An AI agent can answer in clean, confident paragraphs and still quote a superseded price, the wrong contract clause, or a supplier that left the approved list last year. Stanford researchers tested legal research tools sold as hallucination free because they retrieve from a curated database. The tools still gave incorrect or misgrounded answers on 17 to 33 percent of queries (Stanford RegLab, 2024). The fluency comes from the language model. The facts come from whatever the system gives the model, and the way that material is organised decides how often the agent is right.
THE PROBLEM
Why Retrieval Alone Still Produces Wrong Answers
Standard retrieval augmented generation (RAG) finds the passages whose wording looks most like the question and gives those to the model. Similar wording is a weak stand-in for the correct fact, so an agent can cite a real document from your own library and still be wrong.
A typical RAG pipeline splits each document into chunks of a few hundred words and turns each chunk into a numeric vector. When a question arrives, it pulls back the five or ten chunks closest to it. Say a procurement manager asks which sites still buy hydraulic seals from a particular supplier, and at what price. A correct answer needs four sources: the site register, the supplier list, the 2024 framework contract and the March amendment that changed the price. The retriever may return only the original contract, because its wording matches the question best, and the model will then report the old price with full confidence. Nothing in the text it retrieved says that an amendment exists.
WHERE IT BREAKS
Three Ways A Retrieval Agent Gets A Fact Wrong
Most wrong answers from a RAG agent come from one of three retrieval failures. The model is usually doing what it was asked with the text it was given.
The Fact Sits In A Chunk That Was Never Retrieved
A price table on page 40 of a tender rarely uses the same words as a plain English question. When the relevant chunk scores below the cut-off, the model fills the gap from its general training, and that is where invented figures appear.
The Answer Needs Two Documents Joined Together
A question such as “which of our customers are affected by this recall” needs a batch record linked to a delivery record. Vector search returns passages one at a time and cannot tell that two of them describe the same batch, so the model has to guess the join.
Two Versions Of The Truth Look Equally Relevant
Draft and final policies, old and new rate cards, and amended contracts all match the same question. Unless something records which document supersedes which, the agent uses whichever chunk scored highest.
THE FIX
What A Knowledge Graph Adds To An AI Agent
A knowledge graph stores the things your business deals with, such as suppliers, sites, contracts, products and people, along with the named relationships between them. An agent can then follow links like “amendment replaces clause 7” or “site uses supplier” as recorded facts, so it works from structure instead of guessing from wording.
In practice the graph sits beside the document index. Retrieval first finds the relevant entities, then follows their relationships to collect connected facts, and only then pulls the supporting passages. The model receives the current contract, its amendment and the sites it covers as one linked set. Microsoft Research’s GraphRAG study tested this approach on questions about whole collections of around one million tokens. It beat plain vector retrieval on comprehensiveness in 72 to 83 percent of comparisons (Edge et al., 2024).
Read those results carefully. A language model did the judging, the questions asked for summaries, and the score measures how complete an answer was more than how precise its facts were. A graph also has its own costs. A model usually extracts its entities and relationships from documents, and that extraction makes mistakes. Every new contract or policy has to be added to the graph before the agent can use it. A graph is worth building when questions often need facts from several documents and when documents change versions. For a small, stable policy library, well-tuned plain retrieval is often enough.
THE SOURCE OF TRUTH
Some Facts Should Be Read Live From The System Of Record
Figures that change daily, such as stock on hand, ledger balances or ticket status, are most reliable when the agent reads them from the live system at the moment of the question. A copy of them in a document or a graph is a second version, and it starts going out of date as soon as it is made.
Here, connections to your systems matter more than retrieval. With Model Context Protocol (MCP) connectors, an agent can query the ERP, CRM, ticketing or document system directly and quote the value it gets back. A sound design uses each source for what it does well:
| Kind Of Question | Most Reliable Source | Typical Failure With Vector Search Alone | Upkeep |
|---|---|---|---|
| “What does our leave policy say about carry over?” | Document index (RAG) | Rare, if the policy text is current | Re-index on each policy change |
| “Which sites are covered by the amended seal contract?” | Knowledge graph plus documents | Old contract returned, amendment missed | Graph updated with each new contract |
| “How many units of part 4471 are in the Leeds warehouse?” | Live ERP query via MCP | Out-of-date figure quoted from a report | None, read at question time |
| “Draft a purchase order for the shortfall” | ERP write, staged for approval | Wrong quantity committed | A person approves every write |
The last row is the safety net for all the others. An agent that is wrong 2 percent of the time is acceptable for drafting and unacceptable for committing transactions, so every write should wait for a person to approve it.
