AI ties every business claim back to source material that a person can retrieve and check.
It reframes the hallucination problem correctly: the real risk is confident output with no provenance, not just the occasional factual slip.
Grounding is how it works. The model draws only on permitted evidence, and provenance keeps each output connected to that evidence.
The test of it is a single question: “where did this come from?” If you can answer without a scramble, the output is defensible.
It is a broader idea than retrieval-augmented generation. Retrieval finds material; evidence-centric AI also carries provenance, permission-awareness and a record.

What evidence-centric AI means

A general-purpose model is trained to produce the most likely next words. It is extraordinarily good at sounding right. What it is not designed to do is tell you where an answer came from, because in most cases there is no single source: the output is a blend of everything the model absorbed in training. For casual use that is fine. For work an organisation has to defend, it is the whole problem.

Evidence-centric AI inverts the priority. Instead of asking the model to recall, you give it a bounded set of evidence drawn from your own record, and you ask it to reason over that. The output is then not a memory but a derivation: each business claim is produced from specific source material, and it stays linked to that material after the fact. The measure of a good output stops being “does this read well” and becomes “can we show what this rests on.”

That shift changes what AI is for. It stops being an oracle you consult and hope is right. It becomes a way of turning evidence you already hold into work you can trace, check and put your name to.

Why plausible is not the same as defensible

The word “hallucination” has framed the AI reliability conversation, and it has framed it slightly wrongly. It suggests the problem is that models sometimes get facts wrong, and that if they got facts wrong less often the problem would shrink to nothing. That misses the real exposure.

The risk in business is not the occasional wrong number. It is unprovenanced confident output: an answer delivered with complete fluency and no way to tell what it is based on. A wrong figure you can catch is a nuisance. A right-sounding figure that nobody can trace is a liability, because you cannot tell the good ones from the bad ones, and you have signed your name to all of them. Even when the model is correct, an output you cannot ground is a claim you cannot defend.

Plausibility and defensibility are different properties. A plausible output is one that reads as though it could be true. A defensible output is one you can show to be true, or at least show the basis of. General AI optimises for the first. Evidence-centric AI is built for the second. When a client, an auditor, a regulator or a court asks what a sentence rests on, plausibility offers nothing and defensibility offers an answer.

Grounding and provenance: how it works

Two mechanisms carry evidence-centric AI, and it is worth separating them because they do different jobs.

Grounding is the input discipline. Before the model reasons, the relevant evidence is retrieved from the organisation’s own record and given to it as the material to work from. The model is asked to derive its output from that evidence rather than from its training. Grounding is what stops the output floating free of anything real.
Provenance is the output discipline. As the work is produced, the link between each claim and the evidence it came from is preserved, not discarded. Provenance is what lets you walk backwards later, from a sentence in a finished document to the exact source it rests on.

Grounding without provenance is thin: the model may have used good evidence, but you have no way to prove it afterwards. Provenance without grounding is impossible: you cannot preserve a link to a source the output was never actually derived from. Evidence-centric AI needs both, working together, so that the output is not only produced from evidence but stays connected to it for as long as the work lives.

There is a third condition that is easy to overlook. The evidence itself has to be permission-aware. It is not enough to ground an output in source material; the material has to be material the people involved are entitled to use. Grounding an answer in a document someone was never allowed to see does not make the answer defensible. It makes it a breach with a citation. In Square One, retrieval is permission-aware by construction, inherited from the Microsoft 365 foundation the organisation already runs, so grounding never quietly reaches past someone’s entitlements.

The “where did this come from?” test

You do not need a technical audit to tell whether AI output is evidence-centric. You need one question, asked of a specific piece of work: where did this come from?

Ask it of a paragraph in a report, a figure in a submission, a claim in a client-facing document. If the honest answer is “the model produced it” or “someone pasted it out of a chatbot,” the output is not defensible, however good it reads. If the answer is “here is the source, and here is the person who was entitled to use it, and here is the record of it being approved,” the output is evidence-centric, whether or not anyone used that phrase while producing it.

The value of the test is that it is the same question everyone downstream will eventually ask. A client querying an invoice, a regulator reviewing a close-out, a lawyer preparing a defence, a new manager inheriting a file: all of them, sooner or later, ask where a claim came from. Evidence-centric AI is simply the discipline of being able to answer them in minutes rather than weeks.

