Who we work with

Pivotal small cell lung cancer data was presented at ASCO 2025. The brand team needed a position on treatment sequencing, real-world adoption barriers, and trial prioritization while the readout was still current. A live advisory board takes four to six weeks to recruit and schedule.

An advisory board simulation ran about five days after the congress. KOL personas were built from publications, congress presentations, social channels, and practice settings, each synthesized from two to four real profiles. Personas debated positions against each other under an AI moderator rather than answering a discussion guide one by one. Every statement was logged and cite-checked for promotional review.

82% theme overlap with live KOL boards.

Around 40% of themes were new, covering unmet needs and adoption barriers not recorded internally.

Realism rated above 90% by reviewers.

Around 70% lower cost per study, with honoraria, travel, venue, and vendor logistics removed.

A medical affairs team evaluating simulated research for a vaccine portfolio needed a fidelity measure it could defend internally before scaling. A forward-looking study offers nothing to compare against. The team scoped a replication instead: rerun an advisory board it had already completed with real physicians and measure what the simulation reproduced and what it missed.

The completed board was replicated blind. Personas were built for the vaccine portfolio and the simulation ran on the original discussion guide without sight of the live board’s output. The two insight sets were compared line by line, and every simulated insight was classified as matching the live board or net-new. The team evaluated several competing tools in parallel.

100% alignment with the live board’s core conclusions.

106 of 490 insights were net-new, 21.6% of the total, raised by the simulation and not by the in-person board.

The evaluating director recorded that against the other tools assessed, nothing came close on realism, usability, or the ability to simulate meaningful HCP interactions.

Field capacity caps the rate of scientific exchange across an immunology portfolio. Allergists, dermatologists, rheumatologists, nurses, and pharmacists in remote and lower-priority territories had fewer touchpoints and waited days for answers held in approved documents. Every interaction had to be on-label and audit-ready. One-off queries were not captured, so medical affairs had no view of recurring information gaps.

An on-demand AI medical expert went live across the channels the field already uses: iPad in the room, Zoom and Teams, a secure phone line, and QR codes on leave-behinds. It was trained on the client’s on-label content, pivotal trial publications, safety monitoring protocols, and dosing and titration guidance, with escalation workflows for anything outside those boundaries. Multi-agent fact-checking and peer review ran before each response, with multilingual support.

At three months: more than 500 new HCPs engaged, including territories the field rarely reached, at three times baseline engagement efficiency.

Sessions averaged about 18 minutes and covered 50% more questions than baseline. Follow-up loops down 68%. HCP NPS 86.

On-label compliance 100% with zero incidents, 92% of responses cited a primary source, and rep confidence rated 4.7 out of 5.

MSLs recovered three to five hours a week, and 1,300 unique questions captured showed medical affairs the data gaps driving hesitation and the gaps in the approved content library.

Field insight is lost in the note, not in the conversation. Notes are written hours after a visit, in free text, into fields nobody queries. A specific payer barrier reaches headquarters as “good conversation, will follow up,” and brand and medical teams have no countable signal.

Post-call capture moved to voice in the minutes after the visit. The rep describes the interaction out loud and the assistant populates the coded fields for barrier, intent, and next best action, and drafts the follow-up in approved phrasing in the same pass. The rep confirms before anything is written to the CRM, with a full audit trail, role-based access, and client-controlled retention.

Post-interaction documentation time down more than 86%.

Over 80% of records returned with notes rated complete, specific, and carrying an actionable item.

Structured insights reaching brand and medical teams up three to five times per week.

Brand and medical affairs reported higher confidence in field insight, segmented by objection pattern, access friction by site, and interest by patient type.

Specialty access windows are short and unpredictable, which limits how much any single visit can cover. Clinical standards and access conditions change faster than the materials describing them, and payer conditions vary site by site. Preparation for the next call competes with administrative work left from the last one.

A voice-first field companion was deployed as a 12-week project, scoped around the moments before and after a call. Ahead of a visit it handled prioritization within a radius, formulary changes for a named HCP, and coverage points from the last visit. In clinic it returned prior objections, coverage at a named account, and evidence matched to the case in front of the rep. It connected to territory calendars, CRM call history, prescription trends, payer and formulary feeds, and the approved content library, and every CRM write-back required rep confirmation.

85 to 90% weekly active use among eligible field users.

HCP reach per field user up 20 to 30%.

Preparation and documentation time down 25 to 40%.

Tailored scientific discussions up two to four times, meaning conversations that referenced guideline evidence and local access together.

An immunology insights team wanted the simulated comparison run against its own physician panels, on studies it had already commissioned, and scored by its own analysts. Two studies were in scope, a steering committee simulation and a concept test, each of which takes 12 to 16 weeks by conventional methods.

Both studies ran in Simulation Studio inside a single four-week window. Personas were built for the same immunology segments the client’s real panels covered, so outputs were directly comparable. The client’s analysts ran the overlap analysis against their own prior research and classified every insight as matched or net-new.

92 to 97% overlap with the client’s real physician panels, on the client’s own analysis.

More than 19% additional insights the live panels had not produced.

Two studies in four weeks, against 12 to 16 weeks each by conventional methods.

Recruiting five or six rare-disease specialists is hard in any single market and it set the critical path for the launch plan for a global rare disease portfolio across Canada, Brazil, the Middle East, and Japan. Ex-US markets were often dropped from the research plan for that reason. Boards run months apart by different vendors also gave no comparable cross-market read.

