Executive intelligence and an AI workforce for a multi-site urgent care network
A network of urgent care clinics was being run from reports that were accurate the day they were assembled and stale by the time they were read. Catalyze built the layer underneath: one governed definition of every operational metric, an executive intelligence surface that writes the morning brief and answers questions with its sources attached, and a workforce of agents that clears the repetitive work behind the front desk.
Client
A multi-site urgent care networkAnonymized at the client’s requestContext
Urgent care runs on throughput, and throughput is invisible between reports.
An urgent care network lives or dies on operational detail: how long patients wait, how long a visit takes, how many leave without being seen, whether care protocols are actually being followed, whether the front desk captured the information that determines whether the visit gets paid for. All of it is measurable and almost none of it was visible in time to act on. The data existed across the practice-management system, the clinical record, the document pipeline and the phone system — but assembling it into a picture of the network was a manual job, and a manual job only happens on a schedule.
Challenge
You cannot run a network of clinics from a report that is already three weeks old.
The problem was not a lack of data or a lack of effort. It was that every question worth asking required a person to go and assemble the answer, which meant that only the questions important enough to justify that effort ever got asked. Leadership had recurring decks. What it did not have was the ability to ask something new on a Tuesday and get an answer on Tuesday. Meanwhile the highest-volume operational work in the network was being done by hand, one item at a time.
- Operational data split across practice management, clinical records, document intake and telephony
- Network and per-clinic KPIs assembled by hand into recurring decks
- No single agreed definition of a metric, so two reports could legitimately disagree about the same week
- Questions that came up between reporting cycles simply went unanswered
- Compliance exposure — overdue training, care-protocol adherence — visible only when someone went looking
- High-volume repetitive work, from inbound fax triage to duplicate patient records to benefit activation, consuming clinical and revenue-cycle staff
Approach
The semantic layer came before the AI, on purpose.
The fastest way to destroy trust in a system like this is to let an executive ask a question, get a confident answer, and discover the number disagrees with the report they read yesterday. So the sequence was deliberate: consolidate the data, then agree the definition of every metric that matters and encode it once, and only then put a language interface in front of it. The corollary is that every answer the system produces has to show its work — what it queried, how fresh the data was, how many rows it evaluated, and what could make the answer wrong. Provenance is not a feature here; it is the reason anyone believes the output.
System
What was built
Two surfaces on one governed foundation. Leadership gets answers; the floor gets work taken off its plate.
01
Daily executive brief
Generated every morning: network KPIs month-to-date and week-to-date, a written narrative that connects them into an actual read of the day, and per-clinic spotlights measured against the prior comparable period rather than an arbitrary target.
02
Ranked recommended actions
The brief ends with a prioritized list rather than a summary. Each action carries the evidence that produced it, a priority, a named owner function and a timeframe, so the output of the analysis is a decision rather than a chart.
03
Natural-language analysis
An operator asks a question in plain language and gets a dossier: a direct answer, the result set, a chart chosen to fit the shape of the data, the observations worth noting, and suggested follow-up questions.
04
Provenance and uncertainty as output
Every analysis states its query source, data freshness, rows evaluated, generation time and latency — and names the filters and exclusions that could change the conclusion. The system is built to disclose what it does not know.
05
Saved analysis
Any analysis can be named, tagged and re-run against current data. A recurring question becomes a standing instrument instead of a repeated request to an analyst.
06
AI workforce
A directory of named agents, each scoped to one repetitive workflow, each reporting what it completed and how much time it returned — measured against a stated baseline assumption rather than an unexplained number.
07
Governed access
Role-based control over who can query what, with assistant conversations retained and auditable. In a clinical setting an AI interface without an audit trail is not deployable.
Architecture
The system
One foundation, forking into two products. The components named at each layer are what the system actually consists of.
- 01
Source Systems
The operational record, already spread across four categories of system.
- Practice management / EMR
- Clinical documentation
- Inbound fax + document intake
- Telephony
- Object storage
- 02
Data Platform
Consolidation into one warehouse, with validation at the boundary rather than downstream.
