→ Engagement spotlight / Multi-agent systems
A multinational insurer's AI Lab read and ranked every submission by hand. The platform that replaced it now carries 200 active use cases
Each cohort brought more than 150 submissions from 24 markets, arriving as free text of wildly varying quality. Reading a single one took an hour or two, and the lab ran two cohorts a year.
Every one of them sits in a system that can compare it against the others and against what the business already runs. Operating the whole platform costs about $500 a year.
→ The sequence / what the pipeline does
Five stages, and a person at every gate.
01
Structure the input
An agent turns free text into a structured problem statement, asking the clarifying questions the central team used to ask.
02
Make duplication visible
Semantic clustering against every idea submitted and every system already running, so one problem in three vocabularies surfaces once.
03
Score, then stop
Consistent evaluation across strategic alignment, feasibility and likely return, producing a ranked shortlist for a person to act on.
04
Build something to look at
A code-generation agent produces a working prototype on its own isolated virtual machine, so the blast radius is contained by construction.
05
Hand it to the lab
The AI Lab runs the platform. MISSION+ holds no operational role in it.
The pipeline in full / mint marks a human decision
↻ Gate No agent output reaches a funding decision without a named person between it and the decision log.
→ The situation
The same idea was being funded more than once.
The insurer's AI Lab exists to turn employee insight into deployed AI, and by its own measure it was working. Each cohort brought more than 150 submissions from 24 markets, and the lab ran two cohorts a year. They arrived as free text of wildly varying quality, from people with wildly varying AI literacy.
A small central team read all of them. Someone sat with a spreadsheet, interpreted each submission, worked out what the person actually meant, and ranked it against everything else. At an hour or two per submission, the reading alone consumed months of a five-person team's year before a single decision got made.
Two things followed. Ideas that were the same idea, submitted in three countries in three vocabularies, were never recognised as the same idea, so the same research got funded more than once. And because evaluation capacity was fixed while submission volume was not, how far the programme could expand across markets was set by how fast five people could read.
The platform built here became REX. It was designed for this insurer, then packaged as a MISSION+ product and taken to other clients.
The rule the platform is built on
Every ranking the platform produces has a named person standing between it and the funding decision.
→ What was measured
The numbers, and where they came from.
Annual running cost of the platform, against the cost of generating a prototype
US dollars
PlatformPer year, all environments
One prototypeGenerated on an isolated VM
Active use cases
Moved off spreadsheets and onto the platform, where they can be compared against each other and against systems already running.
New environment
A new client environment stands up in three days. That figure is measured on a later deployment rather than estimated from the first one.
Annual running cost
The whole platform, plus about $50 for each prototype it generates. At that price the cheapest way to evaluate an idea is to build a working version of it.
The platform runs on prompt engineering and retrieval alone, with the whole training apparatus left out of the design. On this problem that cost nothing in quality, and it is what holds the figure at $500 a year.
Built on first-party SDKs, with none of the popular orchestration frameworks anywhere in the dependency tree. Dropping that layer removed a class of dependency, audit and upgrade risk that regulated clients care about.
AWS through Bedrock, ECS Fargate, Aurora, SQS and OpenSearch Serverless. GCP through Vertex AI, Cloud Run, Cloud SQL and Pub/Sub. Both first class, selected per environment from configuration, with no forks. That abstraction is why an environment takes three days.
It only finds duplication that has been submitted. Work already running elsewhere in the business that never entered the pipeline stays invisible to it, and in an organisation this size that gap is large enough to matter.
→ Why it matters
The programme is no longer limited by how much a small team can read.
How far the programme could reach used to be set by how fast five people could read. That constraint is gone, so the lab can widen the funnel across markets without adding headcount to the middle of it. A leader now receives the shortlist with working prototypes attached to it.
The judgement at each gate still belongs to a person, which is a design decision the platform was built around. The platform is built so that no agent output reaches a funding decision without a named human between it and the decision log, which is the property that made it acceptable inside a regulated insurer.