Enterprise GenAI Implementation Strategies Across Industries

astrazeneca_/adobe/_allianz_technology · Other · year n/a · source Agentic: no · promotional bias: low · source-backed · extracted with prompt 0.5.6 (schema v2)

Generated from extraction.json — do not edit by hand. Every claim cites verbatim source evidence; stated means the quote directly asserts the claim, implied means it supports it indirectly.

Problem

Domain: Cross-industry enterprise GenAI rollout

Executives from AstraZeneca, Adobe, and Allianz discuss how to move GenAI from pilots to production across research, creative/legal, and code-modernization contexts while managing legal review, culture change, use-case prioritization, model optionality, and scaling decisions.

Constraints:

Legal, work councils, unions, and country-specific approvals can block or accelerate AI rollout.
A successful pilot can create sudden demand before the system is technically or legally ready to scale.
GenAI adoption requires cultural change rather than only tool rollout.
Different use cases need different models, so architecture should avoid hard model lock-in.
Teams need fast experimentation with willingness to scale successes and shut down misses.

Architecture

The panel's shared operating model is to engage legal early, select high-value use cases, run pilots, prepare platform approvals for multiple countries or use cases, share successes through internal culture mechanisms, and architect model-agnostic systems so different models can be swapped for claims, image recognition, handwriting, research, or code modernization.

Diagram confidence: reconstructed (reconstructed from prose).

flowchart LR
  use-case-intake["Use case intake"] --> legal-and-governance-review["Legal and governance review"]
  legal-and-governance-review --> pilot-group["Pilot group"]
  pilot-group --> scaling-decision["Scaling decision"]
  scaling-decision --> model-agnostic-platform["Model-agnostic platform"]
  model-agnostic-platform --> culture-and-adoption-loop["Culture and adoption loop"]

Drag to pan · scroll or use +/− to zoom · reconstructed from the source's described component topology

Components

Component	Role	Interacts with
`Use case intake`	Identifies research, legal/creative, code modernization, and business-unit GenAI opportunities.	`Legal and governance review`
`Legal and governance review`	Handles copyright, licensed content, country approvals, work councils, unions, and platform approvals.	`Pilot group`
`Pilot group`	Initial users who test tools such as AstraZeneca's research assistant.	`Scaling decision`
`Scaling decision`	Determines whether to scale fast, iterate, or fold down based on pilot success.	`Model-agnostic platform`
`Model-agnostic platform`	Allows different models for image recognition, handwriting, claims data, research, or code modernization without changing the whole use case.	`Culture and adoption loop`
`Culture and adoption loop`	Uses stories, hacks, and peer curiosity to spread GenAI practices.	—

Data flow

A use case enters legal and platform review early, runs as a pilot with a focused user group, gathers usage and success signals, scales quickly if it solves real problems, and feeds stories back into a culture of GenAI hacks or adoption practices; model-agnostic architecture lets different models serve different tasks.

“Luckily we had to engaged legal early and we actually could, within a week, one or two weeks, we could scale it to the main countries”transcript, stated

“We also try to be as model agnostic as possible.”transcript, implied

Why this architecture

The panel treats enterprise GenAI as an operating-system change: legal and governance must be involved early, cultural adoption must be cultivated, and model choices must remain flexible because use cases and regulations vary across domains and countries.

“using Gen AI is not just rolling out a tool. I think many organizations make the mistake that they believe”transcript, stated

“you don't become a DevOps organization by just using Jenkins. So it's the same thing with AI. The tool will not do the trick.”transcript, stated

Models used: AstraZeneca research assistant · Adobe Firefly · Amazon Q · GenAI · LLMs · Model-agnostic architecture

Scale

AstraZeneca's research assistant began with 40 pilot users.
“I'm one of the pilot users, one of 40 pilot users.”transcript, stated
AstraZeneca scaled the research assistant within one or two weeks to main countries where legal and labor stakeholders had been engaged.
“within a week, one or two weeks, we could scale it to the main countries where we already had engaged with legal”transcript, stated
The research assistant reached thousands of users overnight after a silent launch.
“Overnight, we had thousands of users starting using this tool.”transcript, stated
One pilot user reported saving weeks or probably months on each literature search.
“tool that saves me weeks, probably months of my time, every time I have to do a literature search.”transcript, stated

Alternatives considered (roads not taken)

Treat GenAI as a normal tool rollout.

