·AgentOlympus ·Product Design

Designing AI for experts

AgentOlympus is BioNTech's internal AI platform. As its sole product designer, I turned a long list of AI ideas into one way of working that five business teams now use, and a design system built to hold 60+ more.

workflowsup from a 2-workflow pilot

usersacross 5 teams, up from 50

faster to launch8 weeks to about 3

less time on evidence synthesisin pilot testing

Decoding Data

RoleSole Product Designer

PartnersPMs · Engineers · Leadership · 2–3 SMEs per workflow

TimelineSince January 2026

Scope5 business teams

01 · The problem

More time assembling evidence than assessing it.

Leadership had identified about 60 places AI could help across commercial, medical affairs, market access, GVAPP and HEOR. In oncology, every claim in a dossier or leadership deck has to trace back to a reliable source, so specialists spent more time gathering, cross-checking and reformatting evidence than analyzing it.

My job was to turn that list into tools experts would trust with real decisions, and to make each new tool faster to build than the last.

The AgentOlympus home screen, showing the workflow launcher — Every team starts from the same place.

02 · The approach

Five teams, the same four jobs.

I started with the people: discussions with each business team, leadership and PMs, then step-by-step journey mapping with 2–3 SMEs per workflow. Very different teams turned out to be doing the same four jobs.

Find

Question an evidence base directly.

Compare

Surface differences across sources or markets.

Synthesize

Turn evidence into structured findings.

Produce

Move findings into reports and decks.

One way of working

So instead of designing five tools, I designed one way of working. The AI prepares the material; the expert decides what is trustworthy and what moves forward.

Before AgentOlympus

Clinical literature

Pricing data

Congress sessions

Read, cross-checked and reformatted by hand, across several tools

AnalysisRecommendationSlide deck

Manual assembly across several tools

With AgentOlympus

Clinical literature

Pricing data

Congress sessions

The platform assembles

FindCompareSynthesizeProduce

The expert reviews, corrects and decides

AnalysisRecommendationSlide deck

The assembly is handled, the judgment stays with the person

What shipped, and in what order

Each release established something the next workflow could reuse.

Step 01 Histology Data Extraction Dossier Mapping Market Research · HEOR

Established reliable evidence extraction.
Step 02 OncoAthena Commercial

Established conversational search and citation.
Step 03 SlideGenie Commercial

Extended findings into presentation-ready output.
Step 04 TPP Builder Commercial

Introduced the reusable assessment pattern.
Step 05 Congress AI Medical Affairs

Brought contributed material into the evidence base for review during medical conferences.
Just launched MR Echo Launched Landscape Assessment Launched

Extended the conversation and assessment patterns already in place.

HEOR: health economics and outcomes research. GVAPP: global value, access, pricing and policy.

03 · What research changed

Three principles, each proven in a pilot

Principle 1 · Landscape Assessment

Lead with structure

GVAPP specialists already knew which categories mattered: standard of care, competitors, pricing and approvals. A blank chat box would have made them retype that knowledge. So the tool opens with those categories filled in by indication, line of therapy and country, and chat handles what the structure can't anticipate.

Result

The method lives in the product, not in the prompt.

Landscape Assessment treatment guidelines compared across USA, EU and UK

Easy access. Guidelines, pricing, competition and HTA sit one click apart, already filled in.

MR Echo showing pre-populated Key Business Questions

Structured in, structured out. The key questions come answered on arrival, each linked to the quote or table behind it.

Principle 2 · MR Echo

Structure what goes in and what comes out

An AI answer is only as good as the data behind it, and only as useful as the form it arrives in. So I designed both ends.

Input

MR Echo takes discussion guides, transcripts, quant reports and respondent profiles in a defined structure, so every study is read the same way. The key business questions market access asks of every study come already answered.

Output

Instead of one long answer, results are broken into digestible parts, with the qualitative and quantitative findings shown side by side and each linked to the quote or table behind it. Landscape Assessment follows the same rule for competitive analysis, biomarker prevalence and epidemiology.

Result

Teams start from a first draft they can scan and verify, not a wall of text.

Principle 3 · Congress AI

Put review inside the workflow

Conference material is unverified, and experts warned that not every upload should become evidence. Uploads now wait in Pending review until the team approves them for AI use.

Result

Governance became a product behavior, not a policy document.

Congress AI dashboard showing documents awaiting review

Review before trust. Uploads wait in Pending review until the team lets them in.

04 · A decision I changed

Confidence scoring: keep the signal, change how it's shown.

Early answers displayed a confidence score as a percentage. But once one answer showed 90% and another 40%, the obvious question was: why show the 40% one at all? The number suggested a precision experts couldn't verify.

So confidence scoring stayed, but moved behind the scenes. Instead of a percentage, it now decides which answers are marked AI-recommended. Experts can select more, and every answer cites its source so they can check it themselves.

Result

The same confidence signal, shown in a form experts can act on and verify.

Before

AnswerConfidence 40%

Community physicians were more cautious and wanted routine testing in place before changing practice.

After

AnswerAI-recommended

Community physicians were more cautious and wanted routine testing in place before changing practice.

