How to stop AI hallucinations

The consequence: retrieval helps enormously and does not finish the job. Stanford's RegLab benchmarked three commercial legal research tools, all of which retrieve from real source law before answering. Lexis+ AI and Ask Practical Law AI still hallucinated on more than 17 percent of benchmark queries, and Westlaw's AI-Assisted Research on more than 34 percent (Stanford HAI). Those products have shipped changes since, and the specific scores will age. The shape of the result will not. Bolting retrieval onto a model reduces invention rather than removing it, and the marketing around those tools ran well ahead of the measurements.

So the question to carry into a vendor call is not whether the system hallucinates. It is whether you find out when it does.

The four mechanisms, and what each one leaves behind

Each mechanism catches what the one before it misses, which is why order matters. Grounding narrows what the model can say. Citations expose what it did say. Abstention handles the questions grounding could not serve. Evals tell you whether any of it still works after last Thursday's change.

The most common half-built configuration in the market is grounding with no abstain path: an accurate-looking system with no way to admit a gap, so it fills one.

Grounding: make it read before it answers

Grounding means the system retrieves from your documents and data first, then answers only from what it found. On one production system we built and still operate, that is more than 1.3 million documents, split into more than 29 million retrievable chunks. A question about a clause in an agreement gets answered from that agreement, not from a general model's impression of what such agreements usually say.

The failure mode moves rather than disappearing. Retrieve the wrong document and the model answers faithfully from it, sounding exactly as certain. That is why most of the engineering goes into the retrieval layer, not the model. Whether to retrieve at all or train a model on your data is a real fork with a real answer, covered in RAG versus fine-tuning; most business questions want retrieval. RAG is short for retrieval-augmented generation. It means the model looks the answer up before it writes. Building that layer is what AI agents and grounded RAG is.

Citations: make every claim point at a source

A citation earns its keep only if a human can follow it and a machine already has. On the same system, answers pass a seven-stage verifier that checks each claim against the source it cites before anything is served. A claim that cannot be traced fails closed instead of shipping with a hedge.

The residual failure is the citation that is real but does not support the sentence attached to it. That is the most dangerous error in the category, because it survives a casual spot check. Ask any vendor whether verification confirms the cited passage contains the specific value, or merely that a citation exists.

Abstention: give it a way to say it does not know

An honest "I do not have that" beats a confident wrong answer in every workflow where somebody acts on the output. On the same production system, abstain precision holds at 1.000: every time it declined to answer, declining was correct. Zero cross-tenant leaks across both embedding spaces, so one client's data never appeared in another client's answers.

Read the limit of that number, because a vendor quoting it will not. Abstain precision measures whether refusals were justified. It does not measure how often the system should have refused and answered anyway, and it does not measure how often it refuses something it could have handled. Always ask for the refusal rate next to it. A system that abstains on 60 percent of questions can post perfect abstain precision and still be worthless.

Evals: turn "it seemed fine" into a number

An eval suite is a test suite for answers: graded questions with known-correct outputs, run on every change, with thresholds that block a release. The gate on that production system holds zero golden wrong-values and zero determinism violations. The known-answer set stays right, and the same question keeps returning the same answer. A change that drops a threshold gets reverted, not argued about.

Method is its own subject, and how to evaluate a RAG system covers building the golden set and reading the scores. At buying time the test is blunter. No suite and no threshold means nobody knows whether last week's change made the system worse, the vendor included. That work is LLM evals and reliability, and it is the layer teams skip first and regret longest.

If the budget stretches to two of the four, take abstention and evals. Grounding and citations without them produce a system that is usually right and never tells you which times. Abstention and evals produce a system that is narrower, occasionally frustrating and honest about its own edges. That is the version you can put in front of a customer while you widen it. That ordering also survives a legal or compliance review. The confident-and-unmeasured version does not.

What to demand from any vendor

Take these into the call. Count how many get answered with a number and how many get answered with a paragraph.

1. Show me one answer, then show me the source it came from. Live, on a question you bring, not a recorded demo. If following the citation takes more than one click, your team will not do it.
2. What happens when the answer is not in my data? Watch the system decline. A vendor who cannot produce a refusal on request has probably not built an abstain path.
3. What is your abstain precision, and what is your refusal rate? Both, or neither is meaningful. Precision without a refusal rate hides a system that says no to everything.
4. Does verification check the value, or only that a citation exists? This separates a real verifier from a link decorator.
5. Show me the eval suite and the thresholds that block a release. Ask what the last reverted change was. A team with a working gate has a story ready.
6. Who writes the golden answers? Your domain experts, not the vendor's engineers, or the test measures the wrong thing.
7. What is retrieval recall on our questions, measured before we sign? A short retrieval test on your real corpus is cheap and predicts most of the outcome.
8. When it is wrong in production, who finds out and how fast? Logging, monitoring, a named owner. "The user will tell us" is not an answer.

If most come back as prose rather than numbers, you are looking at a demo with a roadmap attached. The same shape of question, aimed wider than hallucination, is in questions to ask an AI development company.

Where this lands in your budget

These four mechanisms are most of the cost of a production AI system and most of the reason pilots stall. A demo needs a model and a prompt. Production needs retrieval tuned against your corpus, verification, an abstain path and an eval harness with somebody's name on it. Teams price the first and discover the second, which is the pattern behind why AI pilots never reach production.

Budget for the real thing. What an AI agent costs to build breaks down where the money goes, and build versus buy for AI agents covers when an off-the-shelf platform is enough and when the honesty layer has to be yours.

The bar is not a system that never errs. It is a system where every claim has a source. The source is checked before you see it. The system says no when it should. And a number tells you how often all of that holds. Ask for the number. Ask for the failure mode. Ask what gets reverted.