Four Ontelis agents now carry each campaign from its first planning input to a switched-off journey in Salesforce Marketing Cloud. All nine teams now share one campaign contract. The same three approvers sign every diagram, as before.
Before
One campaign, nine teams, retyped by hand
A pharma company’s US campaign operations group runs existing-brand email campaigns for 13 brands, between 200 and 250 a year. Every campaign crossed nine teams. Reaching Marketing Cloud took 40 to 50 calendar days, and the teams between them spent roughly 131 human effort-days on it.
Measured stage by stage, the delay was not in any one team’s work. It was in passing work from team to team by email, often with several stages running at once. With every campaign that costly, brand teams settled for broad emails over targeted, segment-specific ones.
Before Ontelis, each stage as the customer measured it. The stages overlapped, so the rows do not add up to a campaign’s calendar days.
Campaign inputs
Planned 2 days10 to 15
Agency teams of 10 to 12, filling in fields when they could, with no deadline
Journey diagram
Planned 5 days7 to 8 working days
The customer’s solution architect, drawing each journey by hand in Visio
Approval
Planned 2 days3 to 4
The customer’s solution architect, an experience manager and brand, each seeing the diagram but not its inputs
Segmentation
Not plannedabout 22
Audience teams, running test enrolment and building target lists in Data Cloud
Marketing Cloud build
Planned 3 daysabout 3
Deployment specialists, rebuilding the journey from the diagram and the planning form
A single campaign’s inputs were spread across the planning form, the project plan, a handoff checklist, the content and metadata matrices, the consent survey sheet, brand decks and email threads. More than 50 of its 180-plus fields arrived by email.
The journey existed only as prose, typed into a free-text “campaign purpose” field. One person redrew it as a diagram, and a second rebuilt that diagram in Marketing Cloud. Deployment specialists built from the diagram, so any error on it went out to physicians or patients when the journey was sent.
Delivery
113 questions before the first agent
The customer’s long-standing commercialization partner led the program. It kept the client relationship and ran the functional consultation. Ontelis engineers were the build partner and provided solution architecture for the program (how Ontelis engineers take a first process live).
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.
The customer
- Set the gates every phase cleared before build
- Requires a person to validate anything before it deploys
- Its solution architect, an experience manager and brand approve each diagram
- Campaign operations switch each journey on
- Marketing Cloud operations keep business units, sender profiles and data extensions
The partner
- Led the program
- Kept the client relationship
- Ran the functional consultation
- Its team received the knowledge transfer after the week 24 go-live
Ontelis engineers
- Build partner, and solution architecture for the program
- Ran discovery
- Wrote 28 requirement rows into the customer’s URS (user requirements specification)
“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.”
The partner delivered this program on Ontelis as a Delivery Partner.
Ontelis staff and forward-deployed engineers propose but never approve anything in customer production, and they have no standing access. Ontelis supplies the control. The accountable people stay with the customer.
Nothing was built until each phase cleared the customer’s gates: sign-off of the BRD and FRD (business and functional requirements documents), the AI committee’s approval, and sign-off of the technical design and architecture. This deployment was classified GxP none (GxP: the good-practice quality regulations). Cadence: a scrum each week, a sprint review every two weeks, an executive review each month.
Discovery took ten weeks. Ontelis engineers held a working session with each of the eight teams that own a system, used a written questionnaire of 113 questions, and worked from one swimlane covering all nine teams. Four live journey diagrams were broken down into their individual shapes, and no team was asked a question an earlier session had already answered.
Two of discovery’s four findings withdrew kickoff promises in writing. One was that the agent would build audiences itself. The customer’s Data Cloud had no native AI segmentation and audiences already flowed into Marketing Cloud, so the agent checks the segments people build instead. The planning-tool integration was put off, so teams enter campaign fields in Ontelis Workspace. And two kinds of code, kept in the consent survey sheets, widened the contract the first two agents share.
Discovery
Weeks 1 to 10
Ended on the four findings told above.
Design and the customer’s gates
Weeks 6 to 12
Ended when the requirement documents were signed and the AI committee approved.
The first two agents
Weeks 10 to 19
Phase 1. Week 19: the first brands go live on Workspace and the first two agents.
The last two agents
Weeks 12 to 24
Phase 2. Week 24: all four agents live, with hypercare after.
