Read summarized version with
About project
Working time:
2024 – ongoing
Industry:
HealthTech
The service:
Agentic AI, Cloud Security, Managed Services
Tags:
HealthTech / Agentic AI / Generative AI / Cloud Security / Managed Services
Overview
Spike Technologies, Inc. is a New York healthcare technology company. Its product, Health OPS, is an agentic AI voice and workflow automation platform for outpatient rehabilitation clinics: physical therapy, occupational therapy and speech-language pathology. Its agents call payers and patients, verify eligibility, obtain and renew prior authorizations, follow up on claims and run patient outreach, writing every outcome back into the systems the clinic already uses.
A rehabilitation clinic’s revenue cycle runs on the telephone. Eligibility has to be verified before the first visit, and prior authorization has to be obtained before treatment and renewed as the authorised visit count is consumed, because a plan of care is a sequence of visits and each has to be covered. Industry figures for these practices put front-desk time at more than 10 minutes per patient on prior authorization alone, against an average claim denial rate of about 13 per cent, with nearly three-quarters of denials requiring an appeal.
Dedicatted built the platform on Amazon Bedrock AgentCore inside a HIPAA-eligible boundary in us-east-1, and has operated it under a managed DevOps service since it went live on 8 December 2025. Eleven clinics were live by June 2026, at about 9,500 calls a month.
The Challenge
The expensive part of the work is a conversation. Scripted IVR navigation and portal scraping retrieve a fraction of eligibility answers and break whenever a payer changes a page. The costly work is the call with a payer representative about one patient’s authorization: IVR navigation, hold time, a conversation that follows no script, and then re-keying the answer into the EMR.
A product company with no platform. Spike was a venture-backed startup with the domain knowledge and the product thesis, a small engineering team, and no cloud platform or security function. It needed to run many concurrent telephone conversations carrying protected health information and to orchestrate calls into EMR, billing and payer systems safely, without consuming the engineering capacity the product itself needed.
Protected health information cannot leave the boundary. Each clinic is a HIPAA covered entity, and Spike is its business associate. Every clinic’s compliance officer reads the data-flow diagram before signing, so protected health information had to stay inside HIPAA-eligible services under a business associate agreement, verifiable from the architecture rather than promised.
A confident wrong answer is worse than no answer. When an agent tells a patient how many visits remain or what the copay is, the clinic acts on it. A fluent completion of a payer’s response that is not in it is a denial waiting to happen, so nothing the agent says or records could originate in the model.
Our Approach
- Reuse the landing zone, and sign the paperwork before the first call. The platform was built into the AWS Control Tower landing zone Dedicatted had delivered for Spike in 2024, with its management, log archive and audit accounts and preventive service control policies. The AWS Business Associate Addendum was accepted in AWS Artifact at the Organization level on 20 November 2025, before any protected health information entered the account.
- Keep every component that touches PHI inside HIPAA-eligible services. Telephony, speech, models, agent runtime, state and storage were each selected on eligibility as well as capability, so the whole data-flow diagram sits inside one business associate agreement, one audit record and one set of keys.
- Take the agent platform as a managed service instead of building one. Session isolation between clinics, memory with a retention policy, a credential vault, a tool registry and tracing are what Amazon Bedrock AgentCore provides, and they are exactly what a small team should not own.
- Set a turn-latency budget in milliseconds and design against it. A voice agent that pauses gets interrupted or hung up on. Two models, streaming on every hop and private network paths were chosen against a budget of 2 seconds at the 95th percentile.
- Ground every fact the agent states or records. Every statement the agent makes or records comes from a tool result or a transcript, is grounding-checked, and is read back to the payer representative before the call ends.
Our Solution
Telephony and speech. Calls enter and leave through Amazon Connect, which provides contact flows, telephone numbers and carrier connectivity and encrypts the call recordings with a customer-managed AWS KMS key. Amazon Transcribe performs streaming speech recognition with content redaction for PII and PHI, so the stored transcript carries redaction tokens where identifiers were spoken, and Amazon Polly neural voices synthesise the agent’s side of the conversation.
Agents. Each call or workflow task is an Amazon Bedrock AgentCore Runtime session: isolated per session, so one clinic’s call cannot share state with another’s, and long-lived enough to survive a payer’s hold. AgentCore Memory holds in-call and cross-call state under a retention policy. AgentCore Gateway exposes AWS Lambda tools under one schema, for the clinic EMR API, the clearinghouse eligibility API with its X12 270/271 transactions, and the scheduling system, logging every tool call with its arguments and result. Prior authorization requests go to the clearinghouse as X12 278 transactions where the payer accepts them, and by voice where it does not. AgentCore Identity holds the OAuth credentials for those APIs, so the agent never sees a raw secret. AgentCore Observability traces every reasoning step into Amazon CloudWatch, which is what makes an agent’s action auditable.
Models. Reasoning runs on Anthropic Claude models on Amazon Bedrock, with a smaller model managing the conversational turn and a larger one reasoning over the case and drafting the record, both selected by configuration. Amazon Bedrock Guardrails applies sensitive-information filters and contextual grounding checks, so a statement about coverage is checked against the payer response it was derived from. Amazon SageMaker runs the evaluation pipeline over transcripts and tool-call traces and hosts a custom call-outcome classifier that labels every completed call.
Application, state and network. The clinic-facing application and the workflow API run on Amazon ECS on AWS Fargate with Graviton, in private subnets across three Availability Zones. Live call state sits in Amazon DynamoDB, where a turn needs single-digit millisecond reads; durable case and authorization records in Amazon Aurora PostgreSQL; recordings, transcripts and generated documents in Amazon S3 under a customer-managed key per data class. Interface endpoints keep traffic to Bedrock, AgentCore and Transcribe on the AWS network, and outbound reach to EMR, clearinghouse and payer destinations is allow-listed.
Governance and operations. AWS CloudTrail with S3 data events, Amazon GuardDuty with runtime monitoring on the tasks, AWS Config with the Operational Best Practices for HIPAA Security conformance pack, AWS Security Hub, Amazon Inspector on the container images, and CloudWatch Logs data protection policies that mask PHI patterns make up the governance plane. All infrastructure is Terraform applied through the CI/CD pipeline, and Dedicatted operates the account under a managed DevOps service as a subcontractor business associate.
The Results
Baselines were measured at the first two clinics in November 2025, the month before go-live. The results are from May 2026, across all eleven clinics.
Front-desk staff time per prior authorization fell from 12 minutes to 4, against a target of under 5. The 4 minutes is the average across all authorizations and is human review of the exceptions; the routine case is completed by the agent with no staff time.
Eligibility checks completed by the agent with no human involvement rose from 0 to 78 per cent, against a target of 70 per cent, over 6,200 checks in May 2026. The remaining 22 per cent reach a person with the payer’s response already attached.
Conversational turn latency is 1.1 seconds at the 50th percentile and 1.9 seconds at the 95th, inside the budget set at design time.
The blended cost is USD 0.31 per handled call, against a pre-implementation estimate of USD 0.38. Every tier of the platform is consumption-priced, so a new clinic group adds calls rather than infrastructure.
The platform handles about 9,500 calls a month with a peak of 40 concurrent calls. On 15 July 2026 Spike Technologies was acquired by Raintree Systems, an EHR and practice management vendor for rehabilitation providers, whose stated purpose was to bring this platform to its own customer base.
Dedicatted is an AWS Premier Tier Services Partner. We build and operate agentic AI workloads that handle protected health information on AWS, inside the compliance boundary the customer’s own customers audit.
