Brand and omnichannel teams now put their HCP (healthcare professional) questions to an agent in Salesforce. Marketing operations hand a brief to a second agent, which plans the journey, waits for a person’s approval and builds a Draft in Marketing Cloud.
Before
Questions waited days, briefs waited weeks
An omnichannel lead could not plan a brand’s next email, field or digital program without waiting twice: once for an answer about the audience, once for a journey to review. One lead sits on each brand at a US pharma company with a multi-brand specialty care portfolio, accountable for how physicians hear from that brand.
- An HCP audience questionThree to five business days, restarted by every follow-up
- A written campaign briefAbout three weeks to a journey worth reviewing
Dashboards helped only when someone had built one for that exact question. Anything else meant a ticket: a central team of analysts picked it up and wrote SQL for the enterprise warehouse that holds the HCP data. (Analytics & Insights covers this work.)
For a campaign, marketing operations took each written brief and built it by hand in Marketing Cloud: the emails and templates, the entry events, the Data Extensions and the journey paths. Briefs left things out, and approved assets that already existed but could not be found were made again from scratch.
Why a generic AI assistant was never going to pass
Definitions
The numbers had to be right, not just fluent.
One HCP can be associated with several brands and appear in several preference observations. Count rows instead of distinct HCPs, or divide by the wrong population, and a confident answer is simply wrong. An earlier chatbot pilot had failed on exactly this.
Data residency
The data could not move.
The approved HCP dataset lives in Amazon Redshift under the company’s data governance. Any approach that copied it into a new application database was a non-starter for IT and privacy before the first demo.
Medical, Legal, Regulatory
Nothing reaches a physician without MLR.
Promotional content goes through MLR review. An agent able to activate or send a journey on its own would never be approved, however good its output looked. Governance had to be built in, not bolted on.
Reviewers from the customer’s IT, privacy and marketing compliance teams set the controls with Ontelis in weeks 1 and 2. Nothing was coded specially for this customer: the controls are standard Ontelis platform policies, and they bind both agents on every action.
The solution
Two agents. One governed context.
Ontelis forward-deployed engineers worked alongside the brand, omnichannel and marketing operations teams to build two agents on the Ontelis platform. Both draw on the same context: the connected systems, the business definitions the team signed off, and the policies that decide what any agent may read, write and show. The HCP Insight Agent answers inside Salesforce, where brand teams already work. The Campaign Agent builds in Marketing Cloud, where marketing operations already work.
Where people work
Agents
Ontelis platform
Systems of record
How it works
HCP answers on the team’s own definitions
Ontelis is AI infrastructure for pharma commercial operations that maps every system, records how the team decides, names who approves each step and logs every action for replay. Its forward-deployed engineers built two agents with the brand, omnichannel and marketing operations teams. The first, the HCP Insight Agent, takes the brand teams’ questions in Salesforce, through Agentforce. It reads only.
What brand teams can ask
Population and segmentation
Distinct HCP counts and percentage breakdowns by brand, segment, specialty and geography, each on the agreed denominator.
Cross-brand overlap
HCPs on both brands, either brand, or one and not the other, with preference observations restricted to the brands in question.
Channel, content and timing
Channel preferences and performance categories, digital affinity, content preferences, preferred days and sessions. Top five specialties shown, the rest grouped as Others.
Lookup and bounded lists
Find an HCP by name or NPI. Lists return the full matching count plus a preview of up to ten records, never a bulk export.
Scores, explained
Score averages alongside what each score represents, with a plain statement where the underlying methodology is not available.
Real follow-ups
Population, filters and metrics carry forward, so “and of those, who prefers email?” means what the user meant.
Omnichannel lead
Asks a population question.
HCP Insight Agent
Definitions first. IntuitionDB holds the counting rules, denominators and brand cohorts the brand analytics team ruled on, each labelled as a ruling.
HCP Insight Agent
One view. The agent writes its query for one place only, the approved HCP view, and takes that view’s schema from the Context Graph.
HCP Insight Agent
Checked first. Govern lets through a single read-only SELECT or WITH statement on the approved view, fields and functions, and stops anything else.
HCP Insight Agent
Inside the warehouse. Connect runs the query inside the warehouse, Amazon Redshift through Data Cloud zero copy; the rows stay there.
