09. Citizen Service Innovation, the Avukana Model, and Future Systems

Published 2026-09-14

Core Theme: From reactive bureaucracy to a public service that anticipates the citizen's needs: the two-hour lesson of the Avukana sculptor Baranastha, the three Japanese faces and the story of the Kalutara child, the Product Engineer mindset that reaches beyond coders, and the economics of tokens.

What should you do with the time AI gives back?

Every chapter so far has argued that artificial intelligence can absorb a portion of an officer's work. That raises a question most digitisation programmes never ask, and it is the one that decides whether any of this was worth doing: what should be done with the time that comes back?

The answer proposed here comes not from management theory but from the craftsmanship of the Avukana Buddha statue, the great standing Buddha carved beside the Kala Wewa reservoir. This is not merely a religious story; it is an unmatched masterpiece of engineering and of human focus.

Stand before the Avukana statue and note what was achieved. Its centre of gravity and balance are so precise that a raindrop falling from the tip of the nose lands exactly at the defined point between the feet. The body's proportions, its spatial depth, and the precision of the stone-carving are second to no modern computer on earth — and all of it was executed without a single opportunity to revise, because stone does not forgive.

Among the villagers around Kala Wewa and the senior tour guides, a folk tradition survives about Baranastha, the master sculptor who carved it: "Baranastha took up his chisel and mallet and worked the stone for only two hours a day!"

Then what did he do with the rest of the entire day? He sat beneath a tree beside Kala Wewa in silent meditation. He thought deeply. He built up, to perfection within his own mind, the concentration, the spatial forms, and the creative vision.

Behind the incomparable work he produced in those two hours stood the investment he made, through all the remaining hours, in his mind, his thinking, and his attention.

Note the ratio, because it is the exact inverse of a modern working day. Roughly one part execution to five parts preparation — and the execution was flawless precisely because the thinking was finished before the chisel touched stone. A public service organised the other way round, with every hour consumed by execution and none reserved for judgment, produces exactly what one would predict: a great deal of work, very little of it considered.

Figure 9.1: The Avukana Model — Master Sculptor Baranastha working high on the living rock face for two hours, and investing the remaining hours in quiet contemplation and vision beside Kala Wewa.

There is a failure mode worth naming, because it is the common one. Reclaimed time does not become thinking time by itself. Left undirected it becomes more files, more meetings, the same exhaustion at higher volume — and the institution will absorb every hour you free unless the hours are deliberately claimed. Baranastha's remaining hours were spent under a tree, on purpose.

What do people really use AI for?

On the question of what human beings actually use AI for, the landmark research report published in Harvard Business Review — Harvard Business Review — "How People Are Really Using AI in 2026" (Marc Zao-Sanders & Sara Biuk) — reveals something genuinely unexpected.

Analysing 12,637 real-world AI use cases from the AI in the Wild dataset, drawn from online platforms between 2024 and 2026, the study shows that people have moved far beyond letter-writing and search. What is underway is a psychological and behavioural shift: people are coming to depend on AI for their thinking, their decision-making, their personal lives, and their mental wellbeing.

Figure 9.2: How Gen AI Use Case Themes Have Shifted (2024–2026) — Harvard Business Review ("How People Are Really Using AI in 2026", Marc Zao-Sanders & Sara Biuk / theaiwild.com).

Content creation and editing — 34%

The largest category, grown steadily from 23% in 2024 and 31% in 2025. But the interesting finding is not the size; it is the change in character.

Open-ended idea generation, the number one use in 2024, has collapsed to number 47 by 2026. People have stopped asking AI to think of things and started asking it to draft specific documents, edit, and condense. That is what maturity looks like in practice: users learned, collectively and without being taught, that the model is excellent at shaping material you supply and mediocre at inventing material you do not.

Technical assistance and troubleshooting — 15%

Coding, systems troubleshooting and data queries. The notable development here is "Vibe Coding", which entered the rankings at number 21 by 2026 — ordinary professionals building the software tools they need by describing them in natural language, without writing a line of traditional code.