Evidence-centric AI vs RAG and other buzzwords

The field is thick with terms, and several of them circle the same idea from different angles. It helps to be precise about how they relate.

Retrieval-augmented generation (RAG) is a technique: fetch relevant material and feed it to the model before it answers. It is one way to achieve grounding, and a good one. But RAG on its own says nothing about whether provenance is preserved, whether retrieval respects permissions, or whether the output is approved and recorded. RAG can make an answer better informed without making it defensible.
Grounded AI and evidence-based AI are, for most purposes, other names for the same thing described here. The emphasis is on the source of truth being real, retrievable evidence rather than model memory.
Traceable AI and AI provenance name the output side specifically: the ability to follow a result back to its origin. They are necessary to evidence-centric AI but not sufficient on their own, because you can trace an output back to a source that nobody was entitled to use.
Explainable AI is a different concern again. It asks why a model reached a conclusion, in terms of the model’s own workings. Evidence-centric AI is less interested in the model’s reasoning and more interested in the evidence, which is what a business actually has to defend.

The through-line: most of these terms describe a part. Evidence-centric AI is the standard that assembles the parts into something an organisation can rely on. Retrieval finds the material, grounding uses it, provenance preserves the link, permissions keep it lawful, and approval and record make it accountable. Miss any one and you have a component, not a defensible output.

What it looks like in practice

Consider an environmental management plan drafted for a rehabilitation close-out. It contains the sentence “groundwater monitoring returned to baseline within 14 days of completion.” In the ungoverned version, that sentence is assembled from a similar past job and a half-remembered figure, and it goes out under the company letterhead with no way to say what it rests on. It reads perfectly. It is plausible. It is also indefensible, because if the number is ever questioned, there is no answer.

In the evidence-centric version, the same sentence is derived from the organisation’s own record and stays linked to the two sources it rests on: a dated site-diary entry and an accredited laboratory report. Access to those sources is checked against who is entitled to see them. The document passes through a human approval, and that approval is logged. A year later, when someone asks where the 14-day figure came from, the answer takes minutes: here is the diary entry, here is the lab result, here is who approved the plan and when. You can see this exact behaviour on Square One’s Documents surface, where evidence-bound drafting keeps every sentence traceable to its source.

Nothing about that is exotic. The AI still did the drafting, and did it fast. What changed is that the output arrived already carrying its own defence. That is the whole of evidence-centric AI, and it is why the approach sits at the heart of governed work and its AI governance discipline. As intelligence itself becomes cheap, the ability to trust and trace what it produces is the part that stays scarce, a gap we call the confidence premium.

Frequently asked questions

What is evidence-centric AI in plain terms?

It is AI whose outputs are built from retrievable evidence and stay linked to that evidence, so you can always show what an answer rests on. The goal is work that is defensible, not just work that sounds right.

How does evidence-centric AI stop hallucination?

It reframes the problem. Rather than hoping the model gets facts right, it grounds each output in specific source material and preserves the link to it. The exposure that matters is confident output with no provenance, and grounding plus provenance removes it. Even a correct answer you cannot trace is treated as a claim you cannot defend.

Is evidence-centric AI the same as RAG?

No. Retrieval-augmented generation is one technique for grounding: it fetches relevant material before the model answers. Evidence-centric AI is the broader standard. It also requires that provenance is preserved, that retrieval respects permissions, and that the output is approved and recorded.

What is the difference between plausible and defensible output?

A plausible output reads as though it could be true. A defensible output you can show to be true, or show the basis of. General AI optimises for plausibility. Evidence-centric AI is built for defensibility.

What does provenance mean here?

Provenance is the preserved link between a claim and the source it came from. It is what lets you walk backwards later, from a sentence in a finished document to the exact evidence it rests on, months or years after the work was done.

Does grounding an output in evidence make it safe by itself?

Not on its own. The evidence has to be material the people involved are entitled to use. Grounding an answer in a source someone was never allowed to see does not make it defensible. That is why permission-aware retrieval is part of the standard, not an optional extra.

How do I tell if our AI output is evidence-centric?

Ask one question of a specific piece of work: where did this come from? If the answer is the source, the entitled user, and the record of approval, it is evidence-centric. If the answer is that the model produced it, it is not, however well it reads.

Does evidence-centric AI depend on a particular model?

No. It is a discipline about evidence, grounding and provenance, not about which model does the reasoning. Square One is model-agnostic: it uses the right model for each task and keeps the grounding and governance constant as models change.