.png)
A state agency handling 12,000 monthly submissions reduced its processing backlog by 81% with an automation pipeline built to government compliance standards.
A state professional licensing agency responsible for processing applications, renewals, and compliance submissions across 14 license categories — including contractors, healthcare practitioners, real estate brokers, and financial advisors — was receiving approximately 12,000 submissions per month. Every submission arrived as a physical or scanned document package, was manually reviewed by a licensing examiner, and entered into the agency's case management system by a data entry team before any eligibility determination could begin.
Average processing time from submission receipt to eligibility determination was 22 days. The agency's statutory processing commitment was 10 days. A backlog of 34,000 unprocessed submissions had accumulated over 18 months, driven by a combination of volume growth following post-pandemic professional mobility, examiner attrition, and a hiring freeze that had been in place for 14 months. The agency's ombudsman office had received 2,800 formal complaints in the previous fiscal year related to processing delays — a figure that had reached the attention of the state legislature.
Verttx built an AI document processing pipeline operating within the agency's existing IT environment and security perimeter — no cloud migration, no system replacement, no new procurement category requiring legislative approval. The pipeline went live 13 weeks after the initial discovery call. Within 90 days, the backlog had been reduced by 81% and average processing time had fallen to 4 days.
The agency's processing workflow had four sequential steps, each performed manually: document receipt and sorting, completeness review, data extraction and TMS entry, and examiner eligibility determination. The first three steps — the administrative pipeline before a human examiner ever touched a file — were consuming an average of 11.4 days of the 22-day total processing time. Examiners were spending only 10.6 days on the actual eligibility work that required their professional judgement and licence-category expertise.
Completeness review was the single largest time sink. Agency regulations required that submissions include a specific set of supporting documents — educational credentials, background check results, professional references, insurance certificates, and fee payment confirmations — that varied by license category. Examiners were manually checking each submission package against a 14-category completeness checklist, issuing deficiency notices for incomplete submissions, and re-queuing returned submissions when the missing documents arrived. Approximately 31% of submissions were incomplete on first receipt, and the average time from deficiency notice to complete resubmission was 8.3 days — time the file sat in a queue adding to the processing average without any work being done.
The overtime situation had become financially and operationally unsustainable. The agency had authorised unlimited overtime for all 84 licensing examiners and 22 data entry staff for the previous 11 months in an attempt to reduce the backlog. Overtime expenditure had reached $2.1 million annually — a figure that required a legislative budget amendment and that the agency's director had been required to justify to the appropriations committee twice in the same fiscal year. Despite the overtime spend, the backlog had continued to grow because new submission volume was outpacing processing capacity even at maximum staffing effort.
Government AI deployments have constraints that private sector projects do not. The agency's IT security policy prohibited processing of personally identifiable information on external cloud infrastructure — a requirement driven by state data residency law and the agency's existing ATO (Authority to Operate) documentation. Verttx deployed the entire processing pipeline on-premise within the agency's existing server infrastructure, operating under the agency's current ATO without requiring a new security assessment or procurement cycle. This was a deliberate architectural constraint, not a compromise — building within the existing perimeter was the only path to a 13-week deployment timeline in a government environment.
The pipeline's first stage classifies every incoming document by type and license category using a computer vision model trained on 180,000 historical submission documents from the agency's own archive. The classifier identifies the document type (application form, credential certificate, background check result, insurance certificate, fee receipt, or reference letter), the license category the submission relates to, and whether the document is legible and complete enough to process. Classification accuracy on the production document stream is 98.7%.
The completeness checking module then compares the classified document set against the regulatory completeness requirements for that license category — automatically generating a completeness determination and, for incomplete submissions, a draft deficiency notice populated with the specific missing documents. Deficiency notices that previously took an examiner 12-18 minutes to produce manually are now generated in 40 seconds and reviewed by an examiner before sending — a review that takes an average of 90 seconds because the notice is pre-populated and the examiner is confirming rather than composing.
Complete submissions are processed through an extraction pipeline that reads the relevant data fields from each document type — applicant identity fields, credential dates and issuing institutions, insurance coverage amounts and policy numbers, and fee payment references — and populates the agency's case management system directly. The extraction model was trained separately for each of the 14 license categories because the supporting document requirements and field structures differ significantly across categories. Extracted data is validated against the agency's existing reference databases — credential issuing institution records, background check provider APIs, and insurance carrier verification services — before the populated case record is released to the examiner queue.
The 34,000-submission backlog required a dedicated clearance programme running in parallel with the live processing pipeline. Verttx configured a batch reprocessing pipeline that worked through the backlog in priority order — oldest submissions first, with escalation for submissions where applicants had filed formal complaints — processing 2,400 backlog submissions per day through the AI pipeline. The backlog was reduced from 34,000 to 6,400 submissions in the first 60 days of operation, reaching the agency's target operating level of under 5,000 submissions within 90 days.
Average processing time from submission receipt to examiner queue fell from 22 days to 4 days within the first full quarter of operation. The 11.4-day administrative pipeline that had preceded every examiner review was reduced to an average of 18 hours for complete submissions processed through the AI pipeline. Examiners, previously spending significant capacity on administrative triage, now spend their full working day on eligibility determinations — the professional judgement work that requires their expertise and that the agency's statutory mission depends on.
The document backlog fell from 34,000 to under 5,000 submissions within 90 days — an 81% reduction. Formal complaints to the ombudsman office related to processing delays fell from 2,800 in the previous fiscal year to an annualised rate of 340 in the first year post-implementation — an 88% reduction. The agency met its 10-day statutory processing commitment for 94% of submissions in the first full quarter, the first time it had met that commitment in three years.
All overtime authorisation was withdrawn within 60 days of full deployment as processing capacity caught up with incoming volume. Annual overtime expenditure of $2.1 million was eliminated entirely. The agency's director presented the outcome to the appropriations committee as the first budget amendment in three years that reduced rather than increased the agency's personnel cost line. The deficiency notice workflow — which had previously consumed 31% of examiner capacity on incomplete submissions — now consumes 4%, with examiners reviewing AI-generated notices rather than composing them from scratch.
The complete processing pipeline — the classification model, all extraction models, the validation rules engine, the deficiency notice generator, and the batch reprocessing infrastructure — was transferred to the agency's IT team at handover, deployed on the agency's own infrastructure with no external dependency on Verttx for ongoing operation.
We had been told by two technology vendors that fixing this required replacing our case management system. That was an 18-month project we did not have time or budget for. Verttx built around what we had and had the pipeline running in thirteen weeks. The backlog that had been the subject of legislative hearings was gone in ninety days. — Deputy Director of Operations, State Licensing Agency
.png)
Average processing time fell from 22 days to 4 days. The 34,000-submission backlog was reduced by 81% within 90 days. The agency met its 10-day statutory processing commitment for 94% of submissions in the first full quarter — the first time in three years. Ombudsman complaints related to processing delays fell 88%. All overtime authorisation was withdrawn within 60 days, eliminating $2.1 million in annual overtime expenditure. Deficiency notice production time fell from 12-18 minutes to 40 seconds per notice.
.png)
.png)