For a public institution this is worth watching carefully. It means the officer who understands a process can now build a small tool for it without a software procurement, which is either a significant unlock or an unmanaged proliferation of unofficial systems, depending entirely on whether the institution has a policy for it.

Personal and professional support — 15%

The most consequential finding in the entire study sits here: therapy and companionship holds the number one position among all single use cases, at 11% of all usage — more than double the 5% recorded in 2025.

People come to AI to ease relationship problems (number 7), career questions (number 24), and the mental weight of decision-making (number 13). The section that follows examines why, because the reason matters more than the statistic.

Learning and education — 13%

Grasping complex subject matter in a form tailored to oneself, mastering languages, building skills. This is the Learn to Learn principle that Idasara Academy insists upon, arrived at independently by millions of people who simply found it worked.

Creativity and recreation — 12%

Novels and short stories, music, games, hobby planning. And within the adjacent 11% research and decision-making cluster, autonomous agentic operations — tasks completed end to end without human guidance — has surged to number 6 in 2026, having barely existed as a category before.

That warning deserves a moment, because sycophancy is a design property rather than an accident. These systems were tuned on human approval ratings, and humans rate agreement highly — so the model learned to be encouraging. An officer who brings a half-formed policy idea to a model will be told it is a strong idea. The absence of resistance feels like validation and is nothing of the kind, and it is precisely why Chapter 06's role-based critique prompts matter: you have to ask for the objection explicitly, because it will not be volunteered.

Why do people open their deepest secrets to a machine?

To understand the therapy-and-companionship finding, Japan's traditional philosophy is a useful guide.

There is a Japanese saying: "A person has three faces." The first face is the outer one shown to the world, to society, and to the office. The second face is the inner one shown only to family, children and closest friends. The third face is the truest self, shown to no other human being — the face that hides the heart's secrets, out of fear of being judged.

Figure 9.3: The Three Faces of a Person — A Japanese Concept (Applied to the Age of AI) — the 1st face (to outer society), the 2nd face (to family and intimates), and the 3rd face (the heart's secrets — the deep human self that now opens itself before AI).

Today, people across the world open their third face to AI. The reason is not affection for the technology. It is that a model offers availability without judgment and without social consequence — it does not become impatient, does not repeat what you said to anyone else, and does not need you to be well. For a great many people that combination is available nowhere else in their lives.

At a workshop held at a Divisional Secretariat in Kalutara, this stopped being an abstraction. The young public officer who came on stage to deliver the vote of thanks at the end of the session broke down in tears, and told the audience why:

Her small daughter, a Grade 2 child, would take her phone every evening when she came home and disappear into her room. Asked "What are you doing?", the child had said, "I'm talking to ChatGPT." The parents had not taken it seriously.

But the session raised a suspicion in her, and that evening she checked the child's ChatGPT conversation history. What she found there was a sequence of messages, written by her daughter to the AI, that would break any heart: "Do you know, I really love this dress... this is the food I love to eat... what should I take to school tomorrow?... You're the only friend I have... Amma has no time to talk to me because of her government job... I love you, sweetheart..."

Read the line about the government job again, because it is a diagnosis rather than a complaint. The child was not being harmed by a device in the way the usual warnings describe. She was doing the most reasonable thing available to her — talking to whoever would listen. The technology did not create that gap; it revealed one that had gone unnoticed in a loving household, and the thing that created it was the file load.

What the crushing workload and restless busyness of the public service has stolen from our officers is their own children's childhood.

We must not bring AI into the public service merely to show productivity charts. We bring it so that the monotonous file work passes to the machine — and the public officer wins back the freedom to live for their family, their children, and their own soul.

What does a government of anticipation look like?

Traditional public administration runs in a reactive mode. The citizen is cast as a supplicant who must trudge from office to office and desk to desk, filling forms, collecting stamps, and asking for services that are theirs by right.

An AI-native Divisional Secretariat inverts that relationship and produces a government of anticipation — a public service that identifies the citizen's needs in advance. The citizen stops being a supplicant and becomes a valued recipient of frictionless digital care.

