07. Single Window Government Architecture, RAG, and Public Finance Auditing

Published 2026-09-14

Core Theme: In public finance and auditing, what matters is not conjecture but the hard Financial Regulations (FR), the procurement guidelines, and unbreakable audit trails. How do you guarantee accuracy through RAG — the technology that connects your own government documents to AI?

Where does the Single Window idea come from?

The architectural model proposed in the previous chapter for reorganising a Divisional Secretariat is the Single Window Platform. It is worth knowing that this is not a novel idea invented for Sri Lankan administration, because that fact is itself an argument for adopting it.

The Single Window concept was born in customs and international trade. The World Trade Organization adopted it under Article 10.4 of the Trade Facilitation Agreement to dissolve the bureaucratic barrier facing international traders. Before it, a shipment arriving at a port required separate paperwork submitted separately to customs, the port authority, the health inspectorate, the standards institute and the import-export controller — each with its own forms, its own queue, and its own capacity to halt everything. The Single Window replaced all of it with one digital submission.

Notice how exactly that describes a citizen at a Divisional Secretariat. Separate paperwork to the land division, the pensions division, the social services division; the same identity card photocopied for each; any one of them able to stop the file. The problem is structurally identical, which is why the solution transfers — and why adopting it is importing a proven global pattern rather than experimenting on citizens.

Figure 7.1: The DSS Single Window Architecture — the 5 precision prompt vectors for building the Service Orchestration Layer.

Designing such a platform is not one prompt. It is five, each interrogating the design from a different angle, and the order matters: what it does, why it is worth doing, how the work flows, how the systems connect, and how the people are brought along.

Vector one: the "what" — the policy architect prompt

The first prompt establishes functional scope. The goal is to unify the siloed divisions — pensions, land registrations, social services — behind a single Service Orchestration Layer, built from four core modules: a unified citizen portal, an intelligent task routing engine, a compliance module enforcing state law, and executive dashboards.

The hard constraint in this vector is the one that determines whether the whole project succeeds: the system must not be a trade counter. It must be 100 per cent citizen-centric, organised around the citizen's life needs rather than the office's internal divisions. This distinction sounds abstract and is entirely practical. A portal organised by department requires the citizen to know which department owns their problem — which is precisely the knowledge they do not have, and the reason they are standing in the wrong queue.

Vector two: the "why" — the chain-of-thought justification

The second prompt builds the economic case, and it uses chain-of-thought deliberately, because a number without visible reasoning persuades nobody who has to approve a budget.

The goal is to calculate the waste of the traditional counter-based file system and establish the national productivity return. The reasoning chain runs: count the number of times a rural villager must ride a bus thirty kilometres because of duplicate data entry and hand-carried files; compute the lost daily wages and hours; and from that establish the direct value to the provincial economy of cutting office visits by seventy per cent.

The important move here is the choice of unit. The cost of the current system is not measured in office efficiency but in the citizen's bus fares and lost wages — a figure that is large, real, and never appears in any government account because nobody in the office pays it. Making that cost visible is what converts a convenience argument into an economic one.

Vector three: process logic — the operations manager prompt

The third prompt redesigns the workflow itself, converting a sequential file system into a parallel state machine. The worked example is the timber felling and transport permit.

The redesigned flow runs as follows. The citizen applies digitally. The application is verified against the National Identity Card database and a unified citizen profile is created. Digital review is then dispatched simultaneously to the Grama Niladhari, the Range Forest Officer and the land clerk, rather than passing from one to the next. The instant all approvals land, the digital permit is issued by SMS. And if any officer delays beyond 48 hours, an alert escalates automatically to the Divisional Secretary.

Two design decisions in that sequence carry most of the benefit. Parallel dispatch is what collapses the elapsed time, because the total is now the slowest single review rather than the sum of all three. And the escalation rule is what makes the improvement durable: without a timer, a parallel process silently degrades back into a sequential one as files wait in whichever inbox is busiest.

Vector four: technical interoperability — the system architect prompt

The fourth prompt addresses how distributed systems exchange data, and it carries an explicit prohibition: do not pile all government data into one monolithic master database.

That constraint deserves emphasis, because the monolith is the intuitive solution and the wrong one. A single database holding everything about every citizen is a single point of failure, a single target for attackers, and a single administrator's mistake away from catastrophe. The alternative is connection without consolidation — REST APIs, JSON data formats, OAuth 2.0 security credentials and webhooks, maintaining live links between the Registrar General's Department for births, local government bodies for permits and taxes, the Land Registry, and an identity proxy for the NIC. Each system keeps custody of its own records and answers questions about them. Nothing moves to a central vault.

