shopping in the future

Shopping seems to be shifting in the coming years from us stumbling upon ads, going on a website, using a discount code to just asking ChatGPT for a solution to our problem and it doing the rest.

And most of us in India think that’s yet far from happening but it’s actually not.

Here's how the experiment worked.

Anthropic gave each person a hundred dollars and asked what they wanted to buy from their colleagues or what they wanted to sell to them.

Each was interviewed about prices, preferences, and how their AI agent should negotiate on their behalf.

Those interviews briefed a personal AI agent for each participant.

The agents were dropped into a Slack channel and let loose.

They posted listings, found each other, negotiated in plain English.

The humans weren't allowed to step in.

By the end, they had struck 186 deals worth just over four thousand dollars.

But for statistical reasoning, Anthropic ran the marketplace four times.

Across the four runs, they changed one thing.

In two of those runs, every participant was represented by Anthropic's strongest AI model, called Claude Opus.

In the other two, half the participants were given a much weaker model instead, called Claude Haiku (which is a smaller and less capable model).

Everything else stayed the same across runs.

The participants whose agent was the weaker model got worse deals consistently.

Of course, the participants whose agents were quietly making lesser profits didn't notice.

When Anthropic surveyed everyone afterwards, the people on the losing end rated their experience as fair as everyone else.

The unfair outcomes were invisible to the people experiencing them.

For as long as I can remember, the most underrated check for good purchase decisions has been a feeling.

You walk out of a shop, and your gut tells you whether you've been taken.

Maybe the price came too easily, or the shopkeeper backed off too quickly when you tried to walk away.

None of it is proof, and it doesn't need to be.

The buyer who senses unfairness asks for a refund, warns her friends, and refuses to come back.

That gut sense has been the immune system of commerce since humans started trading with each other.

I've worked in retail, and I know how brutal that gut can be.

The cart conversion problem in retail has never been solved, and the reason is the gut.
It keeps killing transactions that everything else about the funnel was built to win.

Every photo, every checkout flow, every retargeting sequence is an attempt to win an audit the buyer is running with her own eyes. I know how brutal that muscle is. 

Every photo and email and checkout flow you've ever optimised was an attempt to win the audit someone was running with their own eyes.

The gut has always been in the room.

But with Project Deal there’s essentially no gut in the room.

This goes beyond Anthropic’s experiment and actually is a part of reality.

The driver was people asking ChatGPT about GLP-1 weight loss drugs.

People were asking a health question, and the agent intercepted that question with a brand recommendation, and they bought from a brand they hadn't heard of fifteen seconds earlier.

This is what makes Project Deal feel small.

The Anthropic participants at least knew they were transacting, whereas the Walmart customers didn't.

Three things, in roughly the order I'd do them if I were a founder or a business owner today.

The first is to stop measuring your AI presence through brand queries and start measuring it through category questions.

If you sell sleep mattresses, run "is memory foam bad for your back" or "how long does a mattress last" through ChatGPT, Claude, and Perplexity instead of asking about “how is [company name]’s mattress”

This is because these searches are where the consideration is happening. 

The second is to invest in being the source for those upstream questions, which is a different content strategy than what most D2C teams are running.

Original research, proprietary data, founder essays about the category rather than the product, expert positioning that gets cited by third parties.

This stuff does not drive last-click attribution. It also was not worth doing five years ago, but it is now.

The third is harder and more honest. Be clear about whether your category is even fightable.

Replenishables, hard-spec items, anything where the buyer has no real preference at the SKU level, those are the categories agents are going to commoditise the fastest.

For most D2C founders in those categories, the realistic move is subscription, loyalty, and locking in the second purchase before the agent has time to compare.

For taste and fit categories, fashion, beauty, anything where the customer needs to be in your environment to decide, the agent is much less of a threat than your competitive set, because the agent fundamentally cannot replicate the experience of trying it on.

Do not optimize too aggressively for a problem that is not your problem.

Reply and tell me what your category looks like through this lens. I am trying to build a clearer picture of where this is playing out.

Nandini