Government Isn't A Hyperscaler

On Monday, Alphabet fell about 5%, dragging the communication services sector down with it. Memory chip stocks plunged in Asia overnight and the selling crossed the Pacific by Tuesday morning. Beneath the price action, the cause was specific and revealing: investors are starting to ask whether the enormous sums being poured into artificial intelligence will earn their keep, and reports of senior talent leaving Alphabet's AI teams added to the unease.

That question — will the AI capex pay off — is the right question for shareholders. It is the wrong question for a government.

What the market wobble actually says

The sell-off was narrow, not broad. The Russell 2000 closed above 3,000 for the first time even as the big tech names tumbled. That tells you the market is rotating, not collapsing. The doubt is concentrated in a handful of names whose valuations had assumed the AI trade would compound for several more years without a pause.

The question being asked is simple. Amazon, Microsoft, Alphabet and Meta have collectively guided to roughly 25 billion in capital expenditure for 2026. Their combined free cash flow is forecast to fall to about billion in the third quarter — a decade low. At that ratio, the cash being spent is no longer comfortably backed by the cash being earned. That is what the nervous side of the market is pricing.

Why this question doesn't translate to government

Public administration does not sell tokens. It does not have shareholders. Its “revenue” is taxes and its “product” is rights, services, and the slow accretion of public trust. The whole logic of equity-style ROI — free cash flow divided by capex — is structurally absent.

And yet, watch any AI tender written by any government this year and you will see private-sector language smuggled in. Pilots. Use cases. Productivity gains. Cost savings per FTE. We are evaluating public AI investments in the only vocabulary the consultants brought to the room, and the result is that everyone ends up arguing over a metric nobody in the citizen-service chain actually cares about.

From inside a national tax administration, having worked on conversational AI for citizens at scale, I can say something specific. The right number was never “queries handled per rupee of compute.” It was something closer to: did the next confused taxpayer get an accurate answer, in the language she actually speaks, in time to act on it, without having to travel to a counter? That is an outcome question. It is not a capex question.

An outcomes-first lens

If the market is pricing AI on cash flow, the government should price AI on three different things:

• Service latency. How much faster does a citizen get an answer, a refund, or a decision after the system is deployed?
• Service reach. How many more citizens, in how many more languages and pin codes, can the same service touch without adding counters?
• Officer leverage. How much higher-order work can the same officer do because the system has absorbed the routine?

None of these are radical. They are simply restatements of what the public sector exists to do. The discipline is to write them into the procurement document before the vendor walks in, not after the dashboard is built.

A proposal: separate the two stacks

Here is a concrete proposal. A government should split its AI estate into two stacks and evaluate them by two completely different rules.

The internal stack — drafting, summarisation, case retrieval, file review, internal search across decades of orders — can be evaluated commercially. Hours saved, error rates, contractor-substitution ratios. That is fair, because the work being replaced was already priced.

The citizen-facing stack — multilingual chatbots, eligibility navigation, guidance on a new statute, status tracking on a refund — must be evaluated as service delivery. The metric is not unit economics. The metric is whether a person previously locked out of a service is now inside it.

Confuse the two stacks and the second one will always lose to the first on a finance committee's spreadsheet. It will be quietly defunded the moment an AI valuation cycle turns. And, as Monday's tape made plain, the AI valuation cycle will turn.

The public sector organisations point

Professor Michael Ting's course on the Analysis of Public Sector Organizations at SIPA made an argument that has aged into something close to a law. Public agencies serve multiple principals — legislatures, ministers, courts, auditors, citizens — and the metrics they adopt silently decide which principal they end up serving. Adopt private-sector ROI metrics for public AI and the system will end up optimising for the auditor's spreadsheet, not the citizen at the window. That is not a hypothetical risk. It is the default trajectory.

What to do this quarter

Three moves, for any department about to sign an AI contract:

• Write the outcome metrics before the price negotiation, not after.
• Build at least one citizen-facing metric whose headline a non-technical minister can defend in question hour.
• Keep funding the citizen stack even when the headline AI trade in the markets is going through one of its periodic doubts. Especially then.

The market is reminding everyone this week that the private bet on AI is a bet. Government is not making the same bet. It is deploying a tool that should be judged by whether the queue moved. The discipline is to keep saying that out loud while the news cycle is busy saying everything else.

#AIinGovernment #PublicSectorAI #DigitalGovernance #AICapex #IndiaAI #Governance #PublicFinance

Stop Building Chatbots.