Vector five: human adoption — the change consultant prompt

The fifth prompt addresses the factor that actually decides the outcome: "fifty percent of the success of any digital transformation rests not on the technology but on human adoption."

It has three parts. Dissolving staff fear means addressing junior clerks' fear of redundancy directly, by re-skilling them as digital auditors and citizen guides — a real role, not a reassuring phrase. Assisted digital kiosks placed in post offices and Agrarian Service Centres serve elders without smartphones, so that the digital window does not become a wall. And a phased rollout runs a pilot in days 1–30, staff training and regional expansion in days 30–60, and full launch in days 60–90.

The phasing is not caution for its own sake. A pilot produces the internal evidence that persuades sceptical colleagues, and it does so at a scale where the inevitable first mistakes affect thirty citizens rather than thirty thousand.

How do you make AI answer from your own circulars?

Ask a plain ChatGPT about a government procurement and it may answer with another country's rules mixed in — confidently, and without signalling that it has done so. What the officer needs is an answer grounded in Sri Lanka's specific Government Procurement Guidelines and Financial Regulations.

The technique for that is Retrieval-Augmented Generation, introduced in Lewis et al., NeurIPS 2020: Retrieval-Augmented Generation for Knowledge-Intensive NLP Tasks. The principle is simple enough to state in one line: stop asking the model what it remembers, and start asking it to read what you give it.

A worked example: two conflicting procurement circulars

Consider a situation every procurement officer recognises. Several circulars issued at different times govern the same procedure: one sets the procurement thresholds, and an amendment issued two years later changes the tender board approval limits. Which governs this purchase?

An officer working by hand would spend a day cross-comparing the two documents, and would be right to be nervous about the result. Inside a RAG workflow the work changes shape. You attach both official procurement circular PDFs to the chat window, and then issue a constrained prompt of the following form.

Role: You are a senior procurement and financial audit specialist of the
Government of Sri Lanka.
Context: Base your answer ONLY on the attached Government Procurement
Guideline (Document A) and the Financial Amendment Circular No. 02 of 2024
(Document B).

Task: Our Divisional Secretariat plans to purchase computer equipment
worth Rs. 3.5 million.
1. Which tender procedure must be called for this (Shopping, or National
   Competitive Bidding)?
2. What is the required composition of the Procurement Committee (DPC),
   and who is the approving authority?
3. Answer citing the exact paragraph numbers of the attached documents.
   Include no general knowledge from outside the documents.

Read that prompt against the five architectural layers from Chapter 05 and every one of them is present: a role, a context restricted to two named documents, three specific tasks, a prohibition on outside knowledge, and a required output format citing paragraph numbers.

The final instruction is the one doing the most work. Requiring paragraph citations does two things at once: it closes the model's room to fabricate, and it hands you a result you can verify in thirty seconds by opening the circular at the cited paragraph. An answer you can check quickly is worth considerably more than an answer that merely sounds authoritative — and the output is now an audit recommendation grounded solely in the real law of the state.

Free tier or paid tier: what is the difference for government data?

One security rule matters more than any other in institutional use, and it is worth stating before anything else in this section.

On a free tier, the terms of service of ChatGPT or any comparable tool generally grant the company permission to use every prompt and document you enter to train its next generation of models. The consequence follows immediately and is worth stating without euphemism: the moment a government letter is pasted into a free tool, that is a state data leak. It does not feel like one — nothing visible happens, no alarm sounds, the officer's work simply gets easier — which is exactly why the rule has to be institutional rather than left to individual judgment in the moment.

On a paid or enterprise tier, costing roughly US$15–20 a month — about Rs. 5,000 to 7,000 — three things change, per the OpenAI enterprise privacy commitments and the Anthropic Trust Center. You receive a contractual guarantee that your data is not used to train the provider's models. You gain access to the most capable current reasoning models rather than the cut-down free versions. And you receive far greater capacity for analysing large files — the PDFs and spreadsheets that most real government work consists of.

Set those two paragraphs beside each other and the institutional conclusion is not a close call. A paid account costs an institution less per month than a single officer's day spent cross-comparing circulars by hand, and it removes a category of legal exposure under the Personal Data Protection Act that no amount of staff training otherwise eliminates. If a government institution is to use AI at all, the head of the institution or the ministry should issue official paid accounts — securing data protection and superior productivity in one decision.

