01. AI for Transforming Public Service: Concepts, Applications, and Guidance for Public Officers

Published 2026-09-14

Core Theme: Artificial intelligence is not a passing technology fashion, nor a software vendor's sales pitch; it is a national reality that has already arrived on your desk. What it takes from you are the repetitive tasks that exhaust you — never your post, and never your human accountability.

Why does this begin at the public officer's desk?

Sri Lanka's public service is the lifeblood of the nation. From the Grama Niladhari division — the village-officer level, the state's smallest administrative unit — up through the Divisional Secretariat, the District Secretariat and the ministries, the lives of thousands of ordinary citizens are decided every day amid the files stacked on public officers' desks. Land permits, birth certificates, Aswesuma and Samurdhi welfare benefits, pensions, licence renewals and requests tied to development projects arrive without pause.

The traditional file-based way of working is now in deep crisis, and it is worth being precise about the nature of the crisis, because it is not laziness or indifference. It is arithmetic. With limited resources, limited staff and a documentary burden accumulated over generations, an ordinary citizen can spend days — sometimes weeks — shuttling between offices for a single simple service. On the other side of the counter, the officer spends the entire day re-checking the same information, filling the same forms by hand, and combing shelves for old circulars.

Notice that both sides of that counter are losing. The citizen loses days of wages and dignity; the officer loses the entire working day to mechanical work that no part of their training was for. Neither outcome is anybody's intention, which is exactly why exhortation has never fixed it. What changes the arithmetic is a tool that absorbs the mechanical portion — and that is what generative AI now offers.

Figure 1.1: Samisa Abeysinghe (Founder & CEO, Idasara Digital) — leadership for the digital transformation of the public service and for artificial intelligence literacy.

This programme began in April 2025, at the invitation of the Presidential Secretariat, as an AI training initiative for public officers. In the year since, it has travelled well beyond Colombo — to Jaffna, Kandy, Matara, Kalutara, Hambantota and Kataragama — reaching more than 10,000 public officers. What follows is the tenth edition of that material, Version 10.0.

The aim is not to intimidate anyone with technical jargon. It is to show, from an engineering and practical standpoint, how a public officer can use artificial intelligence to make the daily workload radically more efficient, how to serve the citizen with greater dignity, and which safeguards are not optional along the way.

What do officers actually say when asked about AI?

Every workshop in this series opens with the same three questions put to the officers in the room, and the answers are consistent enough across the island to be treated as findings rather than impressions.

1. Do you already use ChatGPT, Google Gemini or a similar AI tool on a weekly basis?
2. What do you use it for? Personal needs, your children's education, drafting official letters, translation?
3. What is the single greatest fear or uncertainty about AI sitting at the bottom of your mind?

Three patterns emerge from the show of hands every time.

A majority already use it privately. Most often for their children's schoolwork, to correct an English letter, or out of personal curiosity. The adoption has already happened; it simply happened without anybody deciding it should.

Officially, it remains a secret. Many officers quietly draft office letters with ChatGPT on their personal phones. Not out of laziness — out of exhaustion from a crushing workload. But because there is no institutional policy and no proper training, they do it with a lingering sense of guilt.

That combination deserves attention, because it is the worst of the available situations. The institution gets none of the benefits of deliberate adoption — no guidance on what data may be entered, no shared prompt library, no verification discipline — while carrying all of the risk. And the officers doing the most to keep up with their workload are the ones made to feel they are doing something wrong. Chapter 08 returns to this as Shadow AI, and it is a policy failure rather than a discipline problem.

A deep fear runs through the whole conversation. "Will this take our jobs? If I use it, will our confidential data leak? Will the computer be hacked? Are our bank accounts at risk? Will machines rule the world?"

This fear has a specific and recent local cause, and dismissing it as ignorance would be a mistake. Many Sri Lankan parents put their children through information technology degrees expecting well-paid careers. After 2023, even securing an internship became extraordinarily difficult for those graduates. The suspicion that follows — "did AI do this, and will the same thing happen to our jobs?" — is not irrational. It is an inference from something the officer watched happen to their own family.

So the fear deserves a direct answer rather than reassurance.