HCP Insight Agent
One repair. If the error is of a kind that may be repaired, the agent gets one corrected attempt, and that attempt goes through every check again.
HCP Insight Agent
An answer or a limit. A separate step writes the reply using nothing but what the query returned. When a question is out of scope or the query fails, the reply names the limit and does not make up a value.
Omnichannel lead
Follows up. The follow-up keeps the population, filters and measures.
The rows a query reads are never copied into Ontelis. A list request returns how many HCPs match and shows no more than ten of them; it cannot be used to export the rest. Where a score has no documented method, the answer points that out.
One brief, an approved plan, a Draft read back
The second agent, the Campaign Agent, works for marketing operations. Given a brief, it can come back with questions twice at most; whatever is still unanswered goes into the plan, written down as an assumption. It searches existing assets, read only, so approved ones are reused.
Then it proposes a plan: the journey diagram, the content, the entry rules, and each artifact it will create or reuse. The plan waits on a Workspace decision card, the card a person decides on, bound to that version of the plan; Govern’s word for this is Propose and wait. The approval stores fingerprints of the plan’s content and of everything it depends on; if either changes, the plan needs a new approval.
Agents write the same way every time: plan, check, approval, a write with an idempotency key so a retry cannot write twice, then readback. Here, every Draft is fetched back from Marketing Cloud after it is written, since Marketing Cloud accepting a request does not show the build came out right. One build produced 13 Marketing Cloud artifacts from a single brief, each read back, among them nine emails and a Journey Builder Draft.
Where the agent stops
The Campaign Agent writes only Draft journeys, Data Extensions with no rows, and entry events in Test mode. Govern holds it to draft writes, the class where nothing is sent; a write class is how undoable a write is, and it sets the default autonomy level. Loading audiences, switching journeys on and sending belong to people, each after MLR review, never to the agent.
Agents never sign, attest, approve content, assess an adverse event or release a payment above their limit.
Changes to prompts, flows and policies are versioned, with rollback, and the audit trail records questions, plans, approvals and builds. Context, policy decisions and check results reproduce exactly at pinned versions; model output is stored, not regenerated. The agents use provider adapters for several model vendors, and the customer’s model policy decides which runs.
Customers bring their own model contracts and endpoints, or run open-weight models inside their own environment.
Marketing & Content walks one sample campaign from brief to go-live.
“What won over our compliance team wasn’t the speed. It was that the agent shows its work, stops at Draft, and tells you when it doesn’t know.”
First agent live in week 9
Ontelis forward-deployed engineers were embedded with the customer’s brand, omnichannel and marketing operations teams, working with the customer’s commercialization services partner. No phase closed without evidence.
Controls and rulings
Weeks 1–2
IT, privacy and marketing compliance reviewers set the controls; the HCP view was linked through Data Cloud; brand analytics ruled on counting rules, denominators and cohorts, and the rulings went into IntuitionDB. Evidence: reference questions fixed in advance.
HCP Insight Agent, tested
Weeks 3–7
Checked against the reference queries and tried by omnichannel leads in UAT, with fixes for cohort preservation, formatting and score explanations. Evidence: each fix a versioned release that can be rolled back.
Accepted
Weeks 8–9. First agent live in week 9
Evidence: 14 of 14 live conversation turns passed in the final regression, and all 27 channel-by-day count rows matched the reference exactly.
Campaign Agent, built
Weeks 8–15
Built in a Marketing Cloud business unit set aside for it, and tested on new and reused content, splits, waits and DOCX briefs. Evidence: every build read back.
These are this customer’s weeks. Ontelis promises no duration.
FDE is the Ontelis service in which forward-deployed engineers audit a pharma company’s estate, build the first process with its team and take it to production. Ontelis staff and forward-deployed engineers propose but never approve anything in customer production, and they have no standing access.
Governance
The controls that got IT, privacy and compliance to yes.
