The Context Graph is a knowledge graph of all the business systems a customer connects, and everything inside them, used as a map for agents to navigate.
The Graph never uses a prescribing-derived input or outcome in a decision that gives an HCP something of value. Govern enforces this with a lineage check on the graph.
Queried, not copied
Slow-changing things such as HCPs, territories, contracts and content are stored as nodes.
High-volume rows such as claim lines, calls and invoice lines stay in their source systems and are read by query, through named query templates bound to nodes. Files with no source system are read the same way from your landing zone.
Two dates on every fact
Every fact keeps two dates: when it became true, and when Ontelis found out.
Nothing is deleted. So Ontelis can rebuild exactly what an agent saw, answer incentive compensation (IC) disputes and audits, and let IntuitionDB judge a past decision only on what was known then.
Every edge carries one of three labels. Inferred edges never satisfy a check. Ambiguous edges go to a named owner, whose ruling turns them extracted. Conflicting definitions are all kept with their dates.
One package per step
A context package is the small, permission-filtered, signed bundle an agent receives before a step: facts, rules, suggestions, notes and check results. Every item carries its source, as-of time, edge label and treatment.
An agent sees a value only if all four allow it: the agent’s scope, the user’s Ontelis permission, the source’s own permission and the Mask serve treatment.
Check inputs, rules in force, required approvals, Mask markings and safety flags are never cut. Graph data is for planning; a check that gates an irreversible or external action reads its system of record live.
The Agent Gateway is the one door through which any agent reads context and asks to act. Ontelis governs only what passes through it.
What a change will break
When a source ships a release, Ontelis diffs its new metadata against the graph and lists the agents, rules, checks and connectors that break.
A same_as link can be reversed. Middling scores never link on their own; they go to a steward. Sampled audits catch errors.
The graph reports what the source says, with its timestamp. It does not correct the source.
Each connector declares what it does not read, so a package says “not covered” instead of implying nothing exists.
“The diagram used to be the thing we argued about. Now it shows us where every value came from, and it flags what our own templates got wrong.”
Senior Solution Architect, Campaign OperationsOne of the top 10 global pharma companies
Who decides
Agents do
Read a signed context package before each step.
Get the impact list before proposing a change.
Read context and ask to act through the Agent Gateway.
People decide
Named owners rule on ambiguous edges and conflicting definitions; data stewards confirm middling matches. Each ruling records who made it and when.
Process owners confirm each process model before agents run on it.
System owners sign off the first graph.
An approver other than the builder approves each deploy.
What the Context Graph does not do
Act as an MDM.Veeva Network or OneKey stays the master.
Act as a warehouse.
Act as a BI tool.
Act as a vector index.Embeddings may help find entry nodes, but they never create edges.
Correct a source.
Where it fits
The Context Graph builds step 2 of every run, with IntuitionDB and Mask.
The Ontelis Context Graph is a map of your business systems and what's inside them, built for agents. It records what exists and how it connects: the fields on an SAP purchase order, a Snowflake dbt model, a Vault approval workflow, and how your doctors, territories and contracts relate across them. Agents use the map to find the context a task needs, then read the actual values from the source system at that moment. Each agent gets only what it's permitted to see.
No, the two do different jobs. Your MDM, such as Veeva Network or OneKey, keeps the golden record for each doctor and account. The Context Graph gives agents a map of your systems and what's inside them, so an agent can pull the context it needs for the task in front of it. It links one doctor's records across those systems using hard keys like NPI, state licence and your MDM's ID. A doubtful match goes to your data steward or becomes a data change request to your MDM.
No. The Context Graph maps your systems; it doesn't hold your data. It stores metadata: which systems you have, what objects and fields are inside them, and how they connect. The values themselves, such as claim lines, calls and invoice lines, stay in Snowflake, Veeva or SAP. When an agent needs a value, Ontelis queries the source system on demand, so the agent works from what the system says right now.
Your data doesn't have to be clean, and you don't need an MDM first. The Context Graph reports what each source says, with its timestamp, and doesn't correct it. When two systems disagree, both versions are kept with their dates and the question goes to a named owner on your side, whose ruling is recorded. With no MDM, records link to NPPES, the public NPI registry. Where a connector doesn't read something, the agent is told "not covered".
Before any action that can't be undone or that leaves your company, Ontelis reads the system of record live. The Context Graph is for planning, not for that last check. Consent, for example, is read at the moment an email is sent. If the source is stale, down or contradicted, the action stops and goes to a person. Inside the graph, each kind of fact has its own freshness target, and anything older is flagged as stale wherever an agent sees it.
The Context Graph gives an agent exact links where a vector search gives it similar-looking text. A doctor's CRM record is joined to their prescriptions in your warehouse on NPI, and the agent follows typed links from there. Most of the graph is parsed by rule from schemas, code, permissions and lineage, with no model involved. Links read from prose such as SOPs are labelled inferred, and no check relies on them. Embeddings can help find a starting point. They never create a link.
Ontelis needs your people for sign-offs and rulings. The mapping itself is mostly parsed by rule. The first graph covers the systems behind the processes you choose, and your system owners sign it off. Process owners confirm each process model before agents run on it. Where parsing can't settle something, such as two definitions of new-to-brand, your experts rule and Ontelis engineers write the ruling in. After that, only changes are reprocessed, from each connector's change feed.
Row values such as claim lines stay in their source systems and are read live through named query templates.
Vendor documents
Ontelis provides: a SOC 2 Type II report, an ISO 27001 certificate, a HIPAA business associate agreement wherever Ontelis processes PHI, a data processing agreement with transfer clauses, and a sub-processor list.
Models
Ontelis builds no foundation model. No model is trained on customer data, except an opt-in adapter that never leaves the customer’s data plane.
Your controls
Ontelis’s posture as a vendor is kept separate from Govern, which holds your own controls.