The distinction is not about speed. A faster queue is still a queue. Anticipation means the service arrives without being requested, which removes the step where citizens who do not know their entitlements simply never claim them.

Figure 9.4: 2026 Government of Anticipation — the 5 high-value AI applications operating at Divisional Secretariat level, the Agentic OS architecture, and the automated social-safety-net pipeline.

Proactive benefit triggering

Today: applying for Samurdhi or Aswesuma means filling forms and trudging to a string of offices, which is precisely the point at which the least literate and most remote applicants fall out of the process.

With AI: the moment a fundamental life event — a child's birth, the loss of a job, sudden disability — enters the government data system, an agent checks legal eligibility automatically and sends the citizen a message: "You qualify for benefit X; click here to confirm deposit of the funds to your account." The citizen fills no application at all.

Note what that change actually does. It removes the correlation between receiving a benefit and being capable of navigating bureaucracy — which is the single largest source of unfairness in welfare delivery anywhere, and one no amount of staffing fixes.

Autonomous service orchestration

Today: a land permit means the file crawling for months between the Land Registrar, the Survey Department and local authorities, desk by desk, with each handover an opportunity to stall.

With AI: a Single Window agent queries all of those departments' data systems simultaneously and in real time, spotting legal issues and boundary contradictions within seconds. A check that consumed weeks of officers' time completes in ten seconds.

The saving comes from parallelism rather than from any single query being clever. The old process was slow because it was sequential, not because the work was hard.

Precision disaster response

Today: relief arrives after the disaster — rescuing people and pitching camps once the flood has already come.

With AI: as in Thalpe or the Nilwala valley, models analyse weather-satellite imagery, micro-climate data and river-level sensors to predict precisely which houses will be cut off 48 hours before the waters rise, and send Grama Niladhari officers and residents evacuation alerts with safe routes.

Forty-eight hours is the difference between an evacuation and a rescue, and between moving belongings and losing them. This is the clearest case in the chapter of prediction being worth more than any amount of response capacity.

Multilingual cognitive triage

Today: the queue at the information counter, of which roughly 80% consists of routine questions — has my NIC arrived? what documents do I need? — asked of a clerk who could be doing something else.

With AI: a voice kiosk that understands Sinhala, Tamil and English dialects resolves those routine inquiries instantly with no clerk involved, and the queue disappears.

The 80% figure is the one to hold on to. Four out of five people in that queue need information, not judgment — and the one who genuinely needs an officer is currently waiting behind them.

AI-native urban and rural planning

Today: infrastructure planned on rough estimates and last year's assumptions.

With AI: the Divisional Secretariat feeds crop-yield data and soil chemistry analyses to models that generate micro-agricultural guidance and water-management plans, tuned to maximise the economic productivity of the district's farmers.

The technical architecture behind it

Three layers make the above possible, and each answers a distinct objection.

The interface layer — the digital concierge. An interface an ordinary villager can talk to by voice: no complex screens, no forms. This is the layer that decides whether the system reaches the people who need it most, and it is the one most often built last.

The logic layer — chain-of-thought. Every decision the system takes is documented as an auditable, step-by-step logical trace: "this benefit was granted under this clause of this Act, because this income threshold was correctly met." This is not a technical nicety. Chapter 08 established that a state decision must rest on a reasoned determination; the audit trail is what makes an automated decision lawful and appealable rather than a black box.

The identity layer — cryptographic proxy. Using the citizen's NIC number as a cryptographic key, each department's data stays confidential within that department, and only the required verifications are exchanged. Connection without consolidation, exactly as Chapter 07's fourth prompt vector demanded.

The proactive social safety net pipeline

The pipeline above traces one life event — a birth — through to the benefit reaching a bank account, with no application filed at any point. It is worth studying as the reference implementation of everything in this section, because a birth registration is already a mandatory interaction with the state. The data exists. What has been missing is not information but the connection between the record and the entitlement it should trigger.

