AI for Public Sector and Education
The only sector where the person affected by the model cannot take their business elsewhere. Twelve verticals, written for the official who will have to explain the decision to the citizen it was made about.
The citizen cannot go to a competitor.
Every other sector on this site serves customers who can leave. A benefits claimant, a defendant, a taxpayer under investigation, a child assigned to a school — none of them can. That asymmetry is why public sector AI carries obligations no private deployment does, and why the failures are political rather than commercial. The Dutch childcare benefits scandal wrongly accused roughly 26,000 families of fraud between 2005 and 2019, and the third Rutte cabinet resigned over it on 15 January 2021. The lesson usually drawn is about bias. The more useful lesson is that there was no working route to challenge a decision and nobody able to explain how one had been reached.
This is also the only sector where the EU AI Act simply prohibits things rather than regulating them. Social scoring by public authorities, predictive policing based solely on profiling, emotion inference in schools and workplaces, and untargeted scraping of facial images are banned outright, with those prohibitions in force since 2 February 2025. On top of that, public bodies deploying high-risk systems must complete a fundamental rights impact assessment under Article 27 — an obligation private companies deploying the same system do not carry. We would rather establish which of these applies in week one than build something that has to be withdrawn in front of a committee.
Where AI actually earns its place in public service and education
Ranked by evidence rather than by how often it appears in a digital strategy. The maturity column is our own read; the catch column is what the vendor deck leaves out.
| Where | What it does | Maturity | The catch |
|---|---|---|---|
| Document and case file extraction | Pulls structured facts from forms, applications and correspondence. | Proven | The strongest case across government. High volume, checkable, and it touches no decision about anyone. |
| Citizen enquiry handling and navigation | Answers questions about services, eligibility and process. | Strong | Real value, because public information is genuinely hard to navigate. Must never state an entitlement as a decision. |
| Translation and accessibility | Makes public information available across languages and formats. | Proven | Among the highest-value uses in government, and among the least funded. |
| Backlog triage and routing | Orders queues by urgency and routes work to the right team. | Strong | Effective. Ordering a queue is defensible; removing someone from it is a decision requiring a person. |
| Teaching and administrative assistance | Drafts materials, marks formative work, reduces teacher workload. | Strong | Workload is the actual crisis in schools, and this addresses it without touching assessment. |
| Asset and infrastructure condition | Reads inspection data for roads, buildings and networks. | Strong | Well-evidenced, and the same failure-example problem as everywhere else applies. |
| Fraud and error detection | Flags claims or returns for human investigation. | Good | Works, and it is the exact application that produced the Dutch scandal. The design of what happens next is the whole thing. |
| Demand and resource forecasting | Predicts service volumes for planning and budgeting. | Good | Solid where history is long and stable. Public demand series break at policy changes, which models do not anticipate. |
| Predictive risk scoring of individuals | Scores people for intervention, enforcement or supervision. | Contested | Legally constrained, ethically loaded, and the evidence for accuracy is weaker than the enthusiasm. |
| AI detection of student writing | Judges whether a student's work was AI-generated. | Not defensible | Peer-reviewed evidence puts the false positive rate above 61% for non-native English writers. We will not build it. |
| Emotion inference in classrooms | Infers engagement or affect from students. | Prohibited in the EU | Banned outright in education and workplace contexts since February 2025. |
Three of these are decided before the technical conversation starts
Emotion inference in schools is prohibited in the EU and should not be scoped anywhere. AI detection of student writing fails on evidence rather than on law: a peer-reviewed study in Patterns found seven widely-used detectors misclassified TOEFL essays by non-native English writers as AI-generated at an average false positive rate of 61.22%, with all seven unanimously flagging around a fifth of them, while performing near-perfectly on US student essays — a tool with that profile in an academic misconduct process disproportionately accuses international students of cheating. And predictive risk scoring of individuals sits under prohibitions, high-risk obligations and a body of evidence thinner than its advocates suggest. Everything else on this list is buildable, and most public bodies have not yet done the unglamorous work at the top of it.
Prohibitions, extra duties for public bodies, and the rules on buying
This sector answers to administrative law and the duty to give reasons, the AI Act's prohibitions and its additional obligations on public deployers, education and child protection rules, and public procurement regimes. This is our reading as at September 2026 and we work alongside your legal, policy and information governance functions rather than in place of them.
In force since 2 February 2025
Social scoring by public authorities, predictive policing based solely on profiling or personality traits, emotion inference in workplaces and education institutions, untargeted scraping of facial images, and real-time remote biometric identification in public spaces for law enforcement outside narrow authorised exceptions. These are not risk-managed categories — they are prohibited, with the highest penalty tier attached.
An obligation only public deployers carry
Bodies governed by public law, and private entities providing public services, must complete a fundamental rights impact assessment before first deploying a high-risk system: the deployment process, frequency of use, categories of persons affected, specific harms likely to result, human oversight arrangements, and the complaint mechanism. The completed assessment is notified to the market surveillance authority. A private company deploying an identical system does not have this duty.
Admission, evaluation, proctoring and benefits
Education systems determining admission, evaluating learning outcomes, assigning people to institutions, or monitoring prohibited behaviour during tests are high-risk, as are systems determining eligibility for essential public benefits and services. High-risk obligations apply from 2 December 2027 following the Omnibus agreement of May 2026.