- Snowflake
- Automated ingestion
- Document pipeline
- Data-quality checks
- Freshness tracking
- 03
Semantic Layer
One definition per metric, agreed before anything was put in front of an executive. This is the layer that makes the rest of it trustworthy.
- Governed metric definitions
- Clinic + region hierarchy
- Comparison windows
- Domain boundaries
- Threshold + target registry
- 04
AI / Model Layer
Language over governed metrics — never over raw tables.
- Text-to-SQL against the semantic layer
- Narrative synthesis
- Chart selection
- Threshold + anomaly detection
- Caveat generation
- 05
Product Surfaces
splits into 2Where the system stops being infrastructure and becomes something a person opens. It forks here because leadership and the floor need different things from the same foundation.
Executive Intelligence
Answers the questions leadership asks — on a schedule, and on demand.
- Daily executive brief
- KPI rollups (MTD / WTD)
- Per-clinic spotlights
- Ranked recommended actions
- Question-to-dossier
- Saved analysis library
AI Workforce
Clears the repetitive work that was consuming clinical and revenue-cycle staff.
- Fax classification + routing
- Duplicate record resolution
- Benefit activation
- Executive synthesis
- Throughput telemetry
- 06
Governance and Trust
The layer that makes an AI interface deployable in a clinical operation rather than a demo.
- Role-based query permissions
- Source + freshness on every answer
- Rows-evaluated disclosure
- Auditable assistant conversations
- Stated baseline assumptions
- 07
Operators
The people the system exists to give time and judgement back to.
- Executive team
- Regional managers
- Clinic leads
- Revenue cycle
- Compliance
Implementation
How it was built
Five phases, sequenced so that each one had to earn trust before the next was allowed to depend on it.
Technologies
Data
- Snowflake
- Automated ingestion
- Object storage
- Semantic layer
AI layer
- Text-to-SQL
- Narrative synthesis
- Anomaly detection
- Chart selection
Workflow
- Agent orchestration
- Document classification
- Record deduplication
- Eligibility + benefits
Delivery
- Web application
- Scheduled brief generation
- Role-based access
- Audit logging
- 01
Consolidate
Practice management, clinical, document and telephony data brought into one governed warehouse with automated ingestion and validation at the boundary.
- 02
Define
Every operational metric given one agreed definition, hierarchy and comparison window. Disagreements about what a number meant were resolved here rather than in a meeting six months later.
- 03
Brief
The daily executive brief shipped first — it was the artefact leadership already wanted, which made it the fastest honest test of whether the data underneath was right.
- 04
Open the questioning
Natural-language analysis added once the metric layer was trusted, with provenance, caveats and follow-ups built in from the first version rather than retrofitted.
- 05
Deploy the workforce
Agents pointed at the highest-volume repetitive workflows, each one scoped narrowly, named, and measured on tasks completed and time returned.
Outcome
What the system changed
The platform is in production across the network and generates a brief every morning. The structural change is that the reporting cycle stopped being the limit on what leadership could know: a question that occurs to an executive is now answered in the same session, with its sources attached, and the repetitive work behind the front desk is handled by named agents that report what they did. These are the measures the system is built to move.
- Time to answer — from a request and a reporting cycle to a question and a session
- One agreed definition per metric, so reports stop disagreeing with each other
- Earlier visibility of wait time, length of stay and left-without-being-seen movement
- Compliance exposure surfaced continuously rather than on discovery
- Clinical and revenue-cycle hours returned from repetitive processing to patient-facing work
- Decisions that arrive with their evidence, their caveats and an owner attached
Results
This engagement is covered by client confidentiality, and the operational figures the platform reports are the client's own. We do not publish them. We share measured results — throughput, time to answer, hours returned by workflow — directly, under NDA.
Capabilities involved
What this system is made of.
Related insights
Our thinking on this.
Looking for something closer to your own operation?
All workTell us about the workflow. We'll tell you whether we have built something like it.