Buy Intelligence, Build Governance

By Anushka Appala and Dr. Janio Rosales

One of the most consequential decisions in a government AI project is also one of the least discussed: what do you build, and what do you buy? Get it wrong and you either burn a year training a model you did not need, or you hand a foreign vendor a quiet veto over your public administration. NaciluzIA takes a deliberate position on both halves. Buy the intelligence. Build the governance. The reasoning matters, because it is where a lot of well-intentioned projects go astray.

Buy the intelligence — and keep it replaceable

The general-purpose reasoning that lets a system understand a request in Q'eqchi', translate it, and draft an acuerdo is now a commodity, and a fast-moving one. Enterprise foundation models — Claude, OpenAI, Google, Azure — are improving faster than any single government could match, and they are reachable through secure APIs today. Trying to build a frontier model in-house, in year one, to do work you can rent at higher quality would be a waste of scarce public money and scarce local talent.

So NaciluzIA consumes those models through a secure API layer, and treats them as interchangeable. The platform is provider-agnostic by design: the model behind the advisory layer can be swapped from one vendor to another without touching the rules engine, the ledger, or the commitment gate. This is not a technical footnote — it is the difference between using a vendor and depending on one. A model that raises its price, degrades, or falls out of compliance can be replaced; a model woven into the load-bearing logic of the state cannot.

The reason this is safe to do is the architecture we have described elsewhere: the model never has authority to decide. It proposes; it drafts; it flags. Because a bought, swappable model has no power to approve a benefit or authorize a payment, the usual anxiety about "vendor lock-in over critical decisions" simply does not apply. There is no critical decision for the vendor to capture. You are renting cognition, not sovereignty.

Do not train a giant neural network in year one

The instinct to train something bespoke is understandable and, for the work that actually binds citizens, wrong. The two jobs NaciluzIA has to do well — grounding advice in the real law, and detecting fraud in administrative data — are both jobs where a large, freshly trained neural network is the inferior tool.

For grounding, the answer is retrieval, not training. The advisory layer is anchored by RAG — retrieval-augmented generation — over cited Guatemalan law, with the legal corpus indexed in a vector store (pgvector) so that every draft points back to the statute it rests on. Retrieval keeps the model's output tethered to text a human can check, and it updates the moment the law does. Baking law into model weights would do the opposite: it would make the law hard to inspect and slow to change, which is exactly backwards for a public institution.

For fraud, the honest engineering answer is that classical machine learning wins. On the tabular, structured data that welfare and procurement systems actually produce — records, amounts, dates, identities, relationships — a deep neural network is neither the most accurate nor the most auditable choice. Record linkage finds the duplicate beneficiaries and the same person under three names. Gradient-boosted trees are, in practice, the strongest performers on this kind of tabular data, and they yield feature-level explanations an investigator and an auditor can read. Graph analytics exposes the rings — the shared addresses, the clustered payments, the supplier that keeps winning under different names. None of these hallucinate, all of them explain themselves, and each surfaces evidence for a human to act on rather than a verdict to be trusted.

There is real money behind getting this right. Latin America loses an estimated 4.4% of GDP — roughly US$220 billion a year — to procurement, payroll, and transfer leakage (IDB, Better Spending for Better Lives, 2018), and India's experience shows the recoverable upside: linking welfare payments to verified digital ID cut leakage by 12.7% (BCG, December 2025). You capture value at that scale with detection that stands up to scrutiny, not with a black box.

Build the governance — because that is the part no one can sell you

If intelligence is bought and detection is classical, what exactly does NaciluzIA build? The part that binds. The deterministic rules engine that executes published law is plain, ordinary, deterministic code — versioned, reproducible, and readable, so that "approve if the DPI is valid, the person is living, there is no duplicate benefit, and the threshold is met" runs the same way every time and can be audited line by line. The commitment gate that refuses to act without both a rule result and a human signature is ours. The hash-chained ledger, valid under Decreto 47-2008, is ours. The four-role discipline — AI recommends, policies decide, officials oversee, audits verify — is ours.

These are not things you can license from a vendor, because they are not products. They are the constitutional wiring of how a decision gets made and recorded, and they have to belong to the public. That is why NaciluzIA is an open standard, stewarded by a nonprofit rather than owned by a company: the intelligence can come from anywhere and change over time, but the governance is the sovereign part, and it stays with Guatemala.

Buy the intelligence, because it is a commodity that gets better while you sleep. Build the governance, because it is the one thing that must never be for sale.

Cada decisión, a la luz.

— Anushka Appala and Dr. Janio Rosales