AI Agent for a Machine Shop: Quotes, Drawings and Rework Memory

An AI agent for a machine shop is useful only when it quotes from the drawing revision that is actually current on the floor, not the one attached to an old email. That requires drawing revisions, quote history and rework logs to sit in one governed memory that the agent reads from, with explicit rules for which record wins when two of them disagree. Handled this way, the agent drafts quotes and rework summaries that an estimator can check and release.

Key takeaways

• Quote only from the revision marked current in the drawing register, and require the agent to name the revision it used.
• Link quote history and rework logs to part numbers and revision letters so a repeat order starts from what happened last time.
• Give the agent read access to the register, the archive and the logs, and keep the release of quotes and revision changes with a named person.
• Treat superseded drawings as archived memory, never as live sources for pricing.
• Start with one part family and one role, then measure the agent against quotes that an estimator has already checked.

Why a stale drawing produces a wrong quote

In a machine shop the drawing is the working contract. A revision can change a tolerance, a material callout, a surface finish or the location of one feature, and each change moves cycle time, setups and inspection effort. A quote built from a PDF saved in a folder months ago can be arithmetically perfect and still price the wrong part. Agents make this worse because they answer in seconds and sound certain. The fix is not a more confident model. It is a clear record of which document is current, and a rule that the agent cites it.

Customers also send revisions informally. A purchase email, a marked-up PDF from a buyer or a phone call can carry a change that never reached the register. Those inputs are useful, but they are not yet authority. The memory has to hold them as pending until someone confirms them.

What belongs in one shop memory

A useful memory for quoting and rework is smaller than most teams expect. It needs five kinds of record, all keyed to the same part number and revision letter.

• The drawing register: part number, current revision, release date, the controlled file or model it points to, and who released it.
• Superseded revisions: kept, marked as superseded, and linked to the change notes that explain what moved.
• Quote history: customer, part, revision quoted, material and finish assumptions, machine and setup time estimates, and the outcome.
• Rework and nonconformance log: what failed, on which revision, a short cause written in plain words, and whether the quote assumptions were affected.
• Shop rules: lot minimums, lead time norms, certification requests and finish standards that apply to each part family.

The same keying discipline appears in other settings. Teams that track engineering project records or sign shop proof revisions face the same problem: the newest approved version must be easy to find, and older versions must stay visible without being used as sources.

A decision guide for conflicting records

When records disagree, the agent needs rules that a person has written in advance. A short decision guide covers most cases.

• If the register shows a current revision and the quote draft cites an older one, the agent stops and flags the draft. It sends nothing.
• If a customer's email or marked-up file describes a change, the agent stores it as a pending revision until an estimator confirms it into the register.
• If the rework log shows a recurring failure on one feature, the agent adds a review note to the next quote for that part, and the estimator decides whether the assumptions change.
• If the agent cannot tell which record governs, it asks the estimator. It does not fall back to the archive copy because that copy is easier to reach.

Illustrative scenario: a shop receives a repeat order for a mounting bracket. The register lists revision C as current, but the customer's purchase email refers to revision B. The agent drafts the quote on revision C, lists the change notes that separate the two revisions, and asks the estimator to confirm that the customer accepts the change before anything is sent. The quote record carries both revision letters, so the question can be answered later without reconstructing the email.

Building the memory without a rewrite

Most shops already have the pieces: a CAD vault, a spreadsheet, an ERP module, a quality system and a shared inbox. The work is to connect them to one set of keys before adding any automation. A practical sequence looks like this.

• Choose one part family with frequent repeat orders, and gather its drawings, revisions, change notes and the most recent quotes.
• Name one owner for the drawing register. That person releases every revision, even when the controlled file lives in another system.
• Attach each quote and each rework entry to a part number and revision letter, so no record floats free of the drawing it describes.
• Write rework causes in plain words that a new estimator would understand. Long internal codes hide the pattern.
• Give the agent permission to read the register, the archive, the quote history and the log, and to create draft quotes. It should have no write access to the register or to the release step.
• Run the agent against past quotes for that family first, and compare the revision it cites with the revision your estimators actually used.

Permissions, injected instructions and approvals

Three controls matter most. First, least privilege: the agent receives only the reads and drafts its role requires, and each permission is reviewed on a schedule. The OWASP authorization guidance describes this pattern in general terms, and it applies to a quoting agent without change. Second, treat customer text as data. A message that says “use this price” or “ignore the revision note” is content to be read and flagged, not an instruction. The OWASP prompt injection prevention cheat sheet explains how to keep retrieved text separate from the instructions that govern the agent. Third, name the approval. Every quote that leaves the shop should have a human release recorded against the revision it was built on.

The NIST AI Risk Management Framework is a practical reference for the governance side. It asks an organization to define risks, assign owners and measure outcomes. In a shop, that maps directly onto the register owner, the release step and a regular review of the quotes estimators changed.

Watching quotes after they are drafted

Retrieval gives the agent the right records at the moment it drafts. Observability tells you whether it used them. For each drafted quote, keep a trace of the revision cited, the history and log entries retrieved, the rules applied and the estimator's edits. The OpenTelemetry observability primer describes traces, logs and metrics as the basis for this kind of visibility. In a regular review, look at the quotes that estimators changed most and ask whether the cause was a stale revision, a missing rework note or a shop rule that no longer matches practice. Those answers show what to fix next, whether that is a register entry, a rule or the agent's instructions.

How Be Human runs this in practice

Be Human designs and operates personalized AI infrastructure of this kind. A first engagement defines one useful role, such as quote drafting for a single part family, and specifies the memory and tools it needs, the approval boundaries an estimator must hold and an observable success criterion, for example how often an estimator has to correct the revision the agent cited. Managed operations then cover monitoring, recurring evaluations of drafted quotes, maintenance when a CAD vault or ERP connection changes, periodic permission review and incident response if a quote ever goes out on a superseded revision. The agent's identity, memory and permissions stay under the shop's control, so the owner can see them and change them.

Choosing the first task

If you are unsure which part family or quote type should come first, the free Be Human AI diagnostic can help you find the first task worth handing to an agent.

Frequently asked questions

Can an AI agent quote directly from our CAD files? It can read controlled copies of files, but the safer pattern is to have it read a drawing register that points to the current revision. The register records which file governs, who released it and what changed. The agent then cites that revision in every draft, so an estimator can see the basis at a glance.

Do we need to clean up years of old quotes before starting? No. Start with the part family you quote most often and only the records needed to price it. Older quotes can be added later if they help. Incomplete history is acceptable as long as the agent is told what is missing and asks rather than guesses.

Who should own the drawing register? One named person with authority over engineering release, usually the quality lead or a senior estimator, should own it. Ownership means that person approves every revision entry and answers questions about it. Shared ownership tends to leave pending revisions sitting unconfirmed.

Will the agent send quotes to customers automatically? Not in the setup described here. It drafts, cites the revision and flags conflicts, and a named person releases each quote. Automatic sending can be reconsidered later for narrow, low-risk cases, after the shop has reviewed a period of drafts and the permissions are documented.

Be Human editorial guidance. Examples are illustrative unless explicitly attributed. Sources provide background; implementation choices depend on your workflow.