Rewritten in April 2025
OMB memoranda M-25-21 and M-25-22, both issued 3 April 2025, replaced M-24-10. Agencies designate Chief AI Officers, establish governance boards and publish AI strategies, and apply minimum risk management practices to high-impact AI — pre-deployment testing, impact assessments, human oversight with intervention, and remedies or appeals for affected individuals. M-25-22 governs procurement, covering data portability, vendor lock-in and performance monitoring across the contract.
Older than any AI statute and frequently the binding one
Administrative law generally requires that a decision affecting a person can be explained to them and challenged. A system whose output cannot be explained in terms an official can put in a decision letter is not usable in that decision, regardless of accuracy. This constraint predates every AI framework, applies in most jurisdictions, and is the one most often discovered late.
A separate and stricter regime
Education deployments involve children's personal data, which carries heightened protection, restricted lawful bases, and in many jurisdictions specific codes on profiling and age-appropriate design. Consent from a school is not consent from a family, and that distinction has ended more than one edtech rollout.
What this means for a build
Three things are architectural rather than procedural. Any system touching a decision about a person needs an explanation an official can put in a letter and an appeal route a person can actually use — designed in, not added after, because that is precisely what was missing in the Dutch case. Anything high-risk in a public body needs the Article 27 assessment before first deployment, which means the harms analysis shapes the build rather than documenting it. And prohibited practices need identifying at scoping, because no amount of governance makes them deployable. See AI risk assessment and EU AI Act compliance.
Who we write for
Each page starts from that organisation's own problems, names the regulatory exposure it carries, and routes into the engineering. Depth varies and is stated on each page.
Federal & national government
Case extraction, backlog triage and enquiry handling — built to the impact assessment and appeal standard from the start.
Read more →State & local government
Service request triage, planning and licensing, and asset condition — with a clear line around social care prediction.
Read more →Smart cities
Traffic, environmental sensing and city asset condition — with the surveillance line stated rather than implied.
Read more →Public safety & policing
Evidence review, call handling and resource forecasting — with the prohibitions and our own refusals stated plainly.
Read more →Defence & military
Sustainment, logistics and administration — with an explicit line around targeting and lethal decisions.
Read more →Judiciary & courts
Listing, transcription and case file preparation — with judicial decision support left where it belongs.
Read more →Tax & revenue agencies
Return processing, enquiry handling and compliance risk — designed around the appeal route from the start.
Read more →K-12 education
Teacher workload, attendance analytics and accessibility — with detection and emotion inference declined outright.
Read more →Higher education
Student support, research administration and admissions — with the integrity detection evidence stated plainly.
Read more →EdTech
Adaptive learning, children's data architecture and the classification your customers inherit from you.
Read more →Vocational training
Content adaptation, completion modelling and funding administration — with competence certification left to assessors.
Read more →Nonprofits & NGOs
Grant administration, funder reporting and service support — scoped to what a small organisation can actually operate.
Read more →FAQ
Marked up with FAQPage schema so these answers can surface in search results and inside AI assistant responses.
Do you have public sector and education experience?
Yes in document and case file extraction, citizen enquiry handling, backlog triage, demand forecasting and education operations analytics. Less in policing and defence specifically, and none in classified environments — we would not claim otherwise. Each of the twelve vertical pages states our depth in that area rather than implying uniform expertise across a sector spanning a national tax authority and a primary school.
What is the single most important design decision in a public AI system?
The appeal route, and it is usually treated as a policy question after the build. The Dutch childcare benefits scandal is normally discussed as an algorithmic bias case, and roughly 26,000 families were wrongly accused of fraud over fourteen years, with the cabinet resigning in January 2021. What made it catastrophic rather than merely wrong was that people flagged by the system had no working route to challenge the decision and no official could explain how it had been reached. Design the explanation and the appeal first; the model is the easier half.
Can we use AI to detect whether students used AI?
We will not build it, and the reason is evidence rather than principle. A peer-reviewed study in Patterns tested seven widely-used detectors and found an average false positive rate of 61.22% on TOEFL essays written by non-native English speakers, with all seven unanimously flagging around a fifth of them, while performing near-perfectly on essays by US students. A tool with that profile inside an academic misconduct process systematically accuses international students of cheating. The defensible responses are assessment redesign and process-visible work, and we would rather help with that.
What does Article 27 actually require of us?
If you are a body governed by public law, or a private entity providing public services, and you deploy a high-risk AI system, you must complete a fundamental rights impact assessment before first use and notify it to the market surveillance authority. It covers the deployment process, how often the system is used, which categories of people are affected, the specific harms likely to result, the human oversight arrangements and the complaint mechanism. The practical consequence is that the harms analysis shapes what gets built, because several answers are only acceptable if the system is designed differently.
Where should a public body start?
Document and case file extraction, citizen enquiry navigation, and translation. All three are high-volume, measurable, and touch no decision about any individual, which means they deliver while the harder governance questions are being settled properly rather than after a rollout. Translation in particular is among the highest-value and least-funded uses in government: public information that people cannot read in a language they use is a service failure that technology now genuinely fixes. The decision-adjacent work should wait for the appeal route and the explanation to be designed.
Start with the problem, not the technology.
Thirty minutes, no charge, no deck. Tell us what is going wrong in your organisation and we will tell you whether AI is the right instrument — including when it plainly is not.