Simulation Studio allowed for four advisory boards to run in parallel, each with 20 physician personas built for that market’s treatment guidelines, payer structure, and clinical practice. One discussion guide was held constant across all four markets. The team received full transcripts, a per-market synthesis, and a cross-market comparison, and kept platform access for follow-up questions to the same personas.

The Japan simulation covered all 48 insights from two completed in-person boards and added 29 more, a 60% uplift, including market-specific screening and scheduling considerations.

Four markets were delivered in about the time one human panel takes to recruit, against 12 to 16 weeks per board run sequentially. That is a roughly 75% reduction in research timeline.

Patient support programs are constrained by staffing, time zones, and call volume. Before a voice agent enters one, patient services teams need evidence on the edge cases: the off-label question from a caregiver, the symptom that needs a nurse that hour, and the claim the agent must not make. Internal testing cannot reach the volume and variability of a live program.

An independent panel of 1,000 registered nurses across all 50 states, median 14 years of experience, ran a prospective, multi-scenario, multi-therapeutic evaluation. They role-played patients and caregivers across 100 scenario scripts modeled on real program call logs, including GLP-1 initiation with cold-chain travel, postpartum patients on B-cell therapy, a pediatric CLN2 infusion, and neutropenia precautions during oncology treatment. Each interaction was scored on clinical accuracy, empathy, communication clarity, appropriateness of advice, and risk of inaccurate claim, with compliance observations logged alongside.

Across 8,000 completed interactions, top-box ratings: clinical accuracy 97.8%; appropriateness of advice 97.1%; risk of inaccurate claim 96.8%; empathy 98%; communication clarity 99%.

No harmful misinformation observed. Evaluators recorded that the agent held to its informational role rather than moving into prescriber territory.

Evaluator feedback changed the platform: a closing protocol that invites further questions twice, calibrated pacing on long procedural explanations, anchored name use, a language-preference check at call start, and dual-unit measurements.

Paid acquisition reached patients searching for IBS information, and the web form lost them. Form completion ran at 67%, and qualified-lead yield kept cost per qualified lead above target. Scaling a nurse line raises acquisition cost and still leaves no after-hours coverage.

Paid placements offered a voice conversation with an AI specialist in place of the form. The call followed the structure of a primary-care visit: consent and scope of call, then symptom history against Rome IV criteria, then contraindication screening and suitability for the OTC product class. Each call closed with a next step: a tele-consult booking, personalized education, or lifestyle guidance.

Over 48 hours: 120 voice conversations, six times the form-fill count for the same media spend.

60 clinically qualified leads, a 50% yield and 2.5 times the historic web-form yield.

100% complete record capture, against 67% on forms, and cost per qualified lead around half of recent social campaigns. Calls averaged just over three minutes.

73% of post-call respondents said the conversation felt like a preliminary nurse assessment. The agent provides no medical advice and the product is over the counter.

A brand team had completed qualitative research on its launch creative and still could not tell which elements were driving or suppressing resonance. Headline, primary visual, core claim, data presentation, and call to action all moved together in the feedback. Re-running qualitative across multiple formats did not fit the launch timeline, and HCP ad boards could not test variants in parallel.

HCP personas were built from anonymized qualitative transcripts, specialty literature, and validated profiles across experience levels and practice settings, then calibrated by decision heuristic: evidence threshold, sensitivity to overstatement, and visual versus data-first processing. Personas rated every section of each asset as useful, neutral, or not useful, reported where attention landed, and tagged the reason behind each low rating. They then debated the creative against each other. Outputs were synthesized into usefulness scores, attention maps, disagreement flags, and revision guidance.

Section-level diagnostics in 24 to 48 hours, tied to specific headlines, visuals, and claims.

Attention maps comparable to eye-tracking output, identifying ignored sections and misleading focal points.

1.5% higher cross-channel engagement after revisions and a roughly 2% improvement in post-launch KPIs against forecast on awareness and intent.

Around 67% lower cost than traditional ad boards, 94% perceived realism, and flyers, brochures, and social concepts tested in the same cycle.

Patient questions on starting treatment, what to expect, and side effect management arrived faster than the team could answer them. Coverage ended with the working day; the questions did not. Staffing an after-hours line raises cost per patient without improving the first answer.

An always-on voice agent handled inbound patient conversations, answering questions on medication, side effects, and next steps in plain language and routing to the live team when a question required one. Coverage ran through evenings and weekends without a staffed line.

More than 5,000 patient conversations in the first 48 hours.

9 out of 10 NPS from patients on the experience.

Strong starter-kit uptake following the conversations.

Pre-clinical decisions such as lead nomination, in vivo study design, species selection, and go/no-go into GLP toxicology turn on trade-offs between functions. DMPK reads a short half-life one way, toxicology reads a metabolite concern another, chemistry argues it is fixable, and pharmacology defends the efficacy margin. That disagreement usually surfaces in the committee meeting rather than before it. A wrong species choice or a premature go decision can cost an entire toxicology package.

Scientific personas were built as internal R&D scientist archetypes, grounded in scientific literature, ICH guidelines, and pre-clinical methodology so they reason about assay data, chemical structures, PK parameters, and species-specific biology. Five functions were represented: DMPK, toxicology, medicinal chemistry, pharmacology, and clinical pharmacology. The personas debated the target against each other, each arguing from its own evidence threshold. The output maps where consensus is real and where the disagreement sits, delivered as a structured pre-read before the committee convenes.