The Challenge
A Canadian federal department processed over 40,000 applications annually across multiple program streams. The work was consequential — each application represented a person or organization waiting for a decision that materially affected their plans. And the department was failing them on time.
The intake process was almost entirely manual. Staff opened physical envelopes and sorted documents. They checked each application for completeness against a program-specific checklist, manually entering missing items into a deficiency tracking spreadsheet. They keyed data fields from handwritten or typed forms into a case management system that had been in place for over a decade. They physically moved file folders from inbox to routing queue to reviewer desk. From the moment an application arrived at the department's receiving address to the moment a program officer opened the file for their first substantive review, an average of 23 business days elapsed. During peak intake periods — which occurred twice annually and lasted 6–8 weeks each — the backlog extended to 8–10 weeks before first review.
Applicants had no visibility into where their file sat. The department's call centre fielded thousands of status inquiries each month, most of which staff could not answer meaningfully because the manual tracking system did not provide real-time location information. Service complaints were rising. A formal audit by the department's internal ombudsman had identified the intake processing timeline as a systemic service quality failure.
The department had attempted to fix this twice before. The first initiative, five years earlier, had contracted with a systems integrator to replace the case management system entirely. The project ran two years over schedule, consumed its full budget before completion, and was cancelled. The second initiative, two years later, attempted a more modest digitization of intake forms. It succeeded in producing an online application portal but did not address what happened to applications after they arrived — the data still had to be re-keyed from the portal into the legacy system by hand. The intake bottleneck was unchanged.
When the department engaged Remolda, the organizational posture toward technology projects was one of considerable skepticism. Staff had lived through two failed initiatives. Union representatives had concerns about AI and job security. Leadership needed a different approach — one that produced visible results quickly and did not require replacing existing systems.
The Approach
Remolda's three-week audit produced a finding that reframed the entire problem: the department did not need to replace its case management system. It needed an AI processing layer that sat in front of that system — handling all the manual work that currently delayed getting applications into reviewers' hands, while leaving the adjudication workflow and the existing system untouched.
Phase 1: Audit (3 weeks). We mapped the end-to-end workflow for each of the four primary application types, measured time on task at every step, and interviewed staff at each processing stage. The central finding: 70% of staff time in the intake unit was consumed by document handling, completeness checking, and data entry — tasks that required accuracy and attention but not judgment. The substantive eligibility review, which required program expertise and decision authority, was being delayed by bottlenecks in work that AI could handle.
We also assessed data quality in the existing system, analyzed the document formats and field structures across all application types, and reviewed the department's IT security and data residency requirements. The AI solution would need to operate within Government of Canada cloud security standards, process bilingual documents with equal accuracy, and produce audit-ready logs for every automated action.
Phase 2: Strategy (4 weeks). Rather than attempting all four application types simultaneously, we designed a three-wave implementation beginning with the highest-volume program stream — one that represented 60% of annual intake volume and had the most standardized document formats. Success metrics were defined with the department's leadership before any implementation began: processing time from receipt to first review, data entry accuracy rate, staff satisfaction scores, and applicant acknowledgement time.
We presented the implementation plan to union representatives at this stage. The framing was direct: AI would eliminate the most tedious and error-prone parts of intake work, and staff whose roles were most affected would be redeployed to program review and applicant support — work that required human judgment and that the department valued highly. Union leadership asked hard questions. We answered them with specifics about which tasks would change, which would not, and what redeployment would look like.
Phase 3: Implement (6 months, 3 waves).
-
Wave 1 targeted the primary application type. We deployed an AI document classification engine that received applications from the intake queue, identified document types, verified presence of all required documents against the program checklist, and flagged incomplete submissions for automated deficiency notices. For complete applications, the AI extracted 47 structured data fields and populated the case management system directly — eliminating all manual data entry for this application type. Processing time for a complete application in Wave 1 dropped from 23 business days to 11 business days before Wave 2 began.
-
Wave 2 extended the system to three additional application types, each with its own document structure and data requirements. The AI routing engine was deployed at this stage: based on extracted data fields, the system automatically routed applications to the appropriate program stream and reviewer queue, eliminating the manual sorting step that had required three full-time staff positions.
-
Wave 3 added bilingual processing capability — the system now processes applications in English and French with equal accuracy, using language-aware extraction models rather than translation. Integration with the department's correspondence system was completed, enabling automated status acknowledgements to applicants within 48 hours of submission and proactive deficiency notices within 5 business days.
Phase 4: Empower (parallel). We delivered AI literacy and operational training to 85 staff members across three offices. The emphasis was not on using AI — the system was largely invisible to end users — but on working effectively with AI-processed applications: verifying extracted data, handling exception cases that the AI flagged for human review, and providing structured feedback that improved the system's accuracy over time.
The Results
- 65% reduction in processing time. Average time from receipt to first review dropped from 23 business days to 8 business days. During the most recent peak intake period — traditionally the department's worst service quality moment — the backlog did not exceed 12 business days, compared to 8–10 weeks in the prior year.
- Data entry error rate fell from 4.2% to 0.8%. The AI system's extraction accuracy for structured fields was 96.5%, with confidence-based routing sending uncertain extractions to human verification rather than propagating errors into the case management system. Downstream reviewers reported a material reduction in file quality issues.
- 12 staff redeployed to higher-value work. The three staff positions previously dedicated to manual routing were dissolved, and nine additional positions that had spent the majority of their time on data entry were reoriented to program review and applicant support roles. Staff satisfaction surveys showed improved scores — the redeployed staff reported that their new work was more meaningful and made better use of their program knowledge.
- Applicant experience transformed. Applicants now receive automated acknowledgement of their submission within 48 hours — previously, this took 2–3 weeks. Deficiency notices arrive within 5 business days (previously 4–6 weeks). Call centre volume for status inquiries dropped 34% in the six months following full deployment.
- Bilingual service equity achieved. French-language applications previously moved through a separate, slower manual process. The AI system processes both official languages with equivalent accuracy and speed, resolving a long-standing service equity issue flagged in the ombudsman audit.
Key Lessons
1. Integration beats replacement. Both of the department's previous technology initiatives failed because they attempted to replace the legacy case management system — a high-risk, high-cost approach that required stakeholder alignment across the department's entire workflow. The AI layer approach left the existing system in place and targeted only the manual work upstream of it. Lower risk, lower cost, faster results, and no disruption to the adjudication workflow that program officers depended on.
2. Start with the bottleneck, not the hardest problem. The eligibility review process involved complex policy judgment and was rightly seen as requiring human expertise. But that was not the bottleneck. The bottleneck was intake — mechanical, high-volume, accuracy-critical work that delayed the eligibility review from even starting. Automating the bottleneck had cascading effects: faster intake meant faster first reviews, which meant faster decisions, which meant fewer status calls from applicants.
3. Staff engagement is not a risk mitigation measure — it is a success factor. Departments that have been through failed technology initiatives carry institutional skepticism. Staff who have seen previous projects collapse are not going to adopt new systems enthusiastically unless they understand what changes, what does not, and what is in it for them. The union consultation in Phase 2, the specificity of the redeployment plan, and the quality of the empowerment training all contributed to an adoption rate that exceeded projections. Staff who felt informed and respected became system advocates.
For federal and provincial government departments looking to modernize document-intensive processes without replacing legacy systems, see Remolda's government AI services and our document processing capabilities.