Eighty-two percent. That is the share of inbound citizen queries an AI agent called Bobbi resolved in its first week across three English police forces, without a single hand-off to a human officer. The number matters less for policing than for what it telegraphs to the rest of government: the chatbot era, finally, is ending.

I say this with some scepticism about my own past. Anyone who has helped stand up a citizen-facing assistant inside a large national department knows the temptation to call any conversational widget a chatbot and declare modernisation complete. It is not. Bobbi, alongside Singapore's GovTech work and Estonia's interoperable agent network, signals that the next layer of public-sector AI will look very different from what most administrations are currently buying.

Chatbots answer. Agents act.

The technical distinction is sharper than the marketing suggests. A chatbot replies to a prompt. An agentic system is handed a goal — process this permit renewal — plans the steps, calls the databases, validates against the rules and executes the transaction. One produces text. The other completes a workflow.

This is why a recent World Economic Forum and Capgemini exercise mapping seventy core government functions sorts them more cleanly by workflow than by department. Eligibility assessment, document processing, fraud detection, permit issuance — these cut across ministries; they are the natural unit of analysis for an agent, not an org chart. Agents do not need to break silos. They operate outside them.

The chatbot habit is the trap.

The single most common mistake I see, both in India and abroad, is treating an agentic deployment as a chatbot upgrade. Same procurement template, same vendor, same knowledge base, a slightly cleverer model bolted on the back. That is not what the technology is for. The Capgemini survey of 350 public-sector organisations finds that ninety percent intend to explore or deploy agentic AI within two to three years, while Gartner forecasts that more than forty percent of those projects could be cancelled by 2027 — usually because the agency moved before understanding where the actual value sits.

The value sits in outcomes, not interactions. The right question is not can the bot answer this but what is the citizen actually trying to finish, and how many steps can we collapse into one supervised flow. An estimated one hundred and forty billion dollars in US federal benefits goes unclaimed each year because the application paths are too fragmented for the people who most need them. Most large administrations, India included, will find similar pockets if they look honestly: schemes whose stated reach far exceeds their actual delivery, not for want of policy but for want of plumbing.

What conversational AI for citizens actually teaches.

Citizen-facing conversational systems at national scale teach two unfashionable lessons. First, the volume of routine queries is far larger than any budget anticipated, and far more repetitive: a small set of questions accounts for most of the load. Second, when the citizen needs to complete something rather than learn something, the conversational layer hits a wall. The handover to forms, portals and back-office staff is where the experience breaks.

Agents are precisely the technology for that gap. In a direct-taxes context, an agent can read a notice, retrieve the relevant return, pre-fill a form against rule sets, cross-check historical filings and route only the genuinely ambiguous cases to a human officer. The administrator stays the architect. The agent does the clerical labour that nobody, on either side of the counter, particularly enjoys.

Bounded autonomy, glass-box defaults.

The discipline the field is converging on is bounded autonomy: being deliberate about what an agent is allowed to do, keeping a human meaningfully in the loop, and making every step auditable. The phrase that fits a public-sector context is glass-box governance. A chatbot answers, and the trail is shallow. An agent acts, and each action — every rule consulted, every database queried, every form submitted — should leave a perfect record. Used well, this is more accountable than the human-only system it replaces, not less.

The procurement template should change accordingly. Stop buying chatbots. Buy a workflow agent that ships with an audit log by default, with explicit, narrow permissions per task and a hard escalation rule for any case touching rights or material amounts. Bounded autonomy belongs in the contract, not in a vendor promise.

Where to start.

The most useful first move for any Indian department is unglamorous: pick one workflow that already exists end-to-end on paper or in disconnected portals — refund issuance, grievance redressal, a single notice-and-reply cycle — and rebuild it as an agentic flow with a human checkpoint at the decision. Measure completions, not interactions. Compare cost not to the chatbot it replaces but to the staff hours it returns. That number is what budget committees will eventually demand, and it is the only one that matters.

The chatbot was a useful detour. It taught millions of citizens that they could speak to the state in plain language and get a sensible answer. The agent is what turns that conversation into a completed transaction. The departments that grasp the difference now, and procure for it accordingly, will spend the next decade doing more with steadier headcount. The rest will spend it explaining why their bot still cannot actually do the thing.

#AgenticAI #PublicSectorAI #DigitalGovernance #GovTech #CitizenServices #AIinGovernance #IndiaAI

Who Vets The Frontier Model?

On June 2, the White House quietly published an executive order that almost no one in India will read closely, and that almost everyone in our public sector should. It puts in place an idea missing from every government's AI playbook so far: a sovereign capability to test frontier AI models before they touch the sensitive parts of the state.