These are this customer’s weeks. Ontelis promises no duration.
How it works
Nine teams, one contract, typed once
Each campaign is one run: one execution of a process, which can wait weeks for a person. Four agents carry it in turn. For one brief followed through MLR review, see the campaign walk on Marketing & Content.
Requirement Collection agent
Inputs, gathered once
Agents open one run per campaign when a request arrives, and the run waits on people for as long as the campaign takes. The Context Graph maps the 180-plus fields to 17 business events and nine teams, and records each field’s phase, trigger, current source and future source.
So the Requirement Collection agent knows whose field is due, and when. A team is asked for its own fields only, at the phase that needs them. A late field draws notices that escalate, and once a team’s inputs are complete, the next team hears at once. In Workspace each team sees, enters and confirms only its own fields.
When two sources give different values, the agent notes which one it used and why, and the conflict goes to a person. Inputs that took 10 to 15 days now take a median of 4.
Flow Planner agent
The diagram, drawn from the contract
The Flow Planner agent turns the contract into an editable Visio file built from the customer’s 43 house shapes, with a report explaining each value. IntuitionDB holds the 19 eligibility and suppression rule types and the house conventions the diagram expands from, as rulings. Dates are computed two independent ways, and the layout is seeded, so the same campaign gives the same diagram.
Its acceptance test is done by hand in Visio. Move a send node, and every connector follows it without coming loose. A diagram that took 7 to 8 working days to draw now takes minutes to generate, then is reviewed and edited the same day.
The Context Graph keeps a source, a method and a label on every value; a value two sources disagree on is marked ambiguous and goes to a person, whose ruling is recorded. Pick a value to see its trail.
Six pure stages, then a gate
- SchemaFail fast on contract shape and ID patterns.
- IntegrityEvery touchpoint resolves; offset chains form a DAG.
- SemanticsDates increase; resends carry a filter; nothing after retirement.
- ResolutionDates by two independent paths, weekday and black-out rules applied.
- Diagram modelRules expand from IntuitionDB into decision nodes; tests diff the model.
- LayoutDeterministic and seeded: same campaign, same geometry.
Every generated Visio file passes seven structural checks. Those checks also turned on the customer’s own files. Collecting the 43 house shapes from live production diagrams, they found one whose PDF export had been losing a whole journey wave, through a broken page reference no one had seen.
The same files held two more slips. A contact rule on one brand’s diagram said “Yes → Stop”; the right form, used elsewhere, is “No → Stop”. And templates reused from older campaigns still pointed to a consent platform retired years before. New diagrams do not inherit these: the agent flags each one rather than carry it forward.
“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 Operations, one of the top 10 global pharma companies - The customer’s solution architectExperience managerBrand
Three sign-offs on a sourced diagram
The customer’s solution architect, an experience manager and brand approve each diagram, and nothing after it moves until all three have. They are the same three approvers as before. Govern holds them as decision rights, and its audit trail records how each computed date or label was derived. Reviewers once saw the diagram without its inputs; now the report beside it shows each value’s source and flags each conflict. Approval that took 3 to 4 days now takes a median of 2.
Segmentation Validator agent
Segments checked before enrolment
The audience teams still own segmentation: they build each segment in Data Cloud, working from the diagram once it is approved. Before any test enrolment, the Segmentation Validator agent compares each segment with the diagram: its filters, its suppressions, and who it includes and excludes. It reports any gap it finds. Segmentation that took about 22 days now takes a median of 6.
Journey Builder agent
Pushed in parts, switched on by people
Connect reads the planning exports, consent metadata, MLR references and email QA and tagging data, plus Marketing Cloud and Data Cloud. Marketing Cloud takes or refuses a journey whole, and its error says that the journey failed, not where (Connect: when a source changes or fails).
So the Journey Builder agent checks a journey one part at a time, deploying none of it, and the part that fails names itself. The build borrowed the method from Salesforce Flow deployments. Agents can push a journey only switched off; switching it on is a person’s step. A build that took about 3 days now takes under 1.
Two of the four agents are new work. The other two work alongside Salesforce tools that already ran, adding checks and replacing none of them. All four hand work to each other only in the campaign’s Workspace. The contract also keeps the first two apart: Requirement Collection deals only in sources, Flow Planner only in diagrams. Either can change without touching the other, whether the planning tool’s export changes or the house diagram conventions do. Because of that split, the build absorbed all four discovery findings and never had to be rewritten.
