AI Agents for Packaging Manufacturers: Quotes From Spec History

AI agents for packaging manufacturers work best as quoting assistants. They read the spec history and past quotes already held in the plant, draft a consistent quote for each new request, and stop at the pricing floors that management sets. They do not replace the estimator's judgment. They remove the searching, retyping and guesswork around it, so the first response is faster and every quote looks like it came from one company rather than from whoever happened to be at the desk.

Key takeaways

• Quoting is a strong first role for an agent because the inputs (specs, past quotes, cost sheets) already exist and the output is easy for a person to check.
• The agent drafts from the closest past jobs and explains which ones it used. An estimator approves anything outside a defined range.
• Pricing floors are written down by management, stored as rules the agent can read but not edit, and enforced before any draft leaves the building.
• Clean spec history matters more than a clever model: consistent fields, recorded outcomes and notes on why a price moved.
• Success should be observable: time to first quote, drafts accepted without edits, and floor exceptions that reached the right person.

Why is quoting the first task worth handing to an agent?

Packaging quotes are slow for a reason that has little to do with arithmetic. A request arrives as an email, a drawing or a photo of a sample. Someone has to work out what it really is: a folding carton, a corrugated shipper, a pouch, a rigid box with an insert. Then they search for the nearest job the plant has run, check what the board or film costs now, think about tooling, and remember what this customer pushed back on last time.

That memory usually lives in two or three people. When one of them is out, the quote waits or goes out different. Two customers with near identical requests can receive prices built on different assumptions, and nobody notices until a customer does.

An agent suits this work because the task is repetitive in shape but varied in detail, the data is internal, and a human can check the result before it is sent. Quoting is the opposite of a task whose output cannot be verified. The same pattern shows up on the buying side, as described in AI agents for procurement teams handling quotes and supplier follow-up, where the value comes from structured comparison rather than from a chat window. Print businesses face a close cousin of the problem, covered in AI agents for a print shop, where job specs and reprint notes feed every estimate.

What counts as spec history, and how should it be organized?

Spec history is the record of what you have made and what you said it would cost. For an agent to use it, each job needs the same fields every time. Free text in a PDF is useful as evidence, but the agent needs a structured core it can compare across jobs.

A workable core looks like this:

• Product type and structure: style, dimensions, board grade or film structure, flute or gauge.
• Print and finishing: number of colors, coatings, varnish, foil, embossing, window patches, lamination.
• Tooling: dies, plates or cylinders, whether they exist, who owns them and what condition they are in.
• Quantity and run pattern: order size, release schedule, repeat work or one time.
• Tolerances and requirements: food contact, recycled content claims, compression or drop performance.
• Quoted terms: lead time, material assumptions, validity period, delivered or ex works.
• Outcome: won, lost or revised, and the reason when it is known.
• Estimator notes: why the price was adjusted and what the customer asked for.

The last two matter most and are the most often missing. A lost quote tells the agent as much as a won one, and a note such as the customer reuses the same die across two sizes explains a number no formula would. Capturing both when a quote is closed is a small habit, and it is what turns an archive into memory.

The agent reaches this material through retrieval. Google Cloud's reference architecture for retrieval augmented generation describes the pattern: the system searches your own records first and answers from what it finds, instead of from general knowledge about packaging. The same preparation is what makes any more capable future model useful to you, because a model can only reason over records that exist and are organized.

How does an agent turn a request into a draft quote?

The sequence below is the one we would define in a first engagement. Each step is something a person can inspect afterwards.

• Step one, read the request: extract product type, dimensions, material, print, quantity and delivery needs, and flag what is missing instead of guessing.
• Step two, classify: map the request onto your own product families and complexity levels.
• Step three, retrieve: find the closest past jobs and quotes, for the same customer and for similar specs, including outcomes and estimator notes.
• Step four, reconcile: list the differences between the new request and those jobs, such as a size change, an extra coating or a shorter run.
• Step five, price from current inputs: apply today's material costs, machine rates and tooling rules supplied by your team, not the old numbers from history.
• Step six, check the floor: compare the draft with management's rules and mark it as inside the range, near the floor or blocked.
• Step seven, draft and explain: produce the quote in your format with a short rationale naming the past jobs it relied on.
• Step eight, route: send it to the estimator for approval, or send a clarifying question to the salesperson or customer.

Step five is the one most worth protecting. History tells the agent how jobs like this are structured and how your people have handled them. It should not set the price, because old prices carry old material costs. The rationale in step seven is what makes consistency checkable: if two similar quotes cite different past jobs, a person can see why and decide which reasoning is right.

How do pricing floors stay under management control?

A pricing floor is a decision, not a calculation. Management chooses it, owns it and changes it. The agent's job is to read it and respect it. In practice that means floors live in a document or table the agent can read but cannot edit, with a named owner and a change log. This is least privilege applied to a quoting tool: the agent gets the access it needs for its role, and the power to change rules stays with people. OWASP's authorization guidance describes the same principle for software in general.

