Public sector AI projects often get pitched on a premise that does not survive contact with reality: that automation will replace human judgment. The actual gains, in our experience, come from a different place. They come from rebuilding the queue that surrounds human judgment so that the judgment itself happens earlier, on better-prepared cases, with the right context at hand. The 21-to-6 result is a queue improvement, not a judgment replacement. Understanding the difference matters because most failed public-sector AI projects fail by trying to do the wrong one.
What the 21-day cycle actually contained
When we measure permit pipelines before any intervention, the pattern is consistent. The 21-day median decision time decomposes roughly as follows. Three days from application submission to intake review. Six days waiting in the queue for case-officer assignment. Four days from assignment to first substantive review. Five days for back-and-forth with the applicant about missing information. Three days from final dossier to decision. The decision itself, in our analysis, took less than a day across more than 90 percent of cases. The other 20 days were queue, hand-offs, and missing-information loops. The decision quality was not the constraint. The work surrounding the decision was.
The decision itself took less than a day in over 90 percent of cases. The other twenty days were queue.
Where the OECD framework points
The OECD's 2024 report Governing with Artificial Intelligence examined 200 real-world examples of governments using AI across 11 core functions and identified automated and streamlined service delivery as one of the highest-impact use cases [1]. The European Commission's parallel work, summarized in survey results from 576 public managers across seven countries, found that perceived AI benefits significantly influence adoption, with administrative process improvement consistently ranking near the top [2]. The pattern across the documented public-sector deployments is the same: the gains come from process improvement, not decision replacement, and the deployments that scale are the ones that respect that distinction.
What was automated in the permit case
Three layers. First, intake validation. The application form was redesigned around the actual permit categories, and the automation checked completeness, internal consistency, and basic eligibility before submission. Applications that failed checks could not submit; the applicant got specific guidance about what was missing. This moved the missing-information loop from day twelve to day zero. Second, case routing. Applications were tagged by complexity using rules derived from prior case patterns. Simple cases went to a streamlined queue with shorter SLA expectations. Complex cases went to specialist queues with the additional context they required. Routing was not a decision; it was a load-balancing operation. Third, dossier preparation. By the time a case officer opened a file, the supporting documents were attached, the relevant precedents were linked, and the standard checklists were pre-populated. Case officer time per case dropped, but the cases themselves received more attention per minute spent.
What did not change
Three things stayed exactly the same, and this is the part that matters for legitimacy. The approval thresholds did not change. The legal standards applied to permits remained identical. The human authority over each decision remained intact. No automated decisions were made on substantive permit questions. The case officer's judgment on the merits was not replaced or assisted by AI in any way that affected the outcome. What changed was the queue, the intake quality, and the preparation of the file. The scrutiny applied to each permit was equal or greater than before, just delivered faster.
Why this matters for legitimacy
Public-sector AI projects that automate consequential decisions face a legitimacy problem that no technology can solve. Citizens have a right to human judgment on decisions that affect them. The EU AI Act explicitly identifies AI use in justice administration and public services as a high-risk category subject to strict requirements [3], with the high-risk-system rules entering full applicability from August 2026. Belgian and EU public bodies designing AI deployments in 2026 are doing so against a regulatory framework that distinguishes carefully between decision-support automation and decision-replacement automation. The permit case worked within the first category, not the second. The 71 percent time reduction was achieved while strengthening the position of human decision-makers, not weakening it.
Citizens have a right to human judgment on decisions that affect them. The technology that respects this is the technology that scales.
The data quality problem that almost stopped the project
The single largest obstacle in the permit case was not technology. It was the historical case data. Three years of permit applications had been recorded in inconsistent ways across multiple systems, with free-text fields that did not lend themselves to structured analysis. Routing rules based on historical complexity tagging required the historical complexity to be tagged consistently, which it was not. Six weeks of the eight-month project went into data cleanup. This is the most common cost overrun in public-sector AI work and the one most often invisible at project kickoff. The lesson is to measure data quality before scoping the project, not after.
What scaled, what did not
The intake validation scaled across permit categories with modest configuration changes. The case routing required category-specific rules and could not be ported wholesale. The dossier preparation was the most labor-intensive to extend, because each permit type referenced different precedents and required different checklists. The pattern is general: simple automation generalizes easily, intelligent automation generalizes painfully. Plan for category-by-category build rather than a single platform that handles everything. This is also where most vendor-pitched solutions fail, because vendors are incentivized to sell the platform story rather than the per-category configuration that actually delivers value.
The Belgian and EU context
Two factors shape this work in the Belgian and EU context. First, citizen-facing AI in public administration sits inside a regulatory framework that is more advanced and more prescriptive than equivalent frameworks elsewhere. The EU AI Act, OECD AI Principles updated in 2024, and member-state implementation work all converge on similar requirements: transparency, human oversight, bias evaluation, post-deployment monitoring. Building these in from project inception is more efficient than retrofitting. Second, Belgian public bodies operate in a multilingual environment that adds complexity to any text-processing automation. NL, FR, and EN are all real production languages for federal-level work, and Brussels-region work adds Arabic and Turkish at the citizen-interface layer. Generic IDP solutions struggle with this. Belgian deployments require Belgian-specific language coverage from day one.
The hardest part is not the technology
The hardest part of the permit project was not the technology, the data, or even the regulatory framework. It was the change-management work with the case officers themselves. The reasonable concern was that automation would be a step toward role reduction. The honest answer was that automation removed the parts of the role officers liked least, which left more time for the substantive work most officers found most meaningful. This was true. It also took six months of demonstrated practice before it was believed. Public-sector technology projects without that change-management investment usually fail, not because the technology is wrong, but because the people the technology was meant to support never adopted it. Budget the change work explicitly. It costs more than the build, and it is what determines whether the gains compound.
