Before an AI build quote: list judgments and exceptions to cut rework
Rework on an AI build drops when you write down your people's judgments and exceptions before asking for a quote.
Key points
- Before the quote, write down judgments
- No quantity? The AI stops and hands off
- Pick what people keep; narrow the quote
Contents
Before the quote, write down judgments people make
A&A perspective
Before you ask a vendor for a quote, list the judgments people make in the work today. A judgment here means a call someone makes on the spot, such as whether to accept a discount or whether to promise a delivery date.
From the sources
In January 2026, Anthropic published an article on how to build tests that check what an AI agent does, which it calls evals. An AI agent is a system in which the AI picks its own steps to get the work done. An eval gives the AI an input and checks whether the output passes.
From the sources
That article says two engineers reading the same spec can come away with different readings of how the AI should handle edge cases. It also says teams can spend weeks on AI features that work in early testing but miss expectations nobody wrote down.
A&A perspective
Automate & Augment (A&A), the company writing this article, sees these unwritten expectations as a major cause of rework. On the buyer's side, those expectations live in the heads of the people doing the work. A vendor cannot price judgments that are not written down.
Hypothetical example
Say a four-person company wants AI to take order emails from customers, enter them in the order sheet, and reply with a delivery date. The person in charge sets the date by checking stock and the factory schedule, and passes discount requests to the owner. Neither step is written in any manual.
A&A perspective
Write each judgment on one line with three parts: who decides, what they look at, and what they decide. Asking the person in charge to recall the orders that made them hesitate over the past month is a quick way to collect them.
From the sources
The eval article says that early on, teams can get surprisingly far with manual testing and intuition. It also says that to a team just starting out, more rigorous evaluation can look like overhead that slows down shipping.
A&A perspective
A&A thinks outsourcing changes the math. Every rebuild means a new quote and more days. Even if writing things down takes time, A&A expects it to cost less when it is done before the quote.
Write down where the AI stops and hands off exceptions to people
A&A perspective
Decide that exceptions, such as an order with no quantity, go to a person, and write them as conditions where the AI stops instead of forcing it to handle them. An exception is an order the usual steps cannot settle. Besides an order with no quantity, a request to repeat a discount that was meant for the last order only is one too.
From the sources
In a December 2024 engineering article, Anthropic writes that an AI agent may return to a human mid-task for more information or a judgment. It says agents can pause for human feedback at checkpoints or when they hit a blocker, and that it is also common to add stopping conditions such as a maximum number of iterations.
From the sources
The eval article recommends starting with 20 to 50 simple tasks drawn from real failures. It says to test both the cases where a behavior should occur and the cases where it should not.
A&A perspective
A&A thinks past orders are the shortest route to the exceptions. Collect about 20 orders from the past month that the person in charge called the customer back about or passed to the owner. For each one, write where the AI stops and who it hands the order to.
From the sources
The eval article says a good task is one where two domain experts would independently reach the same pass or fail verdict.
A&A perspective
A&A thinks each exception line can be checked the same way. If the owner and the person in charge read it separately and pick the same hand-off, the line will also make sense to the vendor.
Hypothetical example
In the order-email example, an order with no quantity goes to the person in charge with a drafted question for the customer. A request to repeat a one-time discount goes to the owner with no amount filled in. If the AI copied the discount from past documents, a term that no longer applies would end up in the reply.
A&A perspective
When the stopping conditions are written, the vendor can include the stop-and-hand-off part in the quote. When they are not, the question of what to do with this order comes up after delivery, and the system gets rebuilt.
| Case | Who decides today | What the AI does |
|---|---|---|
| Order with quantity and model number | Person in charge copies it to the order sheet | Enters it in the order sheet and hands the person in charge a draft reply with the date left blank |
| Order with no quantity | Person in charge calls the customer to ask | Stops and hands a drafted question to the person in charge |
| Request to repeat a one-time discount | The owner decides | Hands it to the owner with no amount filled in |
| Requested date earlier than stock allows | Person in charge checks with the factory | Stops and hands it to the person in charge |
| Order from a new customer | The owner decides whether to accept it | Stops and hands it to the owner |
Pick what people keep first to narrow the quote
A&A perspective
Once the judgments and exceptions are listed, split them into those the AI takes and those people keep, and decide what people keep first. A judgment people keep drops out of what the AI is asked to do. The narrower that scope, the fewer assumptions the quote has to rest on.
From the sources
Anthropic recommends finding the simplest solution possible and adding complexity only when needed. It says this might mean not building an AI agent at all.
From the sources
The same article describes workflows as systems that run the AI and its tools along code paths written in advance. For well-defined tasks, it says, they offer more predictability than AI agents, which choose their own way of working.
A&A perspective
In the table above, the AI goes on without stopping only in the first row. Even there, it leaves the delivery date blank and hands the draft reply to the person in charge. The other four rows are stop and hand off, so the AI decides neither discounts nor delivery dates. What the vendor builds stops at sorting the orders and choosing who to hand them to.
A&A perspective
In the order-email example, entering a complete order with quantity and model number into the order sheet looks to A&A like a well-defined task. For the stop-and-hand-off rows too, who gets each case can be written as fixed steps from the table. At quote time, you can ask the vendor whether it can be built as a workflow rather than an AI agent.
A&A perspective
A&A suggests attaching this table to the quote request as it is. The vendor then does not have to invent judgments the table does not contain. If you later want to move a row that people keep over to the AI, you can have it quoted as a separate request.
A&A perspective
Which judgments the AI can take depends on what AI can do today. Discount decisions may one day be handed to AI as well. For the first quote, though, keeping more judgments with people narrows the scope.
Summary
A&A perspective
Rework on an AI build tends to come from judgments and exceptions nobody wrote down. Before the quote, write down the judgments people make, such as discounts and delivery dates.
A&A perspective
Decide that exceptions, such as orders with no quantity, go to people, and write them as conditions where the AI stops. Deciding first what people keep narrows the scope of the quote.
Preparing to cut rework on an AI build means writing down, before the quote, the judgments and exceptions people handle today. What you decide next is which of them people keep, because that sets what the AI is asked to do and what the quote assumes.
FAQ
Can't the vendor write the list of judgments and exceptions for us?
It can draft the list, but your company decides what passes. Anthropic's eval article also says the people closest to the product requirements and users are best placed to define success.
How many exceptions should we collect?
About 20 is enough to start. Anthropic's eval article recommends starting with 20 to 50 simple tasks drawn from real failures.
If I build it myself with Claude Code, do I still need this list?
Yes. When you build it yourself, the list becomes the basis for the instructions you give the AI and for the examples you use to check the result.
Sources & editorial note
Primary pages read for this article. Publication dates below belong to the sources; access dates record our research.
- Building effective agents
Anthropic · 2024-12-19
Accessed 2026-10-24 - Demystifying evals for AI agents
Anthropic · 2026-01-09
Accessed 2026-10-24