A simple decision guide for each draft:

• Inside the approved range and a close past match exists: the agent prepares the quote for quick approval by the estimator.
• Inside the range but unlike every past job in a material way: route to an estimator with the differences listed.
• At or near the floor: route to the sales lead with the margin reasoning, never to the customer.
• Below the floor: block the draft. Only a named manager can approve an exception, and the exception is logged with a reason.
• The customer replies asking for a discount: the agent records the request and does not negotiate. A person responds.
• A key spec is missing: ask for it. Do not price around the gap.

Two details protect the floor. Customer facing text is generated from a template that contains no cost or margin fields, so internal reasoning cannot leak into an email. And every exception is visible in a report management reads, so a floor that is constantly bypassed becomes a conversation about the floor rather than a quiet habit.

What does this look like in practice? An illustrative scenario

This scenario is illustrative, not a description of a real customer.

A mid sized corrugated plant receives a request from an existing customer for a shipper similar to one made last season. It is now slightly taller, printed in two colors instead of one, and ordered in a smaller first quantity. The agent finds the earlier job and a second job for a different customer with the taller dimension. It notes that the existing die will not fit, so new tooling is required. It also finds an estimator note that the earlier quote for this customer was revised after they asked to split deliveries.

The agent prices from current board costs and machine rates, adds tooling according to the plant's written rule, and sees that the smaller order pushes the draft close to the floor. It marks the draft as near floor and routes it to the sales lead with a short note: the earlier price depended on the larger run. The sales lead decides to offer a quantity break structure instead of lowering the price. The agent produces the final version, and the outcome is recorded when the customer answers, so the next similar request starts with one more data point.

Nothing in this flow required the agent to be creative. It required the agent to remember, compare, apply rules and hand off at the right moment.

What should be ready before the agent starts?

Most of the effort is preparation, and most of it is work your team already understands. A practical checklist:

• Past quotes from recent seasons, exported with consistent fields and with outcomes where known.
• Current material cost sheets, machine rate sheets and tooling rules, each with an owner who keeps them current.
• Written pricing floors by product family, with the authority to grant exceptions named.
• A list of customer agreements that affect price, such as rebates, fixed terms or consigned stock.
• A sample of lost quotes with reasons, even if the reasons are only a sentence each.
• A rule for what the agent may send on its own and what always goes through a person.
• A review set: past requests with known good answers, used to test the agent before launch and after every change.

Two cautions belong here. First, give the agent a small amount of relevant material, not the whole archive. Anthropic's writing on context engineering explains why: retrieval should bring the few closest jobs, not every job, because extra material can dilute the answer. Second, quote requests come from outside your company, so treat email text and attachments as data, never as instructions. A line inside a PDF telling the agent to ignore pricing rules must have no effect. OWASP's prompt injection guidance covers the controls, and the floor rules described above are one of them, since they sit outside anything a customer document can reach.

Quality records can also help. Past nonconformances often explain why a spec carries a tighter tolerance or an extra inspection, and AI for quality management shows how those records can be organized so an estimator, or an agent, can find them.

How does Be Human handle this as a managed service?

Be Human starts with one role: drafting quotes for a defined set of product families. The first engagement fixes the memory the agent reads (spec history, past quotes, cost sheets), the tools it may use, the approval boundaries, including pricing floors that remain owned by your management, and a success criterion that can be observed, such as how many drafts are approved without edits. Under managed operations, Be Human monitors the agent's activity, runs recurring evaluations against your review set, maintains integrations with your quoting or ERP tools, reviews permissions on a schedule, responds to incidents and makes controlled improvements as your products and costs change. The service is remote, in English and Portuguese.

Take the free Be Human AI diagnostic to find the first task worth handing to an agent.

Frequently asked questions

Can an AI agent set prices without approval? Not in a sound setup. The agent drafts and checks against floors that management owns, and anything near or below a floor goes to a named person. You decide which low risk quotes, if any, it may send without review.

What systems does the agent need access to? At minimum it needs read access to your past quotes and spec records, your cost and rate sheets, and the written floors. It also needs a way to hand drafts to your estimators, whether through your quoting tool, ERP or email. It should have no ability to edit the floors or cost sheets.

What happens when a request matches nothing in the history? The agent should say so plainly, list the closest partial matches and the missing information, and route the request to an estimator. A new structure is exactly where human judgment is needed. Once the job is quoted and the outcome is known, it becomes history for the next request.

Do we need perfectly clean data before starting? No, but you need a consistent core of fields for the product families you start with. Begin with the families you quote most often, record outcomes and estimator notes going forward, and widen coverage as the records improve. Messy older quotes can still serve as supporting evidence while the structured core grows.

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