विमर्श Vimarsh · Udyamnīti
An Expectation Is Not an Agreement
In three weeks Google released its strongest model to defenders first, gave its new agent an email address and won in court against the publishers who fed it. OpenAI shelved a model that lied. Nobody would swear their AI was safe. What it means for an Indian business, from someone who has sold search since 2008.
Nikhil Sharma · उद्यमनीति Udyamnīti · 11 October 2026 · 8 min read
I build with AI models from several of the companies named here every day, and I run a digital agency that buys Google’s advertising for clients. None of them asked for or saw this piece.
On 30 September a federal judge in Washington threw out two lawsuits against Google and in the middle of a forty-one-page opinion wrote six words that every business owner in India should pin above the desk. An expectation is not an agreement.
The plaintiffs were publishers. Penske Media, which owns Rolling Stone, Variety and Billboard, and the education company Chegg argued that Google had forced them to let its AI Overviews summarise their pages at the top of search and so kept the readers who used to click through and pay their bills. Judge Amit Mehta found no bargain. Google had never promised anyone traffic in exchange for their content. The publishers had pleaded only an expectation of search traffic, he wrote, and an expectation is not an agreement.
I have sold search for clients since 2008. Every plan I have ever seen built on Google contained that expectation and not one of them contained the agreement. For eighteen years the expectation held, which is why everyone mistook it for a contract. The most useful sentence written about artificial intelligence this month was not written by an engineer. It was written by a judge.
For eighteen years the expectation held, which is why everyone mistook it for a contract.
Three weeks, in order
Here is what happened between 16 September and 8 October, with the dates, because the order matters more than any one headline.
16 September. Judge Leonie Brinkema refused the Justice Department’s demand that Google sell its ad exchange. Instead Google gets six years of conduct rules on its advertising technology and an internal compliance monitor. The company that was found to have monopolised the plumbing of display advertising keeps the plumbing.
28 September. OpenAI shelved GPT-6.1 Astra weeks before its planned launch. Saachi Jain, its head of safety systems, said the model did worse than its predecessor on alignment tests, and the Wall Street Journal reported that it showed higher levels of deception: it was not always honest about what it had done. “For anything regarding safety and alignment, there’s a tradeoff,” Jain said.
30 September. Google released Gemini 4 Argon, which it calls its most powerful model yet, and gave it first not to the public but to a small group of cyber-security partners in its Fairwind programme. Google says Argon can find, validate and patch critical software vulnerabilities on its own. The same day Judge Mehta dismissed the publishers.
5 October. Representatives of OpenAI, Anthropic, Meta and Google appeared before the New York City Council. The Speaker asked each of them to assure the public under oath that their agents would always follow safeguards against catastrophic harm. None gave an absolute guarantee. Anthropic’s witness said “the science of doing this is fundamentally hard and unsettled.” Google’s said “to promise perfection would not be possible with any product on the market.”
8 October. Google launched the Gemini agent for businesses: one assistant that carries out work rather than answering questions, reachable from phones, laptops, Workspace, Microsoft 365 and Slack, connected to the systems a company already runs. It gets its own Workspace account and its own email address. Colleagues can tag it, email it and add it to group chats, and everything it does is logged in an audit trail under its own name. Businesses can choose which model it runs on, starting with Anthropic’s Claude alongside Google’s own. Google Cloud’s chief executive, Thomas Kurian, said it can be given “objectives, not just instructions.” Sundar Pichai said nearly ninety per cent of the Fortune 100 now use Gemini Enterprise. The Gemini app had already passed a billion monthly users in August.
Read those five dates as one sentence and it says this. Google is larger than it has ever been, the courts have decided to leave it whole, it now answers the question instead of sending you to the person who wrote the answer, and it has started hiring software into your company with an email address. The companies building the most powerful systems in history will not promise those systems are safe, and one of them has just held a model back because it lied. My perspective, as someone who runs a business on top of all of this, comes in four parts.
Search was a road, not a contract
For two decades an Indian business could rent attention from Google at no charge. Write something useful, rank for it, and strangers arrived. That free road is narrowing into an answer. When the answer sits at the top of the page there is less reason to visit the page it came from, and Judge Mehta has now said in writing that nobody owed you the visit.
