Confirm venue · Speaker program, US · run SPK-0217Logistics agent · 3 of 7 steps · started ET · waiting on the named approver
Waiting on a person
PlanThree writes, irreversible last: SAP purchase order, CRM event, confirmation email to the venue. The email needs its own approval.
Dry runDiff: one new purchase order. Impact list: the CRM event for SPK-0217 and one email to the venue.
ChecksApproved vendor, OIG exclusion list and meal limit pass.
ApprovalBlocked. This agent’s spend limit is zero, so the purchase order waits for the named approver on a decision card in Workspace.
ExecuteSAP · purchase order, create · under the agent’s own Ontelis identity, through the speaker-program agent class’s SAP integration user, with an idempotency key
Read backConfirm the purchase order in SAP.
Ledger entryActor, run, package, check results, approver and result.
Decision cardWrite packageLedger
Approve purchase orderWorkspace
Requested byLogistics agent
ApproverNamed approver for SPK-0217
WriteSAP · purchase order, create
Reason heldAgent spend limit is zero
IdentityAgent’s own Ontelis identity
Next writeConfirmation email to the venue, approved separately
Every agent spend limit starts at zero, so the purchase order waits for the named approver’s decision card in Workspace.
Trusted by 2 out of the 5 largest life sciences companies
Agents do the work around your decisions
Changes to configuration and data in the systems you run on, deployed through your own pipeline once the system owner approves.
Every write follows the same seven steps
1PlanList the targets, actions and arguments.
2Dry runShow the diff and the impact list.
3ChecksRun the check pack, the checks attached to the action type.
Approval, when the autonomy level requires itIf the step’s autonomy level, what an agent may do alone, requires it.
5ExecuteWith an idempotency key, so a retry cannot write twice.
6Read backConfirm the write in the target system.
7Ledger entryAn append-only, hash-chained record.
From Studio to production in five steps
Forward-deployed engineers, partners or your own builders build agents in Studio, the agent editor. An agent reaches production in five steps:
Draft in the dev environment, usually from a template.
Runs start from event triggers you define
A trigger watches one source and opens a run when an event meets its condition. A run is one execution of a process.
A custom trigger is replayed on the last 30 days of events, and only the owner of the target process can make it live.
The Agent Gateway
Connect is how Ontelis touches systems; the Agent Gateway is how agents touch Ontelis.
The Agent Gateway is the one door through which any agent reads context and asks to act. Ontelis governs only what passes through it. Agents connect through an MCP server, an API, an SDK, event streams or a proxy.
“We own the client relationship and the process knowledge. Ontelis gave us a platform and engineers who could turn that into agents our client’s AI committee would actually approve.”
Head of Agentic SolutionsThe customer’s commercialization partner
Who decides
Agents do
Plan, check and carry out every write, then log it.
Detect and route possible adverse events and off-label questions.
Agents never sign, attest, approve content, assess an adverse event or release a payment above their limit.
People decide
The write class sets each step’s default autonomy level. The step owner can tighten it. Lowering it takes the step owner and compliance together, within Govern’s limits.
Each agent’s owner answers for it and, with compliance on controlled steps, approves it for production.
A named decider approves each gated write on a decision card, which shows the agent’s proposal and check results.
What agents never do
Perform steps marked human only.
Sign on anyone’s behalf.E-signatures stay with people.
Send content to HCPs that is not MLR-approved (medical, legal and regulatory review).
Submit call reports or capture sample signatures on a rep’s behalf.
Commit spend outside an agent spend-limit row, or release a payment unless finance and compliance lowered that payment type.
An Ontelis agent is software that carries out defined steps of a pharma commercial process inside the systems you already run, such as Salesforce, Veeva Vault and SAP. For example, one raises the purchase order for a speaker program venue, and another matches an agency invoice to its PO. Each agent has its own identity, a named owner and limits on what it can change. It plans, checks and logs every change it makes, and anything that spends money waits for a person.
Your CRM vendor's agents are built for that vendor's platform, so they work from what the CRM knows. Most commercial work doesn't live in one system. Confirming one speaker program venue takes a purchase order in SAP, an event in the CRM and an email to the venue. Ontelis agents connect to all of those systems and work from the full picture, which is why their results hold up across the whole process. You can keep your vendor's agents too: they can run through the Ontelis Agent Gateway under the same checks.
Yes, once each Ontelis agent's CRM actions are re-mapped to the new CRM and re-certified. Ontelis runs on Veeva Vault CRM, Salesforce Life Sciences Cloud or both at once during a migration, with one set of checks. Before cutover it lists every agent, rule and check the move would break. Your rules and decision history carry over, consent checks are pointed at the new CRM's records, and the compliance tests rerun before you switch.
Yes, through the Ontelis Agent Gateway. Copilot, Agentforce, Claude or an agent you built connects over MCP, an API, an SDK, event streams or a proxy that needs no change to the agent's code. It never gets a Salesforce or SAP credential. It asks for the change, and Ontelis runs the same checks, approval and record as for its own agents, which get no shortcut. An agent that keeps its own system logins gets none of that: Ontelis governs only what passes through the Gateway.
No. Ontelis sets each step's default by how hard the action is to undo. Reading data and drafting a document run without approval. A reversible change, such as a CRM field value, goes ahead and shows up in the owner's digest, where it can be reversed. A purchase order or an email outside the company waits for a named approver, and every agent's spend limit starts at zero. The step's owner can tighten any step; loosening one also takes compliance.
Ontelis is built to catch a wrong step while it can still be undone. Each change is checked against the system of record before it's made, and a stale or unavailable source stops the change and sends it to a person. The agent reads the record back afterwards. On a job across systems, every step has a defined reversal, and the one that can't be undone, such as the email to a venue, runs last. If an agent's error rate passes a set threshold, a circuit breaker pauses it.
Not to promote. Ontelis agents never make promotional calls or send promotional texts to HCPs, and no setting, contract term or admin role changes that. Outbound, voice agents call venues, vendors, agencies, PBMs, distributors and specialty pharmacies on trade matters. When an HCP's office calls in, they handle logistics such as scheduling and event details, transfer medical questions to medical information and send adverse events to safety. Every call opens with an AI disclosure and a request for consent to record.
Ontelis agents run on the models you choose: your own model contracts, open-weight models inside your environment, or endpoints Ontelis has approved if you have no contract. Ontelis builds no foundation model. A model takes a step only if Mask allows that step's data to reach it and the model passed that step's tests. Versions are pinned, and changing a model reruns the tests.
Only if you want to. Ontelis forward-deployed engineers, a partner or your own builders build agents in Studio, the Ontelis agent editor, usually from a template. Whoever builds it, an agent takes the same route to production: it passes its test suites, is registered with a named owner and a risk tier, and needs that owner's approval, plus compliance approval where it works on controlled steps. Then it's promoted to test, and only after that to production.
Agents never hold secrets: a credential broker issues a short-lived token for each call, from the customer’s vault.
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.