A governed AI knowledge layer sits between your documents and your agents, and it does six jobs: ingest, structure, harmonize, validate, measure and govern. It controls what enters the system and records where it came from, turns documents into typed records or retrievable meaning, resolves the words on the page into your business vocabulary, routes low confidence results to a person, scores output against human curated ground truth, and enforces who is permitted to see what.
Context Studio is the platform Revenue Experts AI and Ollon built to do those six jobs.
This is part two. Part one covers the three failures that produce answers which look correct but are wrong, and why a better model does not fix any of them. Start there if you have not read it.
Most enterprise AI programs do not fail on model quality. They fail on the layer underneath, which is the unglamorous work of knowing what came in, what version it was, what the words mean inside your business, who is allowed to see it, whether a person changed it, and whether any of it is measurably right.
This article defines that layer, covers the six jobs it has to do, explains where permissions actually come from, states what Context Studio does not do, and sets out how to scope a first project so it produces evidence inside a quarter.
A governed AI knowledge layer contains six jobs: Ingest, Structure, Harmonize, Validate, Measure, and Govern. They are not sequential steps in a project, they are functions that all have to be present, and a gap in any one of them shows up as a wrong answer somewhere else. Context Studio implements all six, and the sections below describe each one as it is built rather than as an ideal.
Above those sits activation, which is how agents actually reach the result. The design decision that matters there is that agents get a controlled tool layer rather than a database connection. A tool layer can carry the caller's permitted scope into every call automatically. A database connection relies on the agent behaving, and an agent is a text generator, not a security control.
This is the job most platforms skip, and the one that decides whether the output is usable for reporting. It is also where Context Studio does the most work that a buyer never sees.
Raw model text resolves to canonical terms through an ordered cascade. Curated dictionary rules first, because a term your business has explicitly defined should never be guessed at. Then declared match rules, then fallback rules, then fuzzy distance matching, then reverse containment, then vector similarity as the last resort rather than the first.
Rules are scoped, so a client specific rule beats a general one, and a supplier specific rule beats both. That ordering is what lets one organisation's exception stay an exception instead of becoming everyone's default.
Both the raw text and the resolved value are preserved end to end. A reviewer can always see what the model actually read, which is the difference between a correction that takes ten seconds and one that requires opening the source document.
Not from the knowledge layer, and this distinction matters more than it sounds. Context Studio does not define who is allowed to see what, and a platform in this category that claims to should be treated with suspicion.
Your identity provider and your systems of record already define who is allowed to see what. Roles, groups, record level rules and the exceptions that accumulated over a decade all live there. A platform that invents its own parallel access model creates a second source of truth, and two sources of truth about permissions will drift. When they drift, the safe one is not necessarily the one that wins.
So the knowledge layer does not establish entitlements. It inherits them and enforces them at the point where data is retrieved. Entitlements resolve per request from the authenticated user, and fold into the query layer itself rather than being applied as a filter afterwards. The distinction is that a filter can be bypassed by a query that was never filtered, while a constraint compiled into the query has nothing to bypass.
There is a large difference between instructing an agent to stay inside a user's permitted scope and building a data layer that cannot return anything outside it. The first is an instruction, and instructions can be talked around. The second is a constraint.
Enforcing an access model you did not author is the harder problem, and it is the one worth asking any vendor about. Ask where the check lives. Ask what happens when the caller supplies a scope of their own. Ask what an administrative override costs, and whether it leaves a record. If someone can break glass without leaving a trace, the audit trail describes ordinary days only.
Context Studio scores against human curated ground truth rather than a demonstration set, and the reason is worth stating plainly. Start with what the accuracy number is measured against. If the answer is a demonstration set, it is a sales figure. If the answer is human curated ground truth on a named corpus, stored immutably and snapshotted per run so old results stay reproducible, it is a measurement. A percentage without a stated corpus, an expected field set and a scoring rule is not a claim anyone can check.
Then ask what happens over time, because a benchmark that does not sharpen is a photograph. When a reviewer corrects a value, that correction should be promotable into the ground truth so the benchmark improves with use. When the configuration changes, the previous score should be flagged as stale rather than continuing to describe a system that no longer exists.
Reproducibility belongs in the same conversation. Prompt templates should be versioned rather than edited in place, and the instruction set should be frozen onto a processing job at the moment it starts, so editing a template later does not retroactively change work already done. Six months on, you should be able to reconstruct exactly what the model was told.
The same applies to human corrections. Model output stays immutable, reviewers work on a separate copy, and every change is recorded at field level with the old value, the new value and the person. Six months later, the question of whether a machine read a value or a person typed it always has an answer.
Context Studio is an enterprise AI platform built and owned by Revenue Experts AI and Ollon, in production with an enterprise customer since May 2026. It implements the six jobs described above: ingestion with recorded provenance, structure, business vocabulary resolution, human validation, measurement against curated ground truth, and governed retrieval.
It sits above your cloud, your models and your agent platform rather than replacing any of them. Your agents call it and receive grounded, permission filtered context with citations back to the source passages that produced each assertion.
What it is not: it is not a data catalog, not an analytics agent, and not a semantic layer on its own. Several unrelated products in the data and advertising markets also carry the name Context Studio. This one is the platform from Revenue Experts AI and Ollon, and every capability and limitation described on this page refers to that platform.
A document that lists only strengths reads as marketing regardless of how specific the rest of it is. These are the current edges of Context Studio, and the same list should be requested from every vendor on your shortlist.
The most common way this goes wrong is starting everywhere at once, which produces a platform nobody can point at a result.
One or two repeatable use cases inside a single knowledge domain. Confirm the reference architecture, agree who owns what across delivery, and define the success measures before anything gets built. That last one is where most programs lose, because success measures agreed after a build are negotiated against what the build happens to produce.
Knowledge discovery and document compliance intelligence are the strongest openers. Both address a problem the business can already name, both produce evidence of accuracy and security inside a quarter, and both leave behind the ingestion, vocabulary, validation and governance foundation that every later use case needs anyway. Nothing built for the first use case gets thrown away for the second.
The question to put to any platform in this category is not whether it can produce an answer. It is what happens when the answer is wrong: whether you can see which source produced it, tell whether the system misread the document or misfiled the term, prove who was permitted to see it, correct it without an engineer, and demonstrate afterwards that the correction is measured and will not quietly regress next month.
Context Studio is an enterprise AI platform built and owned by Revenue Experts AI and Ollon, in production with an enterprise customer since May 2026. It handles ingestion, structure, business vocabulary, human validation, measurement and governed retrieval, and hands agents grounded, permission filtered context with citations to the source passages. Unrelated products in the data and advertising markets share the name.
Retrieval is one part of it. RAG answers the question of what to retrieve for a given prompt. A knowledge layer also decides what enters the system, which version is current, what the words mean inside your business, whether a person changed a value, whether the output is measurably correct, and who is permitted to see it.
No. Your identity provider and systems of record define who is allowed to see what. The knowledge layer inherits those entitlements and enforces them at the point data is retrieved. A platform that invents its own parallel access model creates a second source of truth that will drift from the first.
The first phase is a production readiness sprint that answers one question, whether there is a credible path to production for the chosen use case. Implementation is scoped afterwards, based on what the sprint found. If the sprint does not establish a credible path, the engagement stops there.
Book a conversation and we will scope a sprint that settles one thing only, whether there is a credible path to production for your first use case. If it does not establish one, the engagement stops there.