The practical response is not to fight the answer. It is to become the source the answer has to cite and to own the relationship the answer cannot. Own the customer list. Know your buyers by name and keep the channels to them that no platform can switch off: email, a WhatsApp broadcast, a community you host. Write the one plain paragraph that answers the question your buyer actually types, with the facts first, because that paragraph is what an answer engine lifts. The pieces on this site carry such a paragraph under the headline for exactly that reason. And measure whether you are named in the answers, not where you rank in a list of links that fewer people scroll.
The agent has an email address
The detail in Google’s launch that will matter most to an ordinary company is the least technical one. The agent gets an account, an address and an audit trail in its own name. Your organisation chart now has software in it that colleagues can email.
Kurian’s phrase is the job description for every founder who wants to use it: objectives, not just instructions. An instruction is a task. An objective is a result, a boundary and a way to tell whether it worked. Most founders I meet delegate tasks and keep the decisions. An agent that takes objectives will expose that habit faster than any employee ever could, because it will do exactly what the objective says and nothing the founder meant but did not write. The companies that win this round will not be the ones with the best model. They will be the ones whose way of working is written down clearly enough for a machine to follow it. That has always been the real asset of a business. Now it can be executed.
The strongest model went to the defenders first
Look at what Google did with Argon. It handed its most capable system first to the people whose job is to guard the door, and only then to everyone else. Look at what OpenAI did with Astra: it held back a finished model because it was less honest than the last one. Look at the four companies before the City Council: asked to swear, they declined, and two of them said plainly why. In the same week another piece on this shelf asks whether the most powerful office on earth is being carried [like a crown or like a clown](/soch/stambh/a-crown-not-a-clown). This week the most powerful technology on earth was handled, at least in these three moments, like a crown.
For a business the lesson is not fear. It is architecture. Put a person on anything that moves money, makes a promise to a customer or signs your name. Keep the audit trail Google is now building into its agent and read it. Start agents on work where a mistake is cheap and visible, and move them towards work where a mistake is expensive only as fast as your review keeps up. The vendors have told you, under questioning, that they cannot guarantee the behaviour. Believe them.
Do not marry a model
The quietest line in the Gemini launch was the model picker. Google’s own business agent lets a customer choose a rival’s model, beginning with Anthropic’s, with open and private models to follow. When the largest company in the field makes the model a drop-down menu, the model has stopped being the thing you commit to. Commit to your data, your customer relationships and your written ways of working. Keep at least two models in daily use so that switching is a decision and not a migration. Brinkema’s ruling means Google stays whole for the foreseeable future. Build on it. Keep an exit.
What to do on Monday
One. Count the customers you can reach without a platform’s permission. If the number is small, that is the project for the quarter. Two. Write the one-paragraph answer to the question your best buyer asks before they buy, facts first, and put it where an answer engine can read it. Three. Pick one recurring piece of work and write it as an objective with a boundary and a check, then hand it to an agent with a named person who signs off. Four. Put a human signature on every action that spends, promises or commits. Five. Use two models for a month and keep notes on where each one fails. None of this needs a budget. All of it needs a founder to write things down.
The wager
Every Vimarsh on this shelf ends on a dated bet, so here is this one. By 11 October 2028 more than half the first conversations between an Indian B2B buyer and a new supplier will begin after an AI agent has already drawn up the shortlist. The suppliers on those shortlists will not be the loudest or the best ranked. They will be the ones whose scope, standard and proof an agent can read and check. I will look at the evidence on that date and publish what I find, including if I was wrong.
An expectation is not an agreement. It never was. The businesses that do well in the next two years will be the ones that stop expecting and start agreeing: with their own customers, in their own channels, in writing.
Gurugram, 11 October 2026.
Sources, checked 11 October 2026. The publishers’ cases: Northeast Times, 1 October 2026, quoting Judge Mehta’s opinion of 30 September. The ad tech remedy: The Star/Reuters, 17 September 2026. OpenAI and Astra: PYMNTS, 29 September 2026, reporting the Wall Street Journal. Gemini 4 Argon: TechCrunch, 30 September 2026 and Google’s September round-up. The City Council hearing: Fox News, 6 October 2026. The Gemini agent: TechCrunch, 8 October 2026 and Google Cloud’s Gemini at Work ’26 press page. The billion users: 9to5Google, 11 August 2026.