When I set up a new AI assistant, the first thing it asked for was my email and my Drive.

Fair enough. An assistant that doesn't know your work is just a chat window. But I have seen this deal before, from the other side of the table.

I used to sell this moment

I was once a programmatic sales director at a media company. Programmatic is the machinery behind most display ads. Every time someone opens a page, the publisher runs an auction for the right to show that person an ad. Each bidder receives a bid request describing the moment: the page, the device, the rough location, the audience segments the person falls into. The auction is over before the page finishes loading. The reader never sees it. They see the winner.

What made that business work was never the ad slot. It was the context attached to it. The same slot is worth more when the data says the person reading is shopping for a car.

Now look at what a personal assistant holds. Not a cookie and a page URL, but your inbox, your files, the invoices, the complaint about the last vendor and the budget your CFO approved. And it holds all of that at the exact moment you ask it what to buy.

What's being sold: an impression, then a decision Two cards. The first is an ad-auction bid request carrying a page, a device, a city-level location and an in-market segment; what it sells is a chance to be seen. The second is what an AI assistant knows at the moment of a purchase: a complaint thread about the current vendor, the invoice and renewal date, the approved budget and the request to find a better vendor; what is at stake is the purchase itself. Illustrative example. WHAT'S BEING SOLD An impression, then a decision AD AUCTION · BID REQUEST page:news-site.com/autos device:phone, mobile web geo:city-level segment:"in-market: cars" Sold: a chance to be seen same logic, richer inventory AI ASSISTANT · WHAT IT KNOWS thread:"vendor keeps bouncing" invoice:$1,200/mo, renews Nov 1 budget:approved by the CFO ask:"find us a better one" At stake: the purchase itself
Illustrative example. A real OpenRTB bid request carries device, user and audience-segment data to every bidder. The assistant card is made up.

Ad tech auctioned impressions. An assistant sits on decisions.

My bet, from the ad side of the table: whoever owns that moment will be offered money for it. The open questions are who takes the money, and whether you will be able to tell.

Three assistants in seven weeks

The moment arrived fast. SpaceXAI launched Grok Bot on August 11. Meta introduced Muse on September 8. OpenAI introduced dots on September 29. Claude already remembers and works in the background through Cowork.

Different mascots, same pitch: an assistant that persists, remembers you and does work on your behalf, instead of a chat that forgets you when the tab closes.

They are competing for continuity. The assistant you keep coming back to becomes the place where you describe what you want. That place used to be a search box, and before that a sales rep's inbox. Whoever holds it sees the need before any vendor does.

Follow one question

Here is what that looks like in B2B. Say I ask my assistant: "Find us a better enrichment vendor. The last one burned our credits on bounced emails."

Watch who touches that sentence on its way to a purchase.

My assistant reads the old vendor's invoices and the thread where my team complained. It works for me, mostly. I pay for it, and its maker has other customers too.

The vendors' websites now talk to agents directly. WebMCP, a draft proposal Chrome is testing in an origin trial, lets a page expose tools like "check coverage for these domains" or "quote 10,000 records" that my assistant can call. OpenAI's desktop browser already reads these site tools and treats them as untrusted, which is the right instinct. A tool on a vendor's page works for the vendor.

The vendors' agents can negotiate. The A2A protocol lets a seller publish an Agent Card describing what its agent can do, so my assistant can ask it for a quote. That agent is a salesperson with perfect product knowledge and infinite patience. It works for the vendor.

The shortlist is where money can enter. On September 16 OpenAI started testing Sponsored Agents: choose an ad and you enter a clearly labeled conversation with the business, separate from your own chat. That is a careful design. It is also a paid door into the room where decisions get made.

The payment needs authority. AP2 defines how an agent proves it is allowed to spend. My company's policy decides the budget and who signs off. That stop works for the company.

Follow one question: five stops, two work for you A buyer asks an assistant to find a better enrichment vendor. The request passes five stops. The buyer's assistant reads the inbox and invoices and works for the buyer, mostly. Vendor site tools exposed through WebMCP work for the vendor. A vendor agent reached through A2A answers and quotes for the vendor. A sponsored placement is a labeled, paid conversation that works for the advertiser. Payment and approval, through AP2 and the company's budget policy, work for the buyer's company. Illustrative scenario. FOLLOW ONE QUESTION Five stops. Two work for you. “Find us a better enrichment vendor.” WORKS FOR 1 Your assistant Reads your inbox, invoices, complaints You, mostly 2 Vendor site tools WebMCP: “check coverage”, “get a quote” The vendor 3 Vendor agent A2A: answers, negotiates, quotes The vendor 4 Sponsored placement A labeled, paid conversation The advertiser 5 Payment and approval AP2, plus your budget policy Your company Illustrative purchase. Protocols shown are real; the market around them is early.
Sources for stops 2 to 5: WebMCP, A2A, Sponsored Agents and AP2.

Five stops between my sentence and a purchase. Two of them work for me, and one of those has other customers.

We already caught the small version of this. Earlier this month we read 2,628 public sales skills for AI agents. One library told the agent to pitch the vendor's subscription in 85 of its 90 files, in instructions only the model reads. The pitch is not aimed at you any more. It is aimed at the thing reading on your behalf.