PROVE IT
How To Test An AI Agent Before You Trust It
Measure accuracy on your own questions before deployment, using a fixed test set and a score for each answer. A demo on hand-picked questions tells you very little about the error rate your staff will see.
A workable test set is 100 to 200 real questions collected from the people who will use the agent. Each question needs its correct answer and the document or record that proves it. Include questions with a false premise, such as one that asks about a contract that does not exist. A well-grounded agent should say it cannot find one, and the Stanford team used exactly this kind of question to expose fabricated answers. Score two things separately: whether the answer is correct, and whether the cited source actually supports it. Run the set again after every change to the model, the documents or the graph.
Regulators expect this kind of evidence where people are affected. The UK Information Commissioner’s Office has said that the accuracy expected of generative AI output depends on how the output is used, and that the bar is higher when it informs decisions about individuals (ICO consultation on accuracy). A scored test set is the simplest way to show which bar you set and whether the agent meets it.
HOW ARSA WORKS
How ARSA Builds Grounded AI Agents
ARSA builds private RAG systems, agentic workflows and MCP integrations that run on infrastructure the customer controls. Every engagement starts with a two-week feasibility assessment run against your actual data. That assessment, with your documents in hand, is the place to decide whether your questions justify a knowledge graph and what it would take to keep one current.
Document ingestion, embedding and the vector store all sit on your infrastructure, and no document reaches an external API at any point. Models are open-weight, under permissive licences, and chosen for each workload. If your acronyms, document taxonomy or schema formats confuse a general model, ARSA can fine-tune it on your hardware, and the training data stays with you. MCP connectors reach ERP, CRM, ticketing and document systems, including systems ARSA did not supply. Every write an agent prepares waits for a person to commit it. That approval gate is built into everything ARSA delivers and cannot be switched off. More detail is on the LLM and AI agent development page.
There are limits you should know about. An open-weight model small enough to run on premises is less capable than the largest hosted models on the hardest reasoning tasks. If your data is allowed to leave the building and raw capability matters most, a cloud model may be the better choice, and ARSA will tell you so in the assessment.
| Stage | From | Duration |
|---|---|---|
| Feasibility assessment | $4,500 | 2 weeks |
| Pilot deployment | $20,000 | 8 weeks |
| Production programme | $60,000 to $300,000+ | 12 weeks+ |
If you contract within 90 days, the assessment fee is deducted from the project fee. ARSA works with clients in the United States, the United Kingdom, the UAE, Singapore and Australia. List prices are in USD and invoicing is available in USD or EUR. Where visa conditions prevent ARSA engineers from travelling, a qualified local partner installs under ARSA supervision. Every engagement stage is listed on the services page.
FAQ
Frequently Asked Questions
Does RAG Stop An AI Agent From Hallucinating?
It reduces hallucination and does not remove it. Retrieval gives the model relevant text, but if the wrong passage comes back, or a newer version is missed, the model will answer confidently from what it was given. Independent tests of commercial RAG tools have measured error rates of 17 to 33 percent on hard questions.
When Is A Knowledge Graph Worth The Extra Work?
A graph pays off when answers often depend on facts spread across several documents, or when documents replace each other, as contracts, amendments, rate cards and policies do. For a small library of stable documents, a well-tuned retrieval pipeline with good chunking is usually enough and cheaper to maintain.
Can An AI Agent Read Live Figures From Our ERP Instead Of Documents?
Yes, through MCP connectors that let the agent query the ERP, CRM or ticketing system at the moment of the question. For figures that change daily, this is more reliable than any copy kept in documents or a graph. Any write the agent proposes should still wait for a person to approve it.
How Do We Measure Whether An Agent Gets Facts Right?
Build a test set of 100 to 200 real questions with known answers and sources, and include some questions with false premises. Score whether each answer is correct and whether its cited source supports it, as two separate scores. Run the set again after every change to the model or the data, so a drop in accuracy shows up before your users find it.
How Should We Choose An AI Agent Development Company?
Ask three questions: how they will measure accuracy on your data, where your documents and model will run, and what stops an agent writing to a live system without approval. A partner who answers those in writing before building anything is safer than one who starts with a demo.
Does ARSA Deliver AI Agent Projects Outside Indonesia?
Yes. International list prices are in USD and invoicing is available in USD or EUR. Systems run on the customer’s own infrastructure or on a single-tenant instance in a region the customer chooses.
NEXT STEP
Find Out What Your Agent Gets Wrong Before Your Customers Do
If an assistant already answers questions from your documents, or you are about to build one, the useful first step is to measure where its answers fail and what fix each failure needs. ARSA’s feasibility assessment does that in two weeks and includes a cost model and a go or no-go recommendation. Start with LLM And AI Agent Development or contact ARSA to scope it.