One concrete problem
Start with a service or process that has demand, accountable owners and an observable baseline.
PRACTICAL GUIDE · COSTA RICA · 2026
How to implement artificial intelligence in a municipality
A path for turning a public need into a useful, measurable and governed service—without handing institutional decisions to a black box.
SIX CONDITIONS FOR SUCCESS
What must exist before scaling.
Authorized sources
Define what information the AI may consult, which source remains authoritative and how it is updated.
Human authority
Keep legal, financial, technical and discretionary decisions with accountable officials.
Governed interoperability
Connect existing systems through bounded permissions, clear contracts and traceability.
Testing and safe degradation
Test responses, actions and failures before expanding scope; when uncertain, abstain or escalate.
Measurement from the start
Compare time, resolution, quality, adoption and exceptions against an agreed baseline.
01
Implementing AI is not installing a chatbot.
In local government, a useful implementation connects service, knowledge, data, tools and accountable owners. A conversational channel may be visible, but value appears when the institution can resolve, verify, escalate and preserve evidence.
It is
- A service with a purpose and accountable owners.
- A combination of AI and deterministic tools.
- An operation with sources, permissions and traceability.
- A continuous cycle of testing and improvement.
It is not
- A model connected without limits to all data.
- An automatic replacement for public officials.
- A promise of zero errors or autonomous decisions.
- A pilot abandoned after the demonstration.
The right question is not “which model should we buy?” but “which public service do we want to improve, with what evidence and under whose authority?”
02
Where AI can create value first.
Prioritize a case with real demand, accessible information and an accountable owner. Not every process needs AI, and not every automation needs a generative model.
Resident service and guidance
High volumes of repeated questions, multiple formats and a need for continuous service.
WATCH FORDo not turn general guidance into an official ruling or decision.Procedures, case files and permits
Applications, requirements, corrections, documents, inspections and owners are fragmented.
WATCH FORAI may prepare and verify; the competent authority decides and signs.Institutional knowledge
Regulations, resolutions, manuals and questions whose answers change over time.
WATCH FOREvery answer should preserve its source, validity period and a correction path.Documents and evidence
Manual classification, extraction, review and transfer of information between systems.
WATCH FOROCR or extraction is not truth: sensitive fields require validation.Inspections and field operations
Assignment, geographic evidence, checklists and continuity with the case file.
WATCH FORMap context does not replace a verified inspection when one is required.Operational analytics and prioritization
Queues, deadlines, exceptions and workloads that need follow-up.
WATCH FORA prediction must not hide its formula, data or uncertainty.03
Institutional readiness checklist.
If several answers are “no,” the first deliverable should be diagnosis and design—not a production integration.
- There is a defined operational problem, not merely a desire to “use AI”.
- The process has an accountable owner and people who can validate the outcome.
- Official sources, their validity and their owners are identified.
- The institution knows which data is public, internal, personal or sensitive.
- Systems to be consulted or updated and their permissions are known.
- There is a baseline: volume, time, errors, abandonment or workload.
- Human escalation criteria and context preservation are defined.
- There is a path to operate, correct, monitor and eventually retire the system.
04
An architecture that preserves accountability.
Layers may vary, but boundaries should remain clear: conversing, consulting, executing, recording and deciding are different responsibilities.
Residents and channelsWeb · messaging · documents
→ Agentic assistanceUnderstands · guides · coordinates
→ Authorized sourcesRetrieval · rules · tools
↔ TITAN records and workflowsEvidence · procedures · control
→ Conceptual architecture. The Municipal Agentic Assistance System and the TITAN Ecosystem are complementary, distinct layers. Each institution defines which ones it uses and which integrations it authorizes.
Least privilege
Each integration accesses only the operations and data it needs.
Safe degradation
If evidence is insufficient or a source fails, the system abstains or escalates.
Institutional ownership
The municipality retains authority over its data, rules and decisions.
05
How to evaluate a municipal AI provider.
Do not evaluate a demonstration alone. Ask for evidence about the complete system and what happens after launch.
CriterionQuestionExpected evidence
Operational evidenceCan the provider show a system used in real institutional processes?Public source, sanitized demonstration, references and explicit limitations.
GovernanceWho authorizes sources, actions, changes and decisions?Roles, permissions, approvals and escalation matrix.