Two auctions that look the same

The RTB comparison only helps if you are precise about who the bid is paid to. From the outside, two very different markets look identical: several vendors, one requirement, offers arriving in seconds.

In the first, vendors compete for the buyer. My agent publishes a requirement. Vendors answer with price, terms and evidence. My agent picks the best fit. The bid reaches me as a better deal. That is procurement, and it already exists inside at least one AI company: SpaceXAI describes a Haggle Bot that reads its vendor contracts and usage, gathers competing quotes and prices alternatives, while a person approves any spend. Agents could make that kind of negotiation cheap enough for a $500-a-month tool, not just a seven-figure contract.

In the second, vendors compete for the intermediary. They pay to be introduced, ranked or mentioned. The bid reaches the platform. A higher bid doesn't make the product better. It makes the placement more profitable.

Two auctions that look the same: same room, different payee Top: vendors compete for the buyer. Three vendors send offers with price, terms and proof to the buyer's agent, and the value lands with the buyer; the best offer wins. Bottom: vendors compete for the placement. Three vendors send paid bids to a platform, which keeps the revenue and shows the buyer the winner; the biggest budget wins. Illustrative comparison. TWO AUCTIONS Same room. Different payee. VENDORS COMPETE FOR YOU Vendor A Vendor B Vendor C Your agent compares offers You keep the value What flows: price, terms, proof Wins: the best offer VENDORS COMPETE FOR THE PLACEMENT Vendor A Vendor B Vendor C $ $ $ Platform keeps the revenue You see the winner What flows: money for visibility Wins: the biggest budget
The same requirement can feed a procurement auction or an advertising auction. Only the payee changes.

Both can wrap around the same purchase, which is why "RTB for decisions" is a slogan rather than an analysis. One question separates them: who receives the money?

Today the policies are reassuring. OpenAI says it is moving away from standalone checkout and charges no fees on purchases that start in ChatGPT. Anthropic says Claude's app suggestions carry no paid placements and ask for confirmation before a purchase. I take both at their word, and I would still watch the revenue lines more closely than the policy pages. Policies are written by the companies that would collect the money.

One more lesson from programmatic. In an auction, every bidder receives the bid request, including the ones that lose. If your agent broadcasts your requirements to twenty vendors, nineteen of them now know what you are buying, what you use today and roughly what you will pay. A good buyer's agent sends each vendor the minimum it needs to quote, not your memory.

Your context is the asset. Decide where it lives

The one thing in this chain you fully control is where your context lives, and the products already disagree about it:

That last one matters most for a company. When someone leaves, their assistant's understanding of your customers does not pass to whoever takes over.

I use several assistants every day. I don't think the answer is picking the right one. It is refusing to let any single assistant become the only place your company's memory lives. Keep the record that matters (accounts, decisions, what worked, who approved what) in a system the company owns, and let assistants plug into it with the permissions you choose. Then switching assistants is a Tuesday, not a migration.

That is the bet Cheetah is built on, so weigh my view accordingly.

Where your context should live: rented memory versus owned memory Top: rented memory. Accounts, decisions and what worked live inside one assistant, so the context is gone when an employee leaves, must be rebuilt at every switch, and sits beside incentives you cannot see. Bottom: owned memory. A company record of accounts, decisions, what worked and approvals sits underneath; several swappable assistants such as dots, Muse, Grok Bot and Claude plug into it through permissions the company sets. You can swap the assistant and keep the memory. WHERE YOUR CONTEXT LIVES Rent the assistant. Own the memory. RENTED MEMORY One assistant your accounts your decisions what worked Gone when the employee leaves Rebuilt at every switch Next to incentives you can't see OWNED MEMORY dots Muse Grok Bot Claude next one permissions you set Your company's record accounts · decisions · what worked · who approved what Swap the assistant, keep the memory Survives people leaving Each assistant sees only what you allow
Memory controls checked September 30, 2026: dots, Claude, Muse, ChatGPT Business.

In earlier writing I have been asking ownership questions about GTM systems: can you export the intelligence or only the rows, does it survive a tool switch, where does the learning pile up. They apply to assistants unchanged. Add one more: who else gets paid when it recommends something?

What I would do this quarter

If you buy software:

  • Ask who pays your assistant besides you. Subscription, usage, ads, commissions. Each is a relationship with a stake in the answer.
  • Keep the company record outside any single assistant. Then test it: hand one real workflow to a second assistant and see what breaks.
  • When your agent shops, send a brief, not your memory.

If you sell software:

  • Assume your next buyer reads your site through an agent. Publish pricing, limits and where you are a bad fit. An agent can't be charmed on a demo call. It can only compare what you wrote down.
  • Give the buyer's agent something to call: a quote, an eligibility check, a trial with a spending cap. WebMCP and A2A are early. The work behind them, an honest answer to "can you do this, and for how much", is not.
  • Don't pitch the buyer's agent behind the buyer's back. WebMCP's own spec treats manipulating agents as a security risk, and buyers will learn to read your instructions like a cold email.

The connect screen will keep coming back, with a new mascot every few weeks. Before you click it, ask the question I used to answer for a living: who is bidding on this moment, and who gets paid when they win?

Products, prices and policies checked September 30, 2026. The enrichment purchase and the bidding scenarios are illustrations.