Rejected because: Allianz says adoption requires changing how developers work, not just giving them a tool.

“using Gen AI is not just rolling out a tool.”transcript, stated

Delay legal until after technical development.

Rejected because: Panelists say early legal engagement accelerates scaling and avoids approval blockers.

“Legal's an enabler for speed. We are actually not a speed bump at all.”transcript, stated

Use one model for every use case.

Rejected because: Allianz says picture recognition, handwriting recognition, and claims data may need different models.

“models for different use cases. If you have something where you more or less need picture, scanning recognition”transcript, stated

Train generative image models on broad unlicensed content.

Rejected because: Adobe took a conservative licensed-content position for commercial safety and goodwill.

“we will only train on licensed content and what that then tumbleweeded into was like, "Well, wait a second. That means that our model is commercially safe”transcript, stated

Patterns

`safety-policy-enforcement` (governance)

The panel emphasizes early legal, work-council, union, and country-specific approvals before broad scaling.

“engage early with legal, thinking about the global blockers that are out there.”transcript, stated

`model-routing` (model-ops)

Allianz designs for swapping models by use case without changing the entire application.

“we try to architect in a way that we can change the model without impacting the entire use case.”transcript, stated

`canary-rollout` (deployment)

AstraZeneca used a pilot group, then scaled quickly once legal and stakeholder approvals were in place.

“It's a pilot, it's going to take nine months to build it ready.”transcript, stated

`human-in-the-loop-gate` (adoption)

The organizations rely on scientists, developers, legal teams, and culture loops to decide which tools scale.

“culture of Gen AI hacks where if somebody has a hack, he talks about it, then other people become curious.”transcript, stated

Tradeoffs

Take a conservative legal position on training data.

Chose Licensed content for Firefly. over Training on broader legally uncertain content..

Rationale: Adobe connected licensed training to commercial safety, indemnification, and creative-community goodwill.

“That means that our model is commercially safe and we can indemnify this now.”transcript, stated

Scale or shut down based on pilot evidence.

Chose Move fast when successful and fold down quickly when not. over Continuing every pilot indefinitely..

Rationale: The panel recommends doubling down on success while cutting misses.

“if it is a success, then scale fast. But also as we talked before, if it's not a success, fold it down very quickly.”transcript, stated

Use model-agnostic architecture.

Chose Swappable models by task. over A single model embedded into all use cases..

Rationale: Different tasks require different model strengths and future model choices are uncertain.

“we can change the model without impacting the entire use case.”transcript, stated

Gotchas (what bit them)

Successful pilots can outrun readiness.

Root cause: A tool that solves a real problem may create viral demand before the team considers it scalable.

Resolution: Engage legal and platform stakeholders early enough to scale fast when a pilot works.

“No, no, no, no. This is a pilot. We have 40 users. It's not scalable right now.”transcript, stated

Legal can be miscast as a blocker.

Root cause: Teams often bring legal in late after technical assumptions are already made.

Resolution: Engage legal early and treat legal review as an accelerator for speed.

“Legal's an enabler for speed. We are actually not a speed bump at all.”transcript, stated

Tool rollout alone does not change work.

Root cause: Executives may assume installing a tool is equivalent to adoption.

Resolution: Work with users to change workflows, remove disliked tasks, and spread success stories.

“The tool will not do the trick. So what will do the trick is that you need to fundamentally change the way how we work”transcript, stated

Evaluation

The panel evaluates GenAI rollout through pilot adoption, user stories, speed of scaling, legal readiness, cultural uptake, and whether experiments succeed or should be shut down.

Metrics: 40 pilot users · One to two week scale-up · Thousands of users overnight · Weeks/months saved per literature search · Developer adoption of Amazon Q/code modernization

“Overnight, we had thousands of users starting using this tool.”transcript, stated

“going and inspecting things that did not work.”transcript, stated

Outcomes

AstraZeneca's research assistant went viral from a 40-user pilot to thousands of users overnight.
Adobe's licensed-content policy supported commercial safety and indemnification for Firefly.
Allianz frames Amazon Q/code modernization as a culture and workflow change, not only a tool rollout.
The panel converges on early legal engagement, fast scaling of successes, and model-agnostic architecture.

The strongest quantified scale comes from AstraZeneca's 40 pilot users, one-to-two-week country scale-up, and thousands of overnight users.

“it solved real problems for scientists. It democratized the knowledge and that was really our first big success.”transcript, stated