Horizontal AI Platform Cuts Agent Cost per Active User by 60% and Extends AI Agents from 30,000 to 150,000 Seats with OpenAI on Amazon Bedrock
At a glance
A horizontal AI platform sells a suite of enterprise AI agents, but at about $2.20 per active user per month in OpenAI costs on Azure, finance allowed agents only in the Pro tier. Agents reached 30,000 of 150,000 seats, and large buyers wanted them for every employee.
New Math Data moved the platform’s inference to Amazon Bedrock in five weeks with the same GPT models, matched every Foundry content filter in Bedrock Guardrails, and built evaluation-backed routing that sends each agent task to the lowest-cost GPT-5.6 model that passes. Cost per active user fell to $0.88, under finance’s $1.00 ceiling.
AI agents now ship in every seat, so the platform competes for three enterprise deals that require agents for all employees. It chose AWS over a Microsoft discount, and its model spend grows from about $66K to about $132K a month on Bedrock.
Industry
Use Case
AI Agent Development, Cost Optimization, AI Platform Migration & Implementation, Model Evaluation & Benchmarking
Solution implemented
- OpenAI GPT models on Amazon Bedrock replace Azure endpoints and deployment names behind the platform's existing model gateway.
- Evaluation-backed routing sends planning to GPT-5.6 Luna and generation to GPT-5.6 Sol, cutting cost per task about 60%.
- Bedrock Guardrails match every Foundry content filter, proven on a known-good and known-bad test set.
- Shadow mode in the platform's account mirrors live agent traffic to Bedrock before any customer traffic moves.
- Overnight batch jobs move to GPT-5.6 Luna at standard rates, and embeddings move to a Bedrock-hosted model.
- Every call is tagged by tenant and agent for cost tracking, and Terraform deploys through existing pipelines.
The value equation
- AI agents extend from 30,000 Pro seats to all 150,000 active users, now included in every seat.
- Cost per active user falls from $2.20 to $0.88, below finance's $1.00 ceiling for every seat.
- At 120,000 active users on average, that reduction projects to $1.90M in AI cost avoided in year one.
- Three enterprise deals that require agents for every employee are now open to the platform.
- The platform chose AWS over a Microsoft discount, and model spend now counts toward its AWS commitment.
Company Snapshot
A horizontal AI platform sells a suite of 100+ enterprise AI agents to business teams, and now includes agents in every seat across 150,000 active users at under a dollar per user per month.
Location
United States
Customer Situation
A horizontal AI platform sells a suite of enterprise AI agents to business teams. Its OpenAI models ran on Azure AI Foundry, pay-as-you-go with no Azure commitment, at about $2.20 per active user per month. Finance would include agents in every seat only below $1.00, so agents stayed in the Pro tier and reached 30,000 of 150,000 seats.
Large buyers wanted agents for every employee, and three enterprise deals required it. The platform’s own infrastructure already ran on AWS, but Microsoft owned the OpenAI relationship and offered a discount and credits to keep the workload. No one had tested whether every agent task needed the model it used.
Many AI platforms price agents as an add-on because model cost per user outruns what a base seat can carry, which caps adoption.
NMD Solution
NMD reviewed and found that routing each agent task to the lowest-cost GPT-5.6 model that passes evaluation, plus partner funding and commitment burn-down, outweighed Microsoft’s discount. The solution required a provider layer behind the existing gateway, every Foundry filter mapped to Bedrock Guardrails, and shadow mode and evaluation inside the platform’s account so NMD sees aggregates only. Evaluation showed GPT-5.6 Luna handled planning at a fraction of GPT-5.6 Sol’s cost, which drove most of the savings. NMD also closed the batch, embeddings, and zero data retention gaps, and one platform engineer reviewed pull requests about four hours a week, so the launch stayed on schedule.
What We Delivered
The platform’s agents run on OpenAI on Amazon Bedrock in production, five weeks after kickoff. Cost per active user is $0.88, down from $2.20 and under finance’s $1.00 ceiling, so agents ship in every seat and reach 150,000 active users, up from 30,000. At an average of 120,000 active users, that projects to $1.90M in AI cost avoided in year one. Every content filter is proven, and three enterprise deals that require agents for all employees are now open. The platform works with OpenAI applied AI engineers directly. Model spend counts toward the AWS commitment, and the platform holds the runbook, code, Terraform, and evaluation tooling.
Ready to Transform Your Technology Business?
See how New Math Data can transform your technology business with AWS-powered innovation.