Why RAG and most agentic AI tools fall short on accuracy – and how the Graph Research Labs suite of tools helps organisations close the gap and keep pace with changing regulations.
The first three parts of this series covered the Five Agent Failure Patterns, the knowledge graph, ontology and agent harness that fix them, and the methodology for building the ontology and governing agents in production. This final part describes our own tools at Graph Research Labs, designed to help organisations use graph databases and knowledge graphs to run AI agents that are accurate, governed and ready for production.
Many organisations are asking the same question. With agent frameworks, vector databases, graph databases and guardrail products already available, what is still missing? The answer is the hard part, and it is what architects deal with every day. New technology has to integrate with existing systems, quickly and affordably. Regulations have to be met, and that has to be auditable. And existing staff have to be able to use the new tools, so that everything works together.
RAG improves on a bare model by giving the agent documents and data to draw on. Its limit is that it retrieves by similarity. It finds text that looks like the question, and leaves the agent to decide what that text means and how the pieces relate to one another. GraphRAG narrows the gap by retrieving from a graph, but that graph is usually extracted by an LLM, with no agreed ontology behind it.
Better results come from changing what the agent retrieves from. With an expert-designed ontology and a knowledge graph that conforms to it, the agent follows defined relationships to an answer rather than matching on likeness, and each answer carries the facts it was built from. Retrieval still has a place, in finding the right document for example, but accuracy no longer depends on it.
Each category of agentic AI tool on the market solves a real problem. Most stop at the same point. They help an agent run, retrieve or be checked, but they do not give it an agreed meaning to work from, and they do not enforce an organisation's rules by structure.
| Type of tool | What it does | Where it stops |
|---|---|---|
| Agent frameworks | Run the agent: plan, call tools, manage context | Know the tools exist, not what the business means |
| Retrieval and vector search | Find text that looks similar to the question | No agreed meaning, no connections between facts |
| GraphRAG | Retrieves from a graph an LLM has extracted from documents | No agreed ontology behind the graph, and no rules on what an agent may do |
| Graph databases | Store connected data | Do not supply the ontology, the method to build it or the rules |
| Agent guardrails and AI reviewers | Use a second model to watch the first | Fail probabilistically, like the agent they are watching |
| AI agent evaluation | Measures how well an agent behaves, by reading its logs | Does not measure whether the agent produced the right result |
Without the right tooling, organisations end up stitching these together and filling the gaps by hand. The gaps are where the failure patterns live. To reach the accuracy that mission-critical systems need, those gaps have to be closed, and an organisation can do that with its own team using GRL's suite of tools.
The GRL tools support an organisation's own team at each stage of the work.
GRL Ontology Manager lets a team create, govern and publish its ontologies, with collaboration and a managed lifecycle. GRL OntoLinter checks an ontology for structural and semantic defects before they reach production. A library of ontology accelerators – covering areas such as banking, health and regulatory compliance – means most organisations adapt a proven model rather than start from a blank page. These support the AI-Assisted Ontology Modelling Methodology described in Part 3.
GRL Graph Data Pipeline Service ingests, transforms and loads data from an organisation's IT systems into the graph, mapped to the ontology. GRL Document Integrator does the same for unstructured documents, including Word documents, PDFs, spreadsheets, text files and Markdown files, with scans and images coming soon. It extracts what conforms to the ontology and flags what does not. GRL Graph Insights Explorer lets people query and visualise the result, so the graph can be inspected by the business as well as by agents.
GRL Generators produce working REST APIs, MCP servers and applications directly from the ontology in minutes. That allows an ontology to be production tested with a full data flow in hours, and the cycle repeated until it is ready for release – before any agent depends on it. The outputs are in forms most developers already work with, so an organisation can use its own team and does not need to hire specialists or have its developers learn a graph query language such as SPARQL.
GRL Agent Harness sits around the agent. The agent proposes an action, the harness checks it against the rules in the ontology, and a breach is blocked before it executes, with the rule cited. It also governs what agents read, write and remember through a shared memory. GRL Governance Metagraph records provenance, lineage and audit trails. It captures the outcome of an agent's work as well as every action in between, so every decision can be traced end to end.
For organisations drowning in changing rules and regulations, this is where the GRL tools offer their greatest advantage.
In most organisations, a change to a regulation, a product or a structure has to be found and reworked in every application, every prompt and, now with AI, every agent that depends on it. In a large organisation such as a bank or a government department, the applications alone run into the thousands, and each one has to be changed. That takes months or even years, and by then the rules and regulations have often changed again.
It is a cycle that never ends. I spent decades inside it as an architect and, frustrated by systems that were so slow and expensive to build and keep updated, started GRL to break the cycle.
We did this by making the GRL tools declarative. They read the ontology rather than hard-coding the business into each application or agent. When something changes, the ontology is updated once. GRL Generators rebuild the interfaces that depend on it, and the knowledge graph, checks, and every agent follows.
GRL Generators – Define once. Generate forever.
A change that used to be a major project over months becomes an update over minutes. If an update needs to be reversed, that can be done just as easily. This enables an organisation to respond to a new regulation quickly, while showing a regulator exactly what changed and when, and keeping every agent working to the current rules.
The differences come down to a handful of practical points.
| Area | Most tools today | GRL tools |
|---|---|---|
| Guardrails | Rules written into a prompt, or a second AI checking the first | Built into the architecture. A breach is blocked before it runs, with the rule cited |
| Updates and regulations | Every application, prompt and agent changed by hand | The ontology is updated once, and GRL Generators rebuild what depends on it in minutes |
| Testing | Only possible once everything has been built | Production tested with a full data flow in hours |
| Provenance | Logs of what an agent did | The outcome and every action in between, in granular detail |
| Unstructured data | Documents searched for similar text | Word documents, PDFs and spreadsheets brought into the knowledge graph |
| Developer skills | Graph specialists and SPARQL | Standard APIs, MCP servers and applications an organisation's own developers can use |
| Platform | Often tied to one vendor's database or cloud | Any graph database, any ontology, any environment |
| Connectivity | Often require a connection to a vendor's cloud | Run in an organisation's own cloud, on its own premises or fully air-gapped, with no outside connection needed |
| Existing systems | Replaced or worked around | Sit alongside the model, agent framework and IT systems already in place |
Models and agent frameworks will keep changing, often several times a year. The meaning of an organisation should not have to change with them. Defined once and governed properly, it becomes the foundation every agent works from, and the basis for taking agent accuracy beyond 90%.
That is the argument of this series: why agents fail, what fixes them, how to build it, and the tools that let an organisation do it with its own team.
None of this is out of reach. Most organisations already have what matters most: their people, their data and a deep knowledge of their own business. What has been missing is a way to put that knowledge where AI agents can use it. Graph Research Labs has the tools to knit it all together now: more integrated data, more accurate agents, stronger governance and full traceability.
Learn more by booking a 30-minute conversation at graphresearchlabs.com/book-a-call.
Guardrails and AI reviewers use a model to supervise a model, so they fail probabilistically. The GRL Agent Harness checks each proposed action against rules defined in an ontology and blocks a breach before it executes, with the rule cited and the decision recorded.
GraphRAG improves retrieval by drawing on a graph, but that graph is usually extracted by an LLM without an agreed ontology behind it. The GRL tools start from an expert-designed ontology, build the knowledge graph to conform to it, and add an agent harness that governs what agents do with the result.
No. Evaluation measures how well an agent behaves by reading its logs, but it does not measure whether the agent produced the right result. Accuracy improves when the agent is governed by business rules and data in an ontology and knowledge graph.