Will AI take your job?

The honest answer, with nothing concealed: artificial intelligence will never seize your post. It will inevitably take over certain tasks on your desk — and specifically the monotonous, rule-bound, mechanical portion you already complain about, the portion that wears you down.

The distinction turns on a fact about what a post actually is. Any repetitive task that can be performed digitally on a computer can now be performed by an AI model a hundred times faster than a human. But a post is not a bundle of repetitive tasks. A post is a responsibility, a capacity for judgment, and a grasp of context — and none of those three has been automated.

The dividing line is sharp enough to tabulate, and worth keeping where it can be seen.

Read down the right-hand column and a pattern emerges. Everything there requires either physical presence, human feeling, or legal personhood — three things a model has none of. The empty cell in the left column is the most important space in the table: legal accountability has no machine counterpart, and it is not a gap waiting to be filled by a better model. A computer cannot be held answerable for an act of state, so the signature stays with a person by necessity rather than by policy.

There is a harder implication in this, and it is fair to state it. The officer whose working day consists almost entirely of the left-hand column will find that day changing substantially. The response to that is not to defend the mechanical work but to move deliberately toward the right-hand column, which is where the public service was always supposed to be spending its attention.

Which AI models matter, and who built them?

Several tools are now part of the working landscape, and before using any of them one question should come first: which country built this model, which company, and on what data was it trained?

That question is not idle curiosity. As Chapter 03 shows in detail, the environment in which a model was built determines its language ability, its grasp of our culture and its biases — which means the right tool for a Sinhala letter is not necessarily the right tool for a legal analysis.

Figure 1.2: Anatomy of a Frontier AI Interface — System Prompt, Context Buffer, Decoding Mode, and Multimodal Ingest (the 4 foundational controls every officer must command).

Reading a modern AI interface

A current generative AI interface is not a simple chat box. It is closer to an instrument panel, and most users operate perhaps a fifth of it — usually because nobody explained what the rest of the controls do. Figure 1.2 has three zones worth knowing.

The navigation sidebar, on the left, handles memory and separation. Instant internal search (⌘K) retrieves any previous conversation, draft or saved item semantically rather than by exact keyword. The mode selectors (Chat, Voice, Imagine) move between typing, real-time two-way voice conversation and image generation. History continues an earlier conversation without losing its accumulated context.

The control in this zone that matters most for institutional use is Projects. It keeps different domains of work — "Land Registrar Circulars" in one workspace, "Establishments Code Queries" in another — genuinely separate, each with its own documents and standing instructions, so that context from one never contaminates the other. For an office handling several unrelated categories of file, this single feature is the difference between a useful assistant and a confused one.

The grounding canvas, in the centre, is where the answer appears, and it carries evidence about how the answer was produced. The search marker — Searched for "famous haiku poems examples" — 10 results — indicates that the model did not rely on its static training weights alone but swept the live web, read what it retrieved, and composed its reply from that verified material. Learning to look for that marker is a practical verification habit: its presence tells you the answer has sources you can check, and its absence tells you the answer came from memory.

The response itself arrives structured rather than as an undifferentiated block of prose — headings, the governing rules, the cultural context. That structure is not decoration; it makes the answer auditable, because you can see which part of it to challenge.

The command strip, at the bottom, is where the least-used and most valuable controls live. Extended reasoning mode (Think Harder) is the most significant recent shift in these systems: for a complex legal or financial question, it forces the model to work through many reasoning steps internally before answering, rather than replying immediately. For any question with real consequences, this is the setting to use, and it is routinely left off.

Alongside it, multimodal document input (the attachment icon) feeds scanned circulars, PDF reports, photographs or spreadsheets directly to the model for analysis — the mechanism behind almost every serious application in this book. Automatic model selection (Auto) routes simple queries to fast, cheap models and complex ones to powerful reasoning models, which is cost control performed for you. And voice input allows instructions to be spoken rather than typed, which matters more than it sounds for officers who are faster in Sinhala speech than in Sinhala typing.

Seen properly, this is not a toy. It is an instrument panel for file inspection, circular comparison and decision support.