InteroperabilityHow does it integrate with systems the municipality already uses?Documented interfaces, technical identity, tests and reversibility.
Data and privacyWhat data leaves, where is it processed and how long is it retained?Data inventory, minimization, retention and accountable parties.
Quality and safetyHow are answers, actions, attacks and provider failures tested?Test cases, abstention, logs, alerts and recovery.
Post-launch operationsWho handles errors and maintains knowledge and integrations?Owners, response levels, monitoring and change management.
PortabilityCan the institution retrieve its data and continue if it changes provider?Exports, documentation, formats and an exit plan.
06
Reference path: from need to operations.
The schedule is indicative. Data, procurement, integrations and authorizations may extend it. The goal is to advance based on evidence—not deadline pressure.
- 01
Reference weeks 1–2
Understand and bound
Map the service, actors, sources, risks, permissions and baseline. The output is a problem definition and acceptance criteria. - 02
Reference weeks 3–5
Design and test
Build the minimum workflow with controlled data, normal scenarios, exceptions and human escalation. - 03
Reference weeks 6–8
Pilot with accountable owners
Operate with a bounded group, record failures, correct sources and verify that actions respect their limits. - 04
Reference weeks 9–12
Move to governed operations
Approve scope, owners, monitoring, continuity, support and measurement. Expand only when evidence supports it.
07
Measure the change—not merely model usage.
A metric should include unit, formula, period, source, exclusions and cutoff date. “Messages processed” alone does not mean better public service.
Service
First response time · resolution · abandonment · channel availability
Quality
Grounded answers · correct abstentions · errors · corrections
Operations
Cases processed · manual work avoided · queues · exceptions · rework
Adoption
Residents served · active staff · recurrence · satisfaction
Governance
Auditable actions · incidents · escalation · correction time
BASELINE→OBSERVED CHANGE→CAREFUL ATTRIBUTION→CONTINUOUS IMPROVEMENT
08
Documented case: Puriscal and TITAN.
The Municipality of Puriscal’s official land-use page embeds the TITAN municipal form and publicly links to the product page. This demonstrates the presence of the entry point in an institutional operation; it is not, by itself, an independent impact measurement.
PRIMARY SOURCE
Municipality of Puriscal
Official land-use page with the “TicoNeural TITAN technology” reference.
09
Frequently asked questions before starting.
Should a municipality start with a chatbot?
Not necessarily. Conversation can be a useful interface, but the project should begin with an operational problem, its sources, owners and boundaries. In some cases the best first step is internal knowledge, documents, inspections or a digital case file.
Can AI approve procedures or make municipal decisions?
It may guide, prepare, verify and execute authorized tasks. Legal, financial, technical and discretionary decisions should remain with the competent authority and be documented.
Must existing systems be replaced?
No. An implementation can work around existing software through APIs, databases, files, events or authorized adapters. When a legacy system must be modernized, the migration requires its own controls.
How do you prevent AI from making up an answer?
Bound sources and tools, test retrieval, require references where appropriate, and define abstention and escalation rules. No design eliminates every error; continuous measurement and correction are therefore required.
How long does a municipal implementation take?
It depends on the process, data, integrations, authorizations and procurement. This guide presents a twelve-week reference path for a bounded capability, not a universal delivery promise.
What should be covered by the agreement or contract?
Purpose, scope, sources, roles, data processing, ownership and portability, service levels, security, testing, metrics, change management, incidents and exit. Legal wording must be validated for each institution.
10
Sources and scope.
This guide combines the author’s technical experience with public frameworks and sources. References are provided for verification and further reading; inclusion does not imply endorsement of TicoNeural by those organizations.
- 01Costa Rica National Artificial Intelligence Strategy 2024–2027MICITT · Costa Rica
- 02Framework for governing with artificial intelligenceOECD
- 03AI Risk Management FrameworkNIST
- 04Costa Rica Law 8968: personal data protectionSistema Costarricense de Información Jurídica
- 05Costa Rica Law 9986: public procurementSistema Costarricense de Información Jurídica
- 06Land-use service: public reference to TicoNeural TITANMunicipality of Puriscal
NEXT STEP
Start with one concrete municipal need.
TicoNeural can help map the process, identify sources and risks, and define a measurable first scope before committing to a larger implementation.