There is a particular kind of generosity that should make a procurement officer nervous: the kind that arrives priced at almost nothing. When a frontier AI lab offers a government its flagship model for a token sum, the gift is rarely the product. The gift is the dependency — the slow accretion of workflows, data, and institutional habit that turns a pilot into plumbing. Across the United States in 2026, that pattern is unfolding in plain sight, and it is worth watching closely, because India is about to face the same choices on its own terms.
The pitch is seductive and, in fairness, partly true: AI can make public services faster, cheaper, and less miserable to use. But convenience and capture are arriving together. The job for anyone who cares about good government is to separate the two — to take the efficiency without surrendering the accountability. That is harder than it sounds.
The land grab
The public-sector push is broad and fast. According to industry reporting compiled by Build Fast with AI in mid-2026, xAI reportedly offered its Grok model to federal agencies at a nominal price — the kind of headline figure designed to remove cost as an objection entirely. OpenAI, the same reporting notes, signed a defence-related deal earlier in 2026, planting a frontier lab firmly inside the national-security apparatus. And beyond the marquee contracts, agencies have begun quietly reaching for consumer-grade tools — using something as ordinary as ChatGPT to analyse documents and crunch audits.
That last detail matters most, because it reveals the real shape of the trend. This is not a tidy, top-down procurement story. It is also a bottom-up one: civil servants, overwhelmed and under-resourced, pasting sensitive material into whatever model is closest to hand. The official deals set the strategic direction; the unofficial use sets the actual risk surface.
The ‘nominal price’ tactic deserves a name, because it is a recognisable play. Loss-leader pricing into a high-switching-cost buyer is how platforms are built. Once an agency’s analysts learn one model’s quirks, once its prompts and pipelines are tuned to a particular vendor, once a year of casework lives inside that ecosystem, the next contract is not really a negotiation. It is a renewal.
The case for
None of this would be happening if the upside were imaginary. It isn’t. The case for AI in government is genuinely strong, and it would be dishonest to wave it away.
Start with cost. Public budgets are finite, and a great deal of government work is exactly the sort of high-volume, text-heavy drudgery that language models handle well: summarising case files, triaging citizen queries, drafting standard responses, reconciling records, and — yes — combing through audits for anomalies a human would take weeks to find. Done well, this is not about replacing public servants. It is about freeing them from the parts of the job that burn time and morale, so they can spend more of it on judgement and citizen contact.
Then there is service delivery. Anyone who has waited in a government queue knows the cost of slowness is borne by the people least able to afford it. AI-assisted intake, multilingual chat support, and faster document processing can compress turnaround times from weeks to days. For a pensioner chasing a payment or a small business stuck in a licensing backlog, that is not abstract efficiency — it is the difference between a system that works and one that doesn’t.
Finally, modernisation. Government runs on legacy systems — decades-old databases, brittle interfaces, undocumented code. AI tools can act as a translation layer over that mess, making old systems queryable in plain language without a multi-year, billion-dollar rebuild. For institutions that struggle to hire scarce engineering talent, that leverage is real and hard to replicate any other way.
The case for caution
And yet. Every one of those benefits comes attached to a risk that is structural, not incidental.
The first is vendor lock-in and dependency. The nominal-price entry is the front end of a long game. When a single private firm becomes the substrate for how a government reads documents, scores applications, and analyses spending, the state has outsourced a piece of its own cognition. Pricing power follows. So does the quiet ability to shape what is possible — and what is not — inside public administration. A defence deal makes this sharper still: critical national functions running on a commercial model whose weights, training data, and roadmap the buyer does not control.
The second is bias, accuracy, and accountability. Consumer AI tools confabulate. They reflect the skews in their training data. When a model summarises a benefits case or flags an audit anomaly, its output is not neutral — it carries error and bias into decisions that change real lives. In the private sector, a hallucination is a bad customer experience. In government, it can be a wrongful denial, a flagged citizen, a misallocated rupee or dollar. Who is accountable when the model is wrong? The vendor points to the terms of service; the agency points to the vendor. The citizen, meanwhile, has no one to appeal to. That accountability gap is not theoretical: Build Fast with AI reports a 42-state attorney-general investigation into AI now under way in the US, alongside an unresolved fight over federal preemption — precisely the oversight uncertainty that aggressive public deployment exposes and amplifies.
The third is data governance and procurement integrity. Civil servants feeding audits and case files into consumer chatbots may be doing it to be helpful, but they may also be exfiltrating sensitive citizen data into a third party’s systems with no contract, no data-protection assessment, and no audit trail. And the nominal-price deal corrodes procurement itself. When the entry price is near zero, the usual disciplines — competitive tendering, cost-benefit scrutiny, exit planning — quietly lapse. A government that cannot say how it would leave a vendor has not really chosen that vendor. It has been chosen by them.
The India read
India arrives at this moment with an unusual advantage and an unusual temptation. The advantage is that the country has already built the most ambitious public digital infrastructure in the world. The temptation is to assume that head start makes the AI question easy. It does not.
The IndiaAI Mission has put real money and intent behind compute, datasets, and domestic model-building, and AI is already creeping into government services — from grievance redressal to agricultural advisories to translation across the country’s many languages. The strategic logic is sound: a country of this scale and linguistic diversity should not run its public administration on models it neither controls nor can interrogate.
The deeper opportunity is to treat AI as a delivery layer on top of Digital Public Infrastructure. India’s DPI — Aadhaar, UPI, the consent-based data-sharing rails — is, at its best, designed around interoperability and the citizen rather than around any single vendor. Bolt AI onto that as a conversational, multilingual interface, and you get something genuinely powerful: a person who cannot read a form can still access an entitlement by speaking, in their own language, to a system that already knows how to verify and disburse. That is GovTech worth building.
But the same DPI ethic must govern the AI layer, or the advantage evaporates. India should resist exactly the nominal-price capture now visible in the US. That means guardrails before scale, not after:
- Model portability by design. No AI deployment should be locked to a single vendor’s API. If you cannot swap the model, you do not own the system.
- Human accountability in the loop for any decision that affects rights or entitlements — with a named official, not an algorithm, answerable for the outcome and a real channel for citizens to appeal.
- Hard data-governance rules on what government data may touch which models, where it is processed, and how consent is recorded — backed by enforcement, not guidance.
- Procurement that prices the exit, treating a ‘free’ or nominal offer as a risk to be scrutinised, not a saving to be celebrated.
The American gold rush is a useful, slightly alarming preview. The technology is real, the benefits are real, and the capture is real too. India has the rare chance to get the sequence right — to demand the guardrails first and let the scale follow. The countries that confuse convenience for sovereignty will discover, a few budget cycles from now, that the cheapest deals were the most expensive ones they ever signed.