How does this connect to the National Digital Economy Blueprint?

The transformation happening on a Divisional Secretariat desk is not an isolated local improvement. It sits inside Sri Lanka's national Digital Economy Blueprint, and Figure 7.2 shows the six layers of that structure, running from physical infrastructure to the citizen.

Figure 7.2: The Sri Lanka National Digital Economy Blueprint — physical infrastructure, Digital Public Infrastructure (DPI), open APIs, vertical applications, multi-channel access windows, and the governance-and-trust framework.

Layer 1: physical digital infrastructure

The bedrock. High-speed broadband networks — 5G, FTTx optical fibre, SD-WAN and private government networks. The Lanka Government Network, the secure high-speed backbone stitching every ministry and Divisional Secretariat on the island into one network. And the compute: the Lanka Government Cloud, national AI data centres, and sovereign hybrid cloud environments.

Nothing above this layer functions where it is absent, which is why rural connectivity is not a separate social programme but a precondition for every service in the layers above.

Layer 2: Digital Public Infrastructure

The layer that changes what a transaction with the state involves, and the most consequential of the six.

The national digital identity, SLUDI, built on MOSIP, establishes citizen identity biometrically. Verifiable credentials and the E-Locker give the citizen a vault where birth certificates, examination results and land deeds are digitally signed and stored — which is what finally removes the photocopy from every counter. DigiSign provides legally valid digital signing under the Electronic Transactions Act. The National Federated Data Exchange is the secure highway along which institutions exchange data without consolidating it. And the G2P benefits rail credits welfare payments such as Aswesuma directly to the citizen's bank account, with no intermediaries.

That last item deserves its own sentence. Removing intermediaries from a benefits payment removes, at the same stroke, every point at which a payment can be delayed, reduced or conditioned by someone standing between the state and the recipient.

Layer 3: open APIs and the Industry 4.0 platform

Government ERP systems, microservices, AI app enablers, and the unified government payment system through GovPay and LankaPay — together with the open API standards through which the private sector and independent software builders connect securely to government services.

The word to notice is open. An API that outside developers can build against is what allows services the government did not think of to be created by people who are closer to the problem.

Layer 4: vertical sector applications

Where the infrastructure becomes something a person uses. On the government side: motor traffic, education, transport, immigration and emigration, Inland Revenue, local government and Divisional Secretariat services, the company registrar. On the industry side: manufacturing, agriculture with smart harvest planning, health with digital patient records, construction, tourism, ports and logistics, banking and finance.

This is the layer corresponding to Layer 5 of the AI stack in Chapter 02 — the only layer where economic value is actually generated, and the layer where Sri Lankan domain knowledge is the scarce input.

Layer 5: access channels and service delivery

Who reaches the services and how. Citizens, tourists, businesses, startup entrepreneurs, developers — and, increasingly, autonomous AI agents acting on someone's behalf. Through mobile apps, government web portals, self-service kiosks, and APIs.

The multiplicity is the point. A single channel, however well built, excludes whoever cannot use that channel.

The cross-cutting pillars: trust, governance and inclusion

These run vertically through all five layers rather than sitting on top, and a failure here invalidates everything above it.

The legal and governance framework comprises the Personal Data Protection Act, the Cyber Security Act, the Digital Economy Authority, and AI ethics governance frameworks. Social inclusion and capacity building cover affordable broadband for all, accessibility for differently-abled citizens, the Code4Lanka programme, and the island-wide AI literacy training delivered by Idasara Academy.

The reason these are drawn as cross-cutting rather than as a sixth layer is worth understanding. Trust is not a feature added at the end. A citizen who does not believe their data is safe will not use the digital channel, and a system nobody trusts delivers no benefit however well it is engineered.

Figure 7.3: Institutional Strengthening & Capacity Building (RECAP) — developing public service capacity through institutional strengthening, skills intelligence, sustainability, digital literacy, and certification.

Figure 7.3 names the part of this that cannot be procured. Institutional strengthening, skills intelligence, sustainability, digital literacy and certification are all investments in people, and they are the reason a well-funded digital programme can still fail. Infrastructure can be bought and installed in a budget cycle. Capability has to be built in the officers who will operate it, and that runs on a slower clock than any procurement schedule.

In the next chapter, we examine the seven gravest ethical and legal risks a public officer faces when using artificial intelligence, the $440,000 lesson Deloitte suffered in Australia, and the Shadow AI that officers are already practicing in secret.