--- title: "What is the agent’s work worth?" description: "An agent can save hours without finishing the whole job. Its value depends on what it contributes and what it leaves for a person to do." canonical: "https://www.rickblalock.dev/scratch-notes/what-is-the-agents-work-worth/" date: "2026-09-12" updated: "2026-09-12" author: "Rick Blalock" --- # What is the agent’s work worth? > An agent can save hours without finishing the whole job. Its value depends on what it contributes and what it leaves for a person to do. Published: 2026-09-12 Catalog: SN-011 Let's take some time to step away from the white collar and Valley perspective of agents and let's enter an owner/operator's world. An agent can prepare an estimate, or push through a detailed work order to 90% completion, with the last mile needing a quick expert review. There's a lot of folks talking about agents like they can do "100% of all your work, run your business entirely!" Some of that is true but humans can't even "fully run a business automated" without the owner's involvement anyway. So the framing can be deceptive and feel uncanny valley. Getting most of the work done and leaving the last mile to a person can be a very good arrangement. It depends on what that last mile involves. The agent might gather the site notes, find comparable jobs, pull together material costs, coordinate inventory, and prepare a draft. The owner reviews the assumptions, changes the crew assignment, and adjusts the margin before sending it to the customer. If that preparation would otherwise have taken hours, the owner can come out well ahead even after reviewing and correcting it. Well, then what is an agent's work worth? What's a person's work worth? It depends! [Tugce Erten and Sarah Wang recently argued](https://a16z.com/you-are-not-a-model-dont-price-per-token/) that AI applications should price the recognizable work they deliver rather than the tokens they consume. They distinguish between selling model access, selling useful work, and selling an attributable business result. They also point out that credits do not help much if they are only tokens hidden behind another currency. While we're all trying to figure out what billing unit to assign to an agent, ask this question: ***What has the agent actually contributed to the job?*** ## An estimate is part of a conversation Consider what happens after the owner sends that estimate. The customer comes back and asks what it would cost to leave part of the work until next year. Now the estimate needs to change. Some materials come out. The crew may need a different amount of time. Splitting the work into separate visits might introduce costs that were not there before. The agent can help work through those changes, with the owner reviewing the result again. A revision is not automatically evidence that the first estimate was wrong. The customer changed the request. Even with a person preparing everything, this back-and-forth can be a normal part of reaching an agreement. The owner's earlier changes need the same distinction. Choosing a different margin may be a judgment call. Adding a material the agent omitted from the original scope is a correction. Learning that access to the building is restricted is new information that may change the plan. And this is all happening while the operator has 4 more visits to make, he's 8 estimates behind, and phone calls are coming in! Humans and agents need to handle all of those situations, but the agent intelligence should distinguish corrections from judgment calls and changes in scope or information. And note that producing another version of the document doesn't automatically count as more value delivered nor it was a mistake or not valuable work. The useful work is helping the owner arrive at an estimate they can stand behind, including the revisions needed to get there, fast and with minimal involvement. ## The last mile still counts There is a big difference between reviewing a useful draft and rebuilding an unreliable one. If the agent shows which previous jobs it used, where the prices came from, and which assumptions need confirmation, the owner has something they can inspect. They can spend their attention on the parts that require experience or a decision. If the same draft contains plausible numbers with no explanation, the owner may have to reconstruct the estimate to find out whether it is safe to use. A finished-looking document can leave almost all of the original work intact, with the added burden of finding someone else's mistakes. The question is whether the work left for the owner is smaller and more manageable than the work they started with. The cost of being wrong includes what happens before someone notices. Saving preparation time matters, but so does making consequential errors easier to catch. A draft that flags an unknown gives the owner a chance to resolve it. Quietly filling the gap with a guess can make the estimate look more complete while making it less useful. “Human review required” is a reasonable boundary for an agent. It still needs to give the human a practical way to review the work. ## What are we paying for? Preparing a useful draft, producing an owner-approved estimate, winning the job, and making a profit are different contributions. An agent might do the first well while the owner remains responsible for approval. Winning the job also depends on the customer's budget, competing bids, and other things the agent does not control. Profit depends on what happens when the company performs the work. It does not have to claim credit for winning the customer to be worth buying. But the promise should be clear. If a business pays for an estimate ready for review, what condition should it be in? Which checks are still the owner's responsibility? If the agent misses something it was supposed to include, does fixing that consume more of the customer's credits? A customer changing the scope is a different reason to do more work than the product repairing its own mistake. In [No one can show me how](/scratch-notes/no-one-can-show-me-how/), I wrote about starting with a job the owner wants gone. That remains a useful direction. But the whole job does not have to disappear before an agent earns its place on the org chart. An owner who gets hours back and can still exercise judgment at the right moment has received something valuable. I would want to know whether the owner can prepare a sound estimate with less effort, handle the customer's changes more easily, and understand what they are approving. If all of that leads to an owner that doesn't have to stay up to 2am working through a backlog of estimates, that's a great start for an agent. --- [All Scratch Notes](https://www.rickblalock.dev/scratch-notes/)