The mechanism is unfussy. Within sixty days, the Director of the National Security Agency, in consultation with a handful of other officials, must develop and maintain a classified benchmarking process to assess the advanced cyber capabilities of AI models and determine the threshold at which an AI model should be designated a "covered frontier model". AI developers are invited, voluntarily, to hand a model to the federal government for up to thirty days before releasing it to anyone else. There is no licensing regime; the order expressly forbids one. But there is now a place where the state can look under the hood before the rest of the economy gets to drive the car.

The more interesting half of AI policy

India's AI conversation is loud about two things: compute and use cases. We talk about GPUs by the tens of thousands, about Bhashini, vernacular chatbots, AI in courts. Both matter. Neither answers the question every senior bureaucrat will face within the next three years.

That question goes like this. A vendor offers us a model that can read a citizen's tax history, draft a notice, recommend an action, and execute it. How do we decide whether this specific model is safe enough to be allowed into our environment? Right now, the honest answer is that we decide on vendor reputation, a procurement scorecard, and the judgment of two or three officers who have been around the file long enough to be trusted. That is not nothing. But it is not a state capability. It is hostage to the personalities and office life of the moment. And it does not scale to a country that runs DBT, GSTN, an income-tax stack, a new criminal procedure, and a thousand state-level applications on the same architectural assumptions.

What the order gets right

The good instinct is the separation between innovation and security. The order is loudly pro-innovation; it explicitly refuses to stifle the industry with regulation and rules out any mandatory licensing of new models. At the same time, the same document sets up an AI cybersecurity clearinghouse, expands a federal Tech Force, and asks CISA to make even covered frontier models accessible to operators of critical infrastructure such as rural hospitals, community banks, and local utilities.

It treats the question of whether a model is safe enough for sensitive deployment as separate from the question of whether it can be sold at all. That separation matters. Most jurisdictions, including ours, collapse these into one bucket called "AI regulation". The result is either over-regulation in the name of safety, or no real safety work at all because the political cost of the first option is too high. The American order suggests a third path: government as a privileged tester, not as a gatekeeper.

It leaves much open, too. A classified benchmark used only by NSA, with no public-sector analogue in the agencies that will actually deploy these models - health, tax, courts, energy - is incomplete. A thirty-day pre-release window is short. And the voluntary nature of the framework will work only if industry sees a clear benefit; the order hints at this through privileged early access for "trusted partners", but the carrot is not yet sharp.

An Indian version, sketched

The better view is that we should not wait. Three concrete moves I believe should be taken, none of which need new legislation.

A state-side model evaluation cell

House it within MeitY or CERT-In, staffed by a mix of officers and contracted technical talent on three-year tours. Its job is not to police the AI industry. Its job is to build benchmarks aimed at the failure modes that hurt government most: prompt injection in workflows that touch personal data, reasoning collapse under adversarial inputs in tax assessment, hallucinated citations in legal drafting or during chatbot functioning, model behaviour under translation between Indian languages. Publish what you can. Keep the rest classified. The point is to build a small group inside the state who, over time, develop instincts no procurement officer can substitute for.

A "fit for sensitive deployment" tag

Not a licence. A signal. A model that has been tested against the cell's benchmarks receives a tag that any department or any operator of critical infrastructure can default to when shortlisting vendors. Industry will adopt the tag because the customer base is large enough to make it worth the engineering.

A pre-deployment review for the highest-risk uses

For deployments that touch tax assessment, criminal procedure, biometric matching, grid SCADA, or welfare targeting or Enforcement Directorate for that matter, require a written review by the cell, the line department, and one external technical reviewer. Thirty days. Default approval if no issues are raised in writing. Reasons recorded either way. This is, ultimately, file work. File work is something the Indian state knows how to do.

The deeper point

The instinct, when a powerful new technology arrives, is to build an external regulator. The better move, more often than we admit, is to build internal capability. Having worked on AI systems that talk to ordinary citizens inside a national administration, I am convinced that what slows public-sector AI in India is not a missing regulator. It is the absence of trusted people inside the state who can look at a specific model and say, with reasons, whether it should be allowed near a citizen's record. We need Officers with AI bent.

The American order, in its limited way, has admitted as much. The interesting thing about June 2, 2026, is not the politics around it. It is the quiet acknowledgement that even in a country with the world's deepest pool of AI talent, the state still needs its own pair of eyes. We need ours sooner than we think.

#AIGovernance #FrontierAI #PublicSectorAI #IndiaAI #StateCapacity #AISafety #DigitalIndia