ChatGPT (OpenAI, USA)

ChatGPT is the most widely used model in the world and the one that started the conversational era. It is exceptionally strong in English and carries the broadest ecosystem of surrounding tools — Advanced Data Analysis, Custom GPTs. It holds simple conversations in Sinhala, but reaches its limits with complex Sinhala grammar, which is worth knowing before entrusting it with a formal Sinhala letter.

Google Gemini (Google, USA)

Gemini is critically important for the Sri Lankan public officer for one specific reason: Google has indexed Sinhala and Tamil websites, news and search data for decades, and that corpus feeds the model. Its grasp of Sinhala is correspondingly broader and more accurate than its American peers'.

It also offers an unusually large context window — the capacity to take in documents amounting to millions of words in a single session. For the officer who needs an entire act, a full set of amendments, or a year of circulars considered together, this is the practical capability that matters.

Claude (Anthropic, USA)

Claude is the model most trusted among the world's senior technical and engineering teams. Its strengths are long legal documents, complex report analysis, sustained reasoning and principled behaviour under its Constitutional AI approach. For serious state-level policy work and drafting analysis, it is the best-suited of the current tools — which is why Chapter 08's verification pipeline places it at the drafting stage.

DeepSeek (DeepSeek AI, China)

DeepSeek is the Chinese open-source model that demonstrated top-tier mathematical and reasoning ability at a fraction of the usual compute cost. Trained on complex Chinese character patterns and broad non-English multilingual data, it shows a distinctive strength with Sinhala and other structurally complex Asian languages — a surprise that Chapter 03 explains.

Meta AI / Llama (Meta, USA)

Llama is the world's leading open-weight model, and its significance for government is sovereignty rather than raw capability. An institution can install and run it on its own internal servers, which means confidential data never leaves the building. For any application where Chapter 08's privacy constraints would otherwise rule out AI entirely, this is the option that makes it possible.

Grok (xAI, USA)

Grok connects directly to the X platform and analyses global news and public conversation in real time. Its value is currency rather than depth — useful where what matters is what is being said right now.

How does this connect to the national plan?

Many officers reasonably assume an AI workshop is a software company's sales pitch. It is not. This is a Government of Sri Lanka programme reaching your desk, and it sits inside a structure worth knowing by name.

The National Digital Economy Blueprint is the government's foundational vision for digitalising the country's economy — the document the rest of this sits under.

PIC-Net, the Public-Impact Champions Network, is the national network turning officers across 95 ministries, state enterprises and provincial councils into champions of digital transformation. Its design assumption is worth noticing: transformation spreads through officers who have done it, not through directives.

The 2026 Budget makes a historic allocation for Digital Public Infrastructure — the MOSIP-based national digital identity, SLUDI, and the National Data Exchange for sharing data between government bodies. This is the layer that will eventually remove the photocopied identity card from every transaction.

And then your division, and your desk. However many cabinet decisions and central bank reports are written nationally, a national plan is won or lost at the counter. To the ordinary citizen, you are the government — not the blueprint. The country moves forward precisely as far as these tools are used to solve the citizen's actual problems, and no further.

What is the Idasara mission?

The founding vision of Idasara Digital is economic empowerment through digital empowerment, and the order of those terms is deliberate.

For the people of a country like Sri Lanka to rise economically, digital literacy and modern skills come first. That is the reason Idasara Academy exists, and it is not a conventional tuition class or lecture hall. Its core is Learn to Learn: how to find knowledge for yourself, how to learn from your own mistakes, and how to build genuine workplace capability through the Employability content stream. The distinction matters in an era when the specific tools worth knowing change every eighteen months — what transfers is the capacity to learn the next one unaided.

Nor is there any reason for anxiety about the technical diagrams in this book being in English. The English Learn Ladder framework lets anyone build language capacity step by step, starting from exactly where they stand today.

The public service is a nation's largest human resource. An officer who arms themselves with these tools is not improving their own productivity alone; they are raising the capability of the institution that every citizen meets when they meet the state.

In the next chapter, we examine in depth how generative AI shattered the complex coding barrier that stood for seventy years — since the birth of the computer itself.