Every public body produces and receives thousands of documents a year. Most of them follow repetitive paths that a machine can prepare and an official can approve.
That sentence contains the distinction everything else rests on. Preparing isn’t deciding. A system can read, classify, check completeness and prepare. Responsibility for the act stays with whoever signs it.
In administrative practice, this distinction isn’t a matter of sensitivity: it’s the condition that allows the technology to be introduced without altering the accountability regime of the procedure.
The highly repetitive workflows
Not all of an agency’s activities lend themselves to this in the same way. The best candidates share three traits: high volume, written rules, a verifiable outcome.
The protocol system and certified email. Every incoming communication has to be read, classified by type and area of competence, and assigned to an office. It’s reading work that repeats identically hundreds of times a month and requires experience, not discretion.
Applications from citizens and businesses. Each type has a list of requirements and attachments. Checking which are present and which are missing is a formal, repetitive and perfectly definable control.
Freedom-of-information requests. Identifying the relevant documents, checking for data that needs to be redacted, preparing the response file. The search is the most burdensome part and the least discretionary.
Recurring acts. Spending determinations, payment orders, serial administrative measures. The structure is stable; what changes are the references and the amounts.
Internal consultation on regulations and procedures. When an operator needs to know how to handle a particular case, today they ask an experienced colleague. That colleague is a scarce resource and, in many agencies, close to retirement.
What the system actually does
Four operations, in order.
Classifies. Recognizes the type of the incoming document and which office is responsible for it. In uncertain cases it doesn’t force a choice: it flags the case and routes it to human review.
Extracts. Turns the content into structured data: the applicant, tax ID, subject, cited regulatory references, amounts, dates. Each extracted piece of data keeps a link to the exact point in the document it came from, so verification is a few seconds of reading rather than a full check.
Proposes. Checks completeness against the requirements for that type, lists the missing documents, prepares a structured summary for the case officer and, where applicable, a draft of the request for additional documentation.
Pre-fills. For serial acts, it populates the template with the correct references, leaving clearly marked the fields that require judgment.
The system issues nothing. It produces prepared material, with the sources alongside it.
The official’s role remains the decision
This point deserves to be spelled out in project documents, not just in presentations.
The official retains the assessment of merits, the review of doubtful cases, and the signature. What changes is where they apply their expertise: no longer on gathering and transcribing, but on deciding.
The nature of the review changes too. With every piece of data linked to its source, verification becomes targeted and documentable. The reviewer opens the reference and confirms, instead of rereading the whole file.
On the organizational side, this has a consequence that needs managing. The remaining work is on average more complex than before, because the simple cases now arrive already prepared. It’s an improvement, but it has to be supported: training needs to cover the system’s limits as much as its capabilities, and procedures need to state clearly what to do when the automatic proposal is wrong.
We’d add a point on the regulatory framework. A system that supports case preparation, doesn’t decide, and operates under effective human oversight sits in a different position from a system that determines access to a service. The classification, though, has to be made case by case, at the start of the project and together with the digital transition officer and the legal function, not after the technical choice has already been made.
The indicators to measure
A project in a public body is evaluated against numbers agreed before launch, while it’s still possible to record the starting situation.
Average processing time by case type. This is the main indicator. It has to be measured by type, because an overall average hides the differences.
Backlog. Number of cases past their legal deadline, and how it evolves. It’s the most externally visible indicator and the one an agency gets judged on.
Completeness errors caught downstream. How many cases get sent back for missing documentation that wasn’t caught at intake. Measures the quality of the initial check.
Classification accuracy. Share of documents routed to the correct office, measured on a manually verified sample. It should also be tracked after release, on a defined cadence.
Operational hours freed up. Not to cut staff, but to document where the time was redirected.
A warning about the fifth indicator: the number of queries to the system isn’t a measure of results. It says the tool is being used, not that the procedure has improved.
A clarification is needed on how to read the third indicator in the first months. A more systematic initial check surfaces at intake anomalies that used to be caught downstream, or not caught at all. The number of flags goes up while the number of rework cases goes down. These are two opposite movements that need to be read together, or the first can be mistaken for a regression.
What this approach doesn’t solve
It doesn’t shorten statutory deadlines or change the rules of the procedure.
It doesn’t solve archive fragmentation. If determinations and resolutions are spread across disconnected repositories, the first job is a mapping exercise, and it isn’t a technological one.
It doesn’t replace the procedural knowledge of experienced people. It can make that knowledge searchable if it’s written down somewhere. If it isn’t, it has to be collected first, and that’s an activity that takes those same people’s time.
And it doesn’t work on low-repetition case types. A complex, rare procedure doesn’t justify the cost of design and maintenance.
In closing
The approach we recommend to a public body is to start with a single case type, high volume and low risk, with a clear owner and measurable starting data. Extend afterward, reusing the rules already validated.
If you have a specific workflow in mind — protocol system, applications, freedom-of-information requests, serial acts — we’re available for a technical conversation about that process: what data it would take, where the controls would sit, and what indicators would be realistic to measure.
