What a build plan answers
A build plan answers four things before we quote a price: what data the agent reads, which decisions it is allowed to make on its own, how we will know it worked, and who owns it after we leave.
Discovery is a fixed-fee engagement, and it ends in that document. The build is then quoted against the plan, so the number you see is attached to a scope you have already read. This piece walks through the four answers using a build that shipped.
The client is Genesco Sports Enterprises. Every month, the finance team rebuilt job-profitability numbers by hand from Sage 50 exports: importing files, reconciling totals across spreadsheets, and re-keying prior-month profit figures. We built a two-pass web app that ingests the exports, computes the job totals, pre-fills the figures the rules allow, and hands back a clean workbook for the judgment calls. A month-end reconciliation that took the CFO about 4 days by hand now runs in about 60 seconds.
A plan with those four answers can be priced. A list of features leaves all four open, and a quote written over those gaps changes once the build starts and the gaps show up.
Data the agent reads
The plan lists every source the agent reads, the state that source is in today, and the person who can get us access to it.
At Genesco the list had two entries: the Sage 50 exports, in a known format and produced on a schedule the finance team controlled, and the prior month's agreed figures, which lived in the last month's workbook. Both were already structured, so the build spent no time on cleaning them up.
Most lists are longer, with a CRM where half the fields are typed by hand, a shared drive of PDFs under three naming conventions, a procedure document last updated two years ago, and an inbox that holds the only record of what a customer agreed to. Each source gets a line in the plan that says what shape it is in, what has to be cleaned or structured before an agent can read it, and who owns that work.
We count the sources and their state during Discovery so the quote can be fixed before the build starts.
Decisions written as rules
Each decision in the process is written as a rule with three outcomes: proceed, escalate, or stop.
Rules usually live in chat threads, in a procedure nobody updated, and in the head of the person who has done the job longest. The plan moves them into writing with named inputs and named outputs, so the agent knows when to act on its own and when to hand the work back to a person.
At Genesco the first pass fills every figure that follows a written rule. Items that need a call go into the handback workbook, and the second pass takes the reviewed workbook as given. The judgment calls stay with the finance team. We drew that line with the CFO, going line by line through a real month, and the reasoning is in judgment calls stay with the finance team.
The test we use is one sentence. If every step and every criterion for a good answer can be written down, the software can own the figure. If a person regularly weighs something nobody ever wrote down, the person keeps it. Writing the rules out surfaces the exceptions early, while they are still cheap to handle, and the number of rules in the plan sets the size of the build we quote.
Measures and guardrails
Every measure in the plan is paired with a guardrail, and both are written before the build starts.
The measure is the outcome in your words: days down to minutes, a backlog cleared, a category of error that stops happening. Genesco's measure was the close itself. The guardrail is what keeps that measure honest: which items need a human approval, what the agent may never write, and what it does when a figure falls outside an expected range. We write both down before the build, so we count errors in the same review where we count the time saved.
Guardrails belong in code. Anthropic, writing about their own commerce agents, keeps every rule that matters outside the model, so an approval typed into a chat window sets nothing and a spending cap is checked against the state after a write. Their reason for the split is that when the model's text goes wrong, "the failure is a misstatement to correct and no action needs reversing."
The plan also says how the agent will be tested. Anthropic builds a test case by constructing the state directly, appending one user message, and grading the final state and the response, with code graders reading which tools ran and what the data holds afterwards. We write client tests the same way, with a negative case beside every positive one, and the suite runs on every change. There is more on this in how to test an AI agent before it goes live, and on designing the review step itself in the human review step in an automated finance workflow.
The owner and the payback date
The plan names one person on your side who owns the build and the date by which it should have paid for itself.
The owner signs off the rules. A named checker reviews what the agent produced each cycle. On the payback date we compare the measures against what the plan said they would be, and that comparison decides whether a second workflow is worth starting. We write those three into the plan before any code is written, so the first build is production work from day one.
At Genesco the owner is the CFO, Mark Tatum. He used the app, told us what to adjust, and got the adjustments. He described the work this way: "Every month our team burned hours rebuilding job-profitability numbers from Sage by hand. Plainpath built us a tool that does the heavy lifting and leaves us just the judgment calls. It has saved us real time and headaches, and they were quick to adjust it when we asked."
A build with nobody asking for changes drifts away from the work, which is the pattern in why most AI pilots fail.
Genesco's figures belong to Genesco. Their close had clear inputs, clear outputs, heavy manual effort every month, and a rule set that could be written down with the people who run it. A process where judgment enters at every step compresses far less, and we wrote about choosing a first one in which process to automate first.
The plain take
The four answers are the thing you are buying in Discovery, and the quote for the build is written against them. The audit comes first and it is free. An agent asks about your work and scores it, which shows you the kind of questions we ask before you spend anything. Discovery follows when you want the plan written.
Frequently asked questions
What is in an AI build plan?
Four answers. The data sources the agent reads and the state each one is in. Every decision written as a rule with a proceed, escalate or stop path. The measures that say it worked, each paired with a guardrail. The person who owns the build and the date by which it should have paid for itself.
How is an AI build priced?
Against the plan. Discovery produces the plan first, so the quote covers a scope you have already read. When the sources and the rules are counted before the build starts, the number has less room to move during it.
What happens in Discovery?
Discovery is a fixed-fee engagement, with the price listed on our Discovery page. A person checks the audit agent's work, digs into your top three processes, and writes the build plan for them.
Do we need clean data before you can build?
No. The plan writes down what shape each source is in and what has to be structured before an agent can read it, and that cleaning work is part of the scope. For Genesco Sports Enterprises the sources were Sage 50 exports in a known format, which kept the build small.
How do you test an agent before it goes live?
We build test cases that construct the state directly, send one message, and grade the end state, with code checking which tools ran and what the data holds afterwards. Every positive case has a negative one beside it, and the suite runs on every change.
References
- The anatomy of effective commerce agents — Anthropic
- The anatomy of effective commerce agents — Anthropic
The plain take
If you are about to pay for an AI build, ask what the plan behind the quote answers. Four answers make a price mean something: the sources the agent reads and the state they are in, each decision as a rule with an escalate path, each measure with its guardrail, and the person who owns the build with a date to check it against. The audit is free if you want to see how we ask those questions before you spend anything.
If one of your processes looks like this, the free audit at /compass is where to start.