Intent recognition accuracy - portal
Prototype ~82%
Final solution ~96%
Funded by the European Union - NextGenerationEU · Government of Romania · National Recovery and Resilience Plan
Results from an NRRP-funded project · Component C9
Research results: AI and RPA/PA for the clinical front office
MONEFY BUSINESS S.R.L. and Steady Dopetech S.R.L. tested the integration of artificial intelligence and automation into sensitive healthcare administrative workflows, with rules, traceability and human validation.
Project
“Adoption of advanced digital technologies within MONEFY BUSINESS S.R.L.”
Validated architecture Control before autonomy
- 01 InputMessage or document
- 02 InterpretationAI structures the intent
- 03 ValidationThe platform applies the rules
- 04 ExecutionRPA/PA runs the safe steps
- RUE code
- 84
- Financing contract
- 84.1.I3.C9
- Contract date
- 13.05.2025
- Programme
- NRRP · C9
Results at a glance
What was measured during piloting
The indicators below retain the approximate nature and measurement context reported in the article.
- ~96% intent recognition accuracy in the portal from approximately 82% in the initial prototype
- ~93% intent recognition accuracy over SMS from approximately 74% in the initial prototype
- 0% invalid slots reaching the patient in the final solution, following backend validation
- ~93% precision of suggestions on claim drafts flagged problems confirmed as genuine during piloting
- ~40% initial review time for an EOB reduction for documents covering multiple claims
- ~36% aggregate billing effort reduction for the measured reference volume
Experimental methodology
Hypothesis → prototype → measured testing → adjustment
The work combined a time-and-motion study of current processes with per-unit measurements obtained during prototyping and piloting. MONEFY provided operational expertise and the live test environment, while Steady Dopetech carried out the technical research and development work.
- 01
Analysis and hypotheses
Mapping real processes and defining the technological uncertainties that needed to be tested.
- 02
Prototyping and evaluation
Building functional prototypes and comparing language models and integration architectures.
- 03
Controlled testing
Measuring accuracy, escalation and edge-case behaviour against representative datasets.
- 04
Operational piloting
Running the application in a live operational environment with authorised users and progressively adjusting rules and thresholds.
Solution evolution
From prototype to operational results
The comparisons preserve the definition of each indicator: accuracy, completion without escalation and effective backend validation.
Conversational assistant
Scheduling through the portal and SMS
Intent recognition accuracy - SMS
Prototype ~74%
Final solution ~93%
Conversations without escalation - portal
Prototype ~55%
Final solution ~78%
Conversations without escalation - SMS
Prototype ~40%
Final solution ~64%
Invalid slots reaching the patient
Prototype ~7%
Final solution 0%
Billing assistance
Advisory AI for claims and EOBs
The component does not autonomously modify or submit claims. It structures information, flags problems with high confidence and prioritises cases that require follow-up, while authorised staff retain the final decision.
Precision of suggestions on claim drafts
The proportion of flagged problems confirmed as genuine during piloting.
Agreement on EOB prioritisation
Agreement with the assessment performed by a billing specialist.
Initial review time
The measured reduction for an EOB covering multiple claims.
Measured benefits
Impact on operational effort
Creating a claim
Before~4 min
After~2 min
Through automatic pre-population of mandatory fields from data already held in the platform.
Creating a copayment invoice
Before~5 min
After~2 min
Through automatic identification of the relevant EOBs and application of billing rules.
Aggregate billing effort
Before~79.50 h
After~51 h
For the reference volume, including associated verification and reconciliation activities.
Solution delivered
Two layers, one control model
AI interprets and recommends. RPA/PA executes standardised steps. The Monefy platform connects them through rules, permissions, events and audit trails.
Interpretation and assistance
- Conversational scheduling assistant in the patient portal
- Self-scheduling and simple request management over SMS
- Advisory suggestions for claim drafts
- EOB structuring and prioritisation of cases requiring follow-up
Monefy backend rules · validations · permissions · audit
Execution and orchestration
- Automated patient notifications and reminders
- Event-driven insurance eligibility verification
- Invoice generation and exception control
- Billing Rules Engine and Posting Rules Engine
- Operational reporting through Metabase
- Central dashboard for automation configuration
Experimentally validated decisions
Rejected alternatives shaped the final architecture
Structured tool calling
Free-text interpretation was replaced by explicitly defined functions that are validated before execution.
The platform remains the source of truth
Availability, permissions, rules and statuses are verified in the backend rather than accepted directly from the model.
Precision over recall
In billing, the component flags only problems for which it has a high level of confidence.
Events and rules, not rigid scripts
Automation proceeds, stops or escalates according to context and risk level.
Control and responsible use
Automation is not allowed to ignore context
The central technical conclusion is not maximum autonomy, but calibrated autonomy: enough to remove repetitive work, bounded wherever sensitive data or financial impact is involved.
Core principle Humans remain in the loop for exceptions and sensitive decisions.
- 01
AI never accesses the database directly; interactions pass through the backend and controlled APIs.
- 02
Authentication, authorisation, permissions, validation and auditing are applied before any action.
- 03
Conversational flows were tested with adversarial messages and prompt-injection scenarios.
- 04
Unclear, incomplete or financially consequential cases are stopped and routed for human validation.
Transferable knowledge
What can be applied to similar projects
The results support a gradual and verifiable model for digitalising administrative processes that handle sensitive data.
AI becomes useful inside the rules
Natural language can be interpreted flexibly, but the result must be validated against structured data and rules.
Automation must accommodate exceptions
Real processes are not fixed sequences; they need thresholds, stopping points, retry and clear escalation paths.
Trust is built through verifiability
Users adopt recommendations when they can see the reason, control the action and continue manually.
Safe implementation is gradual
Low-risk workflows are validated first, then automation expands after measurement and feedback.
Project information
Adoption of advanced digital technologies
The research and development activity was carried out within the project “Adoption of advanced digital technologies within MONEFY BUSINESS S.R.L.”, RUE code 84, financed through the NRRP (National Recovery and Resilience Plan), call PNRR/2024/C9/MIPE/I3.1/GRANTS FOR SUPPORTING ENTREPRENEURS IN THE DEVELOPMENT OF ADVANCED DIGITAL TECHNOLOGIES - MEASURE 1, based on Financing Contract no. 84.1.I3.C9 of 13.05.2025.
- Lead partner
- MONEFY BUSINESS S.R.L.
- Partner
- Steady Dopetech S.R.L.