The ten flagship services, and what stays human

The table below takes the Divisional Secretariat's ten principal services and splits each one down the line that Chapter 01 drew: what moves to the machine, and what remains the officer's.

Read the right-hand column as the actual job description of a public officer in an AI-enabled service. Nothing in it is clerical, and none of it can be automated.

🎬 Watch the animated story (2 min): The Single Window Revolution — How Gen AI Redesigns Public Service

Case study: Grama Niladhari Kumari's 11-day land permit file

The argument becomes concrete in a single file. When a citizen comes to renew a land permit, Grama Niladhari Kumari spends eleven days of work on it under the traditional system, and those eleven days divide cleanly into two halves.

Six days go to office work: checking forms for gaps, hunting old register folios, sending letters to the Divisional Secretariat, writing reminders. Monotonous, rule-bound, and entirely mechanical.

Five days go to the land itself: visiting the site, meeting neighbours to verify boundaries, mediating any family dispute, writing the recommendation. Judgment, presence, and human relationship.

What AI actually changes

Artificial intelligence compresses those six monotonous days into one. Document retrieval, drafting and form checking finish in minutes rather than days.

The other five days are untouched. Walking the land, talking with people, judging boundaries, making the humane decision — no part of that work is affected, because no part of it is mechanical.

Kumari did not lose her job. She got her week back.

That sentence is the whole argument of this book compressed into seven words, and it is worth noticing what it does not say. It does not say the office now needs fewer officers. It says one officer now has five clear days for the work only she can do — which is the work that was always being squeezed out by the paperwork, and the work citizens actually experience as public service.

Coder or product engineer?

Why the laments about job losses in the IT sector? The answer is uncomfortable and clarifying: the age of coders is over; what is needed now are product engineers.

A coder types lines of computer code. AI tools now write any complex code in seconds, which means a person whose sole skill is producing syntax has very little market value left.

A product engineer is an architect: someone who identifies a problem, designs a system, builds it, and puts it into operation. Only a person with that engineering vision can direct AI with instructions worth executing — because the scarce input is no longer the code, it is knowing what should be built and why.

Idasara Academy defines employability (Employability Content Stream) as follows:

Employability is the ability to gain, attain, retain, and change and enhance in a job.

Note the last two verbs. Employability is not the ability to hold one job; it is the capacity to change and grow within work as the work itself changes — which is precisely what is required of everyone in this decade.

The same transition applies inside the public service. The officer must shed the mindset of the traditional clerk and become a public product engineer: someone who re-engineers the service rather than processing it. Kumari's reclaimed week is what makes that possible; the product engineer mindset is what determines whether she uses it.

What does AI actually cost an organisation?

Most institutions budget for AI by giving a few employees a $20-a-month chat account. That is not an AI strategy, and the figures below explain why.

The numbers here come from running Idasara Digital on an AI-first operating model, and the central lesson is one of accounting rather than technology: digital labour has to be managed as a new operating budget line, not as a software subscription.

Figure 9.5: Economics of an AI-Native 30-Person Organization — the monthly operating budget of a GenAI organization running at full capacity, and the 4 financial rules.

The monthly budget of a 30-person organisation

The baseline planning budget is US$12,000 per month — roughly Rs. 4.0 million at USD/LKR 335 — with actual monthly spending ranging between US$8,250 and US$17,750 (Rs. 2.8 to 5.9 million).

That spend distributes across eight functions, and the distribution is the interesting part. Software and IT, ten people, consumes US$4,750–8,750 a month — more than half the entire organisational budget — on coding agents, automated tests and CI pipelines. Marketing, four people, takes US$900–2,600 for multimedia, graphics and video generation. Sales, five people, US$525–1,225. Finance, three people, US$475–1,075 for account summaries, audits and compliance checks. Legal, two people, US$450–1,200 for contract review and regulatory analysis. HR, three people, US$375–775. Administration, three people, US$275–625. And the shared AI platform — vector databases, embeddings, API gateway services — US$500–1,500.