A confident wrong answer is worse than no answer.
| Mechanism | What it does | Why it matters here |
|---|---|---|
| Nullable with a reason | “Not applicable” and “nobody knows” are different states in the contract | Gaps stay visible instead of being guessed |
| Configurable failure mode | Block, annotate the diagram with visible callouts, or ignore, per run | Reviewers see problems where they already work |
| Derivation log | Every computed date or label records its method, inputs and cross-check | Answers “why does this say 21 August?” in an audit |
| Deterministic output | Seeded layout; diagram models snapshotted in tests | Same campaign, same diagram; regressions are diffable |
Results
From 131 effort-days to a median of 46
Each campaign now takes a median of 46 human effort-days, counted across all nine teams, against roughly 131 before. The same headcount now has nearly three times the campaign capacity. Each campaign takes a median of 14 calendar days from first planning input until the journey is ready to switch on, against 40 to 50 before.
One contract did more than any one agent. A value is entered once, by the team that owns it, and the other eight teams work from that value instead of retyping it, reading it their own way or asking for it again. Every sign-off is unchanged.
| Stage | Before | With Ontelis |
|---|---|---|
| Campaign inputs complete | 10 to 15 days | 4 days, with each team asked only for its own fields |
| Journey diagram | 7 to 8 working days | Generated in minutes, reviewed and edited the same day |
| Approval | 3 to 4 days | 2 days, with provenance and flagged conflicts in the report, not in email |
| Segmentation | About 22 days | 6 days, with gaps caught before test enrolment |
| Marketing Cloud build | About 3 days | Under 1 day: pushed deactivated, reviewed and activated by ops |
| Calendar days per campaign | 40 to 50 | 14 |
| Human effort-days per campaign | 131 | 46 |
Results as of October 2026. Method: each after figure is a median over the campaigns run since Phase 2 went live in week 24. Each before figure is the customer’s own baseline for that stage. The before values are ranges, and the stages ran in parallel, so stage figures do not add up to the totals. Other changes in the period may also explain part of the difference.
These are this customer’s figures under the method above, not a promise for another.
Planned next, and the rest of Ontelis
Each item below is a plan, not a result.
Planning-tool integration
Inputs: agent steps are to read from the planning tool and write back to it, so agency teams no longer type a handful of fields into two places.
Native segmentation, orchestrated
Segments: the Segmentation Validator is to direct Salesforce’s own segmentation agent and check its output, once that agent is generally available to the customer.
The Agent Gateway is the one door through which any agent reads context and asks to act. Ontelis governs only what passes through it. If an agent does not come through the Agent Gateway, it is labelled observed and gets no Ontelis checks. Agents: the Agent Gateway for outside agents
Marketing Cloud Next ready
Journeys: a later switch to Salesforce’s next Marketing Cloud should touch one adapter in the journey step; the contract and the first two agents do not depend on the platform.
A graph diff lists what a system change breaks, before anything moves (a system change, walked).
New launches and markets
Scope: launch campaigns, and markets outside the US, are to run on the same contract.
The same eight-step run carries other commercial processes, such as a speaker program or a quarterly PBM rebate validation.
IntuitionDB records the decisions people make at process steps, tests what other choices would have produced, and turns options that beat the benchmark into top rules.
IntuitionDB: how rulings become tested rulesOntelis Oversight is one screen per process in production: live runs, queues, failed checks, overrides, drift, cost and cycle time, with pause and kill switch beside them. A low override rate is not proof of quality.
Oversight: one screen per process in productionSome rules are platform invariants. No setting, contract term or admin role changes them. There is no admin mode that skips Govern. The safety screen reads raw text before any Mask treatment applies. The safety screen runs on all text. The metric firewall applies to every value decision. Regulated steps are never auto-applied.
Govern: the rules no setting changesThe customer owns what is learned: rules, decision records and edge weights stay in its data plane and can be exported. Customers can export the graph, rules, decision records and ledger in open formats at any time.
What the customer owns and can export
Security and data
- Your data
- Ontelis owns no business system of record. It writes into those systems and never replaces them.
- 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.
The customer asked us not to name it. To protect confidentiality we also withhold its partner, its brands and the people quoted, who speak here by role. Results are medians as of October 2026.