None of these were custom code written for one project. They are Ontelis platform policies, configured with the customer’s IT, privacy and marketing compliance reviewers in the first two weeks, and applied to every action either agent takes.
| Control | What it does | Why it matters in pharma |
|---|---|---|
| Read-only query policy | Every generated query is checked before it runs: one SELECT or WITH statement, the approved view, approved fields and functions | An agent cannot change data or wander outside what it was approved to read |
| Zero-copy access | Queries run where the HCP data lives, through Salesforce Data Cloud | No new copy of HCP data to secure, govern or explain to privacy |
| Bounded outputs | Lists return the full count plus at most ten records | Answers inform decisions without becoming an export channel |
| Plan-bound approval | Approval attaches to the exact plan and its content and dependency fingerprints | What was reviewed is exactly what gets built |
| Draft-only writes | Writes stop at Draft journeys, empty Data Extensions and Test-mode entry events | Activation and sending remain human decisions, after MLR |
| Deterministic artifact keys | An identical retry reconciles with what already exists instead of duplicating it | A failed run does not litter Content Builder with near-identical assets |
| Independent readback | Content, fields, sendability, event linkage, journey paths and Draft state are read back after every build | Partial or uncertain builds are reported as exactly that |
| Model-agnostic | Provider adapters for Google Gemini, Anthropic and OpenAI, with deterministic code doing assembly and validation | The customer’s model policy decides, not the platform |
| Versioned releases | Every prompt, flow and policy change is versioned, with retained rollback versions and a full audit of questions, plans, approvals and builds | Replayable evidence for compliance and IT, and a safe way back |
Results
Analysts went back to analysis
| Before | With Ontelis | |
|---|---|---|
| HCP population question | 3 to 5 business days, via analytics ticket | Under 30 seconds, self-serve in Salesforce |
| Follow-up question | New ticket, new wait | Same conversation, context carried forward |
| Business definitions | Varied by analyst and dashboard | One agreed set in IntuitionDB, applied every time |
| Brief to reviewable journey | About 3 weeks | 1 day, including clarification and review |
| Existing assets | Rebuilt when nobody could find them | Found and reused before anything new is built |
| Evidence for review | Screenshots and spreadsheets | PDF plan, journey diagram and readback report for every build |
Three groups work differently now. Analysts spend their time on analysis, not on clearing tickets. Marketing operations review plans rather than retyping briefs. And the brand teams deciding whom to reach count HCPs exactly as the people building the sends do.
Results as of October 2026. How to read these figures. Under 30 seconds is a customer figure from the production deployment. 1 day is a customer figure from a Marketing Cloud business unit set aside for the agent, counted from a final brief. 14 of 14 and 27 of 27 are an acceptance test on a frozen reference set. No further method was stated.
These are this customer’s figures under the method above, not a promise for another.
Planned next, and the rest of Ontelis
Because both agents run on one context, every new capability starts with the definitions and policies already in place.
Into the field
Planned: the HCP Insight Agent opens to field teams for call planning, each person limited to their own territory.
Governed activation
Planned: audiences loaded and journeys switched on, but only after consent and suppression checks and behind the same plan approval.
Consent and content status are read live at send time.
MLR-aware content
Planned: approved claims and safety information placed exactly as the claims library holds them, with no rewording.
More channels
Planned: SMS, push and banner steps. Plans can name them today; none is built until a person approves it.
For IT and privacy reviewers:
Ontelis Mask decides, field by field, what Ontelis stores, what each reader sees and what may leave the data plane, and names who may change those decisions.
Mask: what is stored, shown or sent, per fieldCredentials stay in the customer’s vault, and read and write credentials are separate.
Connect: credentials and connectorsThe Context Graph links records and raises data change requests; Veeva Network or OneKey stays the master. Ontelis never creates or overwrites a golden record.
Context Graph: what it holds, and what it is notZero misses in n clean cases supports, at 95% confidence, only a miss rate under about 3/n.
Evaluations: what a clean pass can claimOntelis does not offer a general chat assistant. Copilot or Claude remain the front door and reach Ontelis through the Agent Gateway. Ontelis governs only what passes through it.
Agents: the front door for outside assistants
Security and data
- 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.
- 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.
- Your controls
- Ontelis’s posture as a vendor is kept separate from Govern, which holds your own controls.
The customer asked us not to name it, and we also withhold its partner, its brands and the people quoted, who speak here by role. The figures date from October 2026 and come from the production deployment and from a Marketing Cloud business unit reserved for the agent.