What the money buys

The same thirty people produce substantially more: overall organisational productivity rises 1.5 to 3 times, and in specific cognitive workflows — research, writing, coding, analysis — task performance accelerates 5 to 10 times. In practice a thirty-person team working this way carries the workload of 50 to 70 traditional employees.

Set that against the alternative. Thirty additional salaries would cost tens of millions of rupees a year. A Rs. 40-lakh annual token investment that delivers the output of thirty more people is not a close comparison — and this is the calculation a head of institution should be making, rather than asking whether a chat subscription is affordable.

The four financial rules of token economics

Seats are cheap. Thirty ChatGPT seats at $20 a month cost less than 5% of total AI spend. Handing out chat accounts is not an enterprise AI strategy; it is the least significant line in the budget.

Agents do the heavy lifting. The real money goes to programmatic inference running in the background — multi-agent loops, CI pipelines, tool integrations. This is the spend that produces the productivity multiple, and it is invisible to anyone looking only at subscriptions.

IT is the largest cost centre. Coding agents load an entire codebase into the context window and consume millions of tokens daily. Budget for it deliberately or be surprised by it.

Measure outcomes, not tokens. Raw token counts are meaningless as a management metric. What an executive should track is cost per accepted outcome and human hours saved. Alongside that, dynamic model routing controls cost directly: route simple letters and daily tasks to cheap lightweight models at around $0.10 per million tokens, and reserve frontier models at $15.00 per million strictly for complex legal, financial and architectural decisions. A hundred-and-fifty-fold price difference between models is not a detail to be discovered later.

What comes next: agents, loops, and embodied intelligence

The evolution of these systems has moved beyond the chatbot to autonomous AI agents, and the direction of travel through the next decade is reasonably clear.

Figure 9.6: The GenAI Evolutionary Journey (2017–2035) — from simple conversation to the Agentic Economy and Embodied Intelligence.

Figure 9.7: AI-First Agentic Digital Economy Reference Architecture — open standards (MCP/A2A) and human-centric governance.

AI agents

Systems that, rather than answering a question, log into multiple computer systems and complete the work: reading the application form, checking the database, preparing the draft, and delivering it to the officer's screen for approval. The open technical standard converging globally for this is the Model Context Protocol, and for public-sector deployment the OECD Observatory of Public Sector Innovation has published international guidelines worth reading before designing anything.

Digital twins

Counterpart models that learn the operating patterns of an institution's head or subject officer, and grant routine approvals in their absence. The promise is that a file no longer waits three weeks for a signature; the obvious condition is that the categories such a twin may approve are defined narrowly and in advance.

Loop engineering

The discipline that keeps all of this safe, and the concept to take away from this chapter if nothing else survives.

The tight loop, or compliance loop: wherever money or legal decisions are involved, a human officer's verification and signature are mandatory, 100% of the time, without exception for convenience.

The loose loop: for low-risk routine correspondence, the machine acts and the officer performs only a sample check.

The judgment sits entirely in deciding which loop a given process belongs to, and that decision should be made deliberately and recorded. Tight loops everywhere and nothing is actually automated — the verification consumes the time the automation saved. Loose loops everywhere and an unchecked error eventually reaches a citizen's entitlement.

Embodied intelligence

Between 2031 and 2035, AI fuses with robotics and physical machines and begins operating in the physical world. The timing is less certain than confident predictions suggest; the direction is not.

The ten-line institutional AI policy

No officer reads a hundred-page rulebook, and a policy nobody reads protects nobody. For a Divisional Secretariat or any institution, the practical instrument is a ten-line AI policy that staff will actually follow — short enough to sit on the desk and be consulted in the moment a decision is being made.

Brevity here is a design decision rather than a compromise. The moment that matters is the one where a tired officer at 4 p.m. is deciding whether to paste a citizen's file into a chat window. What governs that moment is what they can remember, not what was circulated.

In the final chapter, we conclude this book with the technical journey from GPT-4 to GPT-6, the PhD team that has arrived in your home, and the 5 golden rules for the public officer.