# Mapping AI use cases across the public value chain
A mid-size US city spent three years and roughly $2 million (estimate, based on typical municipal IT contracts reported in local government budget documents) building a generative AI chatbot to answer resident questions about parking permits. The chatbot regularly told people the wrong renewal deadline. Meanwhile, the same city's building department was still assigning permit inspectors by hand, using a spreadsheet, when a simple rules-based scheduling algorithm could have cut wait times by half. This is the pattern across public sector AI adoption: budget flows toward visible, flashy tools and away from boring, high-value automation.
This lesson gives you a framework to tell the difference, by walking the actual value chain of a city government.
Borrowed from Michael Porter's value chain concept but adapted for government, this is the sequence of activities a public agency performs to convert policy intent into services delivered to residents:
1. Policy design (what rules should exist)
2. Program and budget planning (how to allocate resources)
3. Permitting and regulatory review (approving what residents/businesses want to do)
4. Citizen services and case management (benefits, complaints, requests)
5. Field operations (inspections, maintenance, public safety response)
AI's value differs sharply at each stage. The mistake most agencies make is applying one tool (usually a chatbot) everywhere, instead of matching the AI type to the task.
Here the task is synthesizing evidence: comparing how other cities regulated short-term rentals, summarizing public comment volumes, drafting first pass language for ordinances.
Good fit: Large language models (LLMs, AI systems trained on text to generate and summarize language) for literature review, comment summarization, and drafting. The OECD's AI in government observatory tracks real deployments like this across member states.
Bad fit: Letting a model recommend policy positions. It has no accountability, and hallucinationhallucinationA hallucination is when an AI model generates output that is fluent and confident but factually wrong, fabricated, or unsupported by its source data.Voir la définition complète → (confidently generated false information) risk is highest when there's no ground truth document to check against.
Budget offices need to predict service demand: how many shelter beds next winter, how much road salt to stockpile, where crime or fire risk clusters geographically.
Good fit: Traditional machine learning (ML), statistical models trained on historical data to predict outcomes. Time series forecasting and geospatial risk models have decades of academic and operational validation. Chicago's food inspection prioritization model (using ML to rank which restaurants to inspect first) is a well-documented public example.
Bad fit: Generative AI. You don't need a model that writes fluent sentences, you need one that predicts a number accurately and explains why.
This is the hook's central example, and it's the stage agencies get wrong most often.
Permitting is fundamentally rule-following: does this building meet setback requirements, is this business license application complete, does this applicant meet income thresholds for a subsidy program. These are deterministic decisions (same inputs always produce the same correct output), governed by statute.
Good fit: Rules engines, robotic process automation (RPA, software that automates repetitive digital tasks like data entry across systems), and optical character recognition (OCR) to digitize paper applications. These are unglamorous but reliably cut processing time.
Bad fit: LLMLLMA Large Language Model is an AI system trained on vast text data to predict and generate language, enabling tasks like writing, summarizing, and answering questions.Voir la définition complète → chatbots interpreting eligibility rules. A generative model doesn't "know" the zoning code, it predicts plausible-sounding text. If the model says a permit is approved and it's wrong, the city owns that liability. The UK's Office for AI and the US GSA's AI guide for government both flag this: use generative AI for drafting explanations of a decision already made by a rules engine, not for making the decision itself.
Simple pseudocode for the deterministic approach:
if applicant.income <= threshold_for_program:
if applicant.documents_complete == True:
status = "eligible"
else:
status = "pending_documents"
else:
status = "not_eligible"No model needed. This is what most "AI failed in government" stories actually are: agencies replacing code like this with a probabilistic model that introduces error where none needed to exist.
This is where chatbots and LLMs genuinely earn their budget, but only for the right sub-tasks: answering natural-language questions by retrieving from verified documents (a pattern called retrieval-augmented generation, or RAG, where the model pulls from an approved knowledge base rather than inventing answers), translating service information into multiple languages, and triaging incoming complaints to the right department.
Boston, Los Angeles, and several EU municipalities have piloted RAG-based assistants for 311 non-emergency service lines (estimate based on public pilot announcements circa 2023-2025). The key design choice: the model retrieves from an approved FAQ or code document, and cites it, rather than generating free-form legal interpretation.
Bad fit: Chatbots without retrieval grounding, deployed as a public-facing "ask anything" interface with no guardrails. This is exactly the parking permit failure from the hook.
Inspectors, sanitation crews, and public works benefit from computer vision (AI that interprets images/video) for tasks like detecting potholes from dashboard camera footage, monitoring water main pressure sensors for leak prediction, or flagging code violations from satellite/aerial imagery.
Good fit: Narrow, well-labeled computer vision models trained for one task (pothole detection, not "general road quality").
Bad fit: General-purpose AI agentsAI agentsAgentic AI refers to AI systems that pursue goals autonomously by planning, taking actions through tools, and adapting based on results, with minimal step-by-step human direction.Voir la définition complète → making autonomous field decisions (e.g., automatically issuing fines from detected violations without human review). Due process concerns under US administrative law, and GDPR-adjacent transparency rules in the EU (General Data Protection Regulation, governing automated decision-making about individuals), generally require a human in the loop for anything with legal consequence.
Inspectors general and audit offices use ML for fraud detection in benefits programs (flagging statistically unusual claim patterns) and procurement anomaly detection (flagging bids that cluster suspiciously). This is one of the best-documented ROIROIReturn on Investment: the ratio of net profit to the cost of an investment. A 300% ROI means each dollar invested returns $3.Voir la définition complète → (return on investmentreturn on investmentReturn on Investment: the ratio of net profit to the cost of an investment. A 300% ROI means each dollar invested returns $3.Voir la définition complète →) cases in public sector AI: CMS (Centers for Medicare & Medicaid Services) fraud detection systems have publicly reported recovering hundreds of millions of dollars annually (estimate, figures vary by fiscal year, see CMS program integrity reports).
Vérification des acquis
1. What is the core lesson illustrated by the city that built an expensive chatbot while permit inspectors were still scheduled by hand on a spreadsheet?
2. According to the framework, why does AI's value differ across stages of the public value chain?
3. In the policy design stage, why are large language models described as a 'research accelerant, not decision-maker'?
4. Select ALL correct answers about the six stages of the public value chain described in the lesson.
Sélectionnez toutes les réponses correctes.
5. Select ALL correct answers about why matching AI type to task matters in government AI adoption.
Sélectionnez toutes les réponses correctes.
Ask three questions before approving budget:
1. Is the underlying decision deterministic (rule-based) or probabilistic (pattern-based)? Rules engines for the former, ML/AI for the latter.
2. What's the cost of being wrong, and who is accountable? High-stakes, legally consequential decisions need human review regardless of model accuracy.
3. Does a generative AI answer need to be grounded in an authoritative document? If yes, RAG architecture is mandatory, not optional.
Agencies that skip this diagnostic tend to buy generative AI for stage 3 (permitting) problems and rules engines for stage 4 (citizen services) problems, exactly backwards from where the value sits.
🎬 [VIDEO: "How AI Is Being Used in Government Right Now" - youtube.com/results?search_query=AI+in+government+use+cases+2025 - search for recent explainer content from government technology channels like GovTech or Code for America covering real municipal AI deployments]