Interview 2 · 00:44 · "…wanted routine testing in place…"

The score decides what is recommended; the citation lets the expert check it.

05 · Designing for scale

A system built for 60+ workflows

Getting to a design language first. Agreeing on a language leadership recognized as BioNTech's took longer than designing any single screen. The direction existed only in conversation, so I separated what the brand had already settled from what was still open, and brought clickable directions instead of static screens. Clickable beat flat every time.

Foundations

Color, type, spacing and base components.

Shared patterns

Layouts and behaviors reused across workflows.

Workflow configuration

What a workflow changes to suit its own job.

From shared to specific

The blocker

The direction existed, but only in conversation.

Leadership's preferences existed, but had not been translated into shared principles.

The decision

Start from what is settled, then offer a choice.

I separated established brand decisions from areas that still required alignment.

What changed it

Clickable beat flat, every time.

Clickable directions produced clearer decisions than static screens.

01 Foundations

Color, type and the base components every workflow inherits.

Six color ramps, a Figtree type scale with Source Serif 4 reserved for AI-written text, 4pt spacing and an eight-step radius scale. The serif gives the AI a distinct voice, so it reads as a tool, not a person.

AgentOlympus design systemFigtree + Source Serif 4 · 6 ramps · 4pt spacing

Color

Jungle GreenBrand primary#005A64

Forest GreenBrand secondary#3C9673

Lime GreenBrand accent#BECE33

Space CrayolaNeutral#294145

SinopiaError#C73F1E

MarigoldWarning#FFA62B

Type

StyleWeightSizeLineTracking H1Medium5003240-1 H2SemiBold6002432-0.5 H3SemiBold60020240 H4SemiBold60016240 Body LRegular40016240 Body MRegular40014200 Body SRegular40012160 AI bodySource Serif 440014200

Components

Buttons

Labels

NeutralBrandSuccessWarningErrorInfo

Corner radius

NoneXSSMLXL2XLRound

Spacing · 4 to 80

02 Shared patterns

Two shared patterns carry most of the platform.

Conversation Scoped chat, a source panel, and citations that open into the evidence.

OncoAthena chat with a SharePoint source tree beside it

OncoAthena. The source panel is a document tree.

MR Echo citation panel with an interview transcript and the cited passage highlighted

MR Echo. The same panel holds a transcript and the cited moment.

Assessment Filter chips over a comparison table, with the same row, header and expand behavior everywhere.

TPP Builder standard of care and emerging benchmark tables

TPP Builder. Standard of care against emerging benchmarks.

Landscape Assessment price benchmarks across USA, Germany and UK

Landscape Assessment. The same table, pricing across geographies.

03 Workflow configuration

Room to differ.

Workflows can depart from these patterns without leaving the system. SlideGenie uses the same foundations for a workflow that starts with a prompt and ends in a deck.

SlideGenie generating a presentation from a prompt

SlideGenie. Starts from a prompt, ends in a deck on BioNTech templates.

Landscape Assessment competitive timeline of approvals, market entry and exclusivity

Landscape Assessment. The same system stretched to a timeline of approvals and exclusivity.

06 · How I delivered

One designer, four workflows in parallel

The first two pilots had three backend engineers and no front-end engineer, so I built coded prototypes with AI tools, and engineering connected them to the backend. The team has since grown to eight backend AI engineers and a front-end engineer, and I still deliver every workflow as a coded prototype, now to front-end. It moved delivery from one workflow at a time to four in parallel, and cut launch time from about 8 weeks to 3.

The coded prototype, live. Built with AI tools and handed to engineering, which connected it to the backend. Switch screens with the bar at the top and click through them; the data shown is illustrative.

Coded prototype

Each workflow is built as a working prototype with AI tools, on the shared design system.

Engineering

Takes the prototype as the spec and connects it to the backend, four workflows at a time.

Two-day workshop

Before launch, with SMEs, business teams and leadership on the working product: what works, what breaks, and where the AI's answers needed to change.

Prioritizing as the only designer

I weigh urgency, impact and reusability, with extra weight on reusability, since 60+ future workflows inherit what I build. When Congress AI had to be ready for a mid-September conference, I set out the trade-off with leadership and paused design on three other workflows. It shipped on time, and its review queue is now a reusable pattern.

07 · Where it stands

From pilot to platform

In about 10 months, AgentOlympus grew from 2 workflows and 50 oncology specialists to 8 workflows and 200+ users across five teams, with Congress AI running live during medical conferences.

2→8workflows

50→200+users

5business teams

~10months, pilot to platform

08 · Reflection

Before the screen, and after it.

What I learned

In expert domains, the hardest design work happens before the screen: deciding what shape the data arrives in, and where a human checks the AI. Designing AI is more than a chatbot. The work is deciding what the AI should handle and keeping enough expert control for the result to stay trustworthy, and a platform becomes one product when shared patterns support very different work.

What I'd do differently

I'd agree on the design language before the first pilot. It took longer than any single screen, and the early workflows had to be brought in line afterwards.

Next project Breaking Money Silos