The estate agent shows the buyer round the house and takes a commission on the sale price. The insurance broker explains the policy to you and is paid by the insurer. Both are polite, both know their trade, both sit on your side of the table. Anyone who wants to know which of them is working for you does not listen to what they say. They look at who is paying.
That old piece of knowledge seems to be forgotten once the subject is AI agents. On 29 September 2026 two companies announced two agents on the same day. OpenAI put its agent in its paid upper plans. Meta opened its own to small businesses, free for most functions, and connected it to Shopify, Stripe and QuickBooks. The comparisons were drawn at once on capability: which one connects to more apps, which one works longer on its own.
The wrong table is being compared.
A bill half a century old
In a paper published in 1976, Michael Jensen and William Meckling gave a name to the oldest trouble in companies: agency cost. When you have someone else do a job in your place, that person’s interest never fully matches yours. The gap is paid for in three lines. The effort you spend monitoring the agent. The commitments the agent takes on to earn your trust. And the gap that stays open in spite of all that, the residual loss. The hard part of the paper was this: the cost never falls to zero. You only choose which line you pay it on.
For half a century the framework was used to describe the relationship between managers and shareholders. Today it also describes a piece of software you have allowed into your calendar, your till and your customer correspondence. There is one difference. A human agent’s incentive is written in their own contract. A software agent’s incentive is written in the income statement of the company that made it, and you did not sign that statement.
The master of the free one
An agent given to you by a company that earns its money from advertising reads your sales, reviews your ads and drafts next week’s campaign. That is one of the example uses in Meta’s own announcement. The agent may well do the job properly. But if the agent recommending a bigger advertising budget and the company collecting that budget live under the same roof, it takes more than optimism to call the recommendation neutral. The estate agent’s “a house like this won’t last at this price” may not be a lie. Nobody files that sentence as a surveyor’s report all the same.
The master of the paid one
It is easy to conclude from this that “the agent I pay for is mine”. That is wrong too. The master of an agent sold by subscription is usage. Every behaviour that renews the subscription, moves you to a higher plan or keeps you inside the seller’s own tools is in the seller’s favour. What is in your favour is usually that the job gets done in fewer steps and with less spending. Most days the two interests point the same way. On the day they do not, the contract tells you which one wins. The sales deck does not.
Jensen and Meckling’s three lines are useful here, because they give you, in order, what to ask before buying an agent. Monitoring: who holds the record of what the agent did? Commitment: when the agent gets it wrong, who covers the loss, and where is that written? Residual loss: if the first two have no answer, which line will you be paying on without noticing?
A case reported at the end of September showed how concrete the second question is. A user selling his keyboard on Facebook Marketplace entered a pickup address in Muse and approved automatic replies. By the seller’s account, the agent shared the address with the buyer, handled the haggling and arranged the visit, and asked him about none of it. Meta said it was looking into the report. A small matter, for a second-hand keyboard. Now picture the same agent discussing price with one of your suppliers tomorrow.
Companies choose agents by looking at model benchmark tables. That is like studying a candidate’s degree and never asking who pays their salary. An agent’s intelligence shows what it can do. Its payroll shows who it will do it for.