Building a Scenario Took Weeks. Now It Takes Minutes.
ScenarioAI sells decision support to defense and government organizations. The work that made their product valuable, building credible scenarios from scratch, was also the slowest part of it.
Tiber built the platform they ship as IDEA: retrieval-augmented generation running Claude models on Amazon Bedrock inside ScenarioAI's own AWS account, where the scenario work their clients pay for now happens through conversation rather than by hand.
The outcome
80%+
Reduction in time and cost against the manual methods IDEA replaces.
According to ScenarioAI customer feedback
- What we built
- IDEA, a retrieval-augmented generation platform for building scenarios.
- Where it runs
- Claude models on Amazon Bedrock, inside ScenarioAI's own AWS account.
- What it ships as
- A subscription ScenarioAI sells to their defense and government clients.
The Challenge
Building scenarios is slow by nature. It was also the bottleneck in their business.
ScenarioAI helps defense and government organizations plan against complex, fast-moving situations. Their product depends on scenarios credible enough to make real decisions on: force structures, terrain, political context, and second and third order effects traced through systems that do not sit still.
Building one by hand took weeks. Most of that time went into assembling the scenario rather than into the analysis the scenario exists to support. The analysis is what clients pay for. It was also the part nobody could reach quickly.
Four constraints shaped the build
The last constraint drove the rest. ScenarioAI was not building an internal tool. They were building something defense clients would buy, which made the data boundary a commercial precondition rather than a preference. Public AI interfaces were never an option.
The Solution
IDEA: expert data, grounded generation, and Claude where the data already lives
Tiber built IDEA, the platform ScenarioAI sells today. It takes the scenario work that used to be done by hand and turns it into something an expert directs through conversation. Five parts make that possible.
ScenarioAI's own AWS account
Purpose-built agents, one per theater or task
Composed by ScenarioAI's own experts, each with its own data scope and operating instructions. No engineering ticket required.
Generated work product
- Road to War narratives
- Orders of Battle
- Country books
- Capability cards
- Intelligence reports
Connectors across structured and unstructured data
Scenario data does not arrive in one shape. Order of battle tables, terrain and ice coverage data, doctrine documents, country reporting and open source material all had to land in one place. We built custom connectors that pull structured records and unstructured documents into a single corpus, with provenance preserved so every output can be traced back to the source it came from.
A retrieval layer over expert-curated data
On top of that corpus we built retrieval-augmented generation (RAG), a method that grounds what the model writes in specific, verified source material instead of in its training data. ScenarioAI subject matter experts curate the underlying PMESII-PT dataset. That framework covers eight dimensions of an operating picture: political, military, economic, social, infrastructure, information, physical environment and time. Grounding generation in curated material is what makes the output defensible.
Claude models on Amazon Bedrock
Claude models run on Amazon Bedrock inside ScenarioAI's own AWS account. Analysts build scenarios through chat, working conversationally against the grounded data, and the same models generate the documents the work product actually consists of: Road to War narratives, Orders of Battle (ORBAT), country books, capability cards and intelligence reports. Bedrock decided the deployment question. It gave ScenarioAI the reasoning depth of a frontier model without a single record leaving infrastructure they control.
Purpose-built agents for each theater
A scenario in the Arctic does not ask the same questions as one in the South China Sea. Rather than ship one general assistant, we built an agent creation experience that lets ScenarioAI compose purpose-built agents for a given theater, domain or task, each with its own data scope and operating instructions. Their experts configure them directly. No engineering ticket required.
Delivered to their clients as a subscription
IDEA is in production today, shipping to ScenarioAI clients as a software-as-a-service (SaaS) subscription, with tenant isolation, the same contained data boundary applied per client, and integration into the command and control (C2) and simulation platforms those clients already run, including Systematic SitaWare HQ and MAK VR Forces.
What Changed
The slowest work moved off the critical path.
IDEA compresses the parts of scenario work that consumed the most time: initial build, gap identification, and iteration. What changed was not only speed. It was where expert attention could go. Analysts stopped spending their best hours assembling baselines and started spending them stress-testing assumptions and exploring edge cases that were previously too slow to reach.
Before
With IDEA
Before
Weeks of expert effort to build a single credible scenario
With IDEA
Initial scenarios generated in minutes and hours
Before
Manual gap analysis requiring full team review cycles
With IDEA
Logical gaps surfaced automatically during generation
Before
Flat, linear scenario narratives with limited branching
With IDEA
Branching futures and alternative pathways explored on demand
Before
Source material scattered across systems and formats
With IDEA
One grounded corpus, with provenance on every output
The Outcome
IDEA was not built to replace their experts. It was built to amplify them.
Analysts
Freed from baseline drafting. More time on assumption challenges and edge case exploration.
Designers
Freed from formatting and structural scaffolding. More time on strategic depth and narrative refinement.
The Company
A platform to sell, not just a faster internal process. Shorter iteration cycles and meaningfully greater output without a larger headcount.
IDEA runs in production today. It is not a pilot or a proof of concept. ScenarioAI sells it to their clients as a subscription, and those clients build scenarios on it as part of their working practice.
The Data Boundary
The boundary is the reason the product could be sold at all.
ScenarioAI operates in environments where trust is not a feature. It is the prerequisite for being in the room. Clients share operational constructs, planning assumptions and institutional knowledge, and they expect every piece of it to stay contained.
Running Claude models on Amazon Bedrock inside ScenarioAI's own AWS account let them give that assurance without caveats. It is the same pattern Tiber packages as Caddie, our secure retrieval-augmented generation deployment on AWS. Data does not leave the boundary, which is what made realistic problems, and therefore useful output, possible.
Looking Ahead
Built once, extended since.
ScenarioAI continues to build on the same foundation. Structured reasoning frameworks, decision trees and multi-step simulation workflows are layered on top of the platform, each one possible because the data and deployment decisions were made correctly at the start.
The lesson here is not complicated. AI becomes useful to a business when it is grounded in that business's own verified data, deployed where the data is allowed to stay, and built into the workflow people already have. Those three conditions do the work. A model is only ever as good as what sits underneath it.
That is the part most AI projects skip. It is also the part that decides whether the result is a demo or a product.
Ready to Build on a Trusted AI Foundation?
Tell us what you are building. We will map a practical delivery path from your current state to production impact.