A small packaging converter can hand five jobs to AI today: reading enquiries and drafting quotes, checking artwork against a checklist, writing specification sheets, planning the weekly press schedule, and writing customer emails. None of them replaces the estimator, prepress operator or planner; each removes the typing and comparing around their decisions.
The catch is that every one of the five needs a check a person runs, and the cost of each is mostly set-up time rather than software, because one chat plan covers all five. Market-size reports and robots on the line matter to big plants; these five jobs sit in the front office of a plant with 10 to 60 people. All figures here are illustrative; swap in your own.
The five jobs side by side
| Job | What the AI does | What stays human | Set-up time | Running cost |
|---|---|---|---|---|
| 1. Quoting | Reads the enquiry, extracts size, board, quantity and finish, drafts the reply | Price sheet, the sums, final price | 1 day | One chat plan, around $20 a month |
| 2. Artwork checks | Compares artwork notes with your checklist, lists gaps | The preflight tool and the sign-off | Half a day | Same plan |
| 3. Specifications | Turns a messy brief into a clean spec sheet | Technical limits and tolerances | Half a day | Same plan |
| 4. Scheduling | Proposes a week plan from open jobs and constraints | Machine knowledge, the final call | 1 day | Same plan, plus a spreadsheet |
| 5. Customer emails | Drafts status, proof and delay messages | Promises and dates | 2 hours | Same plan |
The running costs overlap because one chat assistant, such as ChatGPT or Claude, does all five. Prices change, so check the vendor page; OpenAI's pricing page listed ChatGPT Plus at $20 a month when checked in November 2026, and that is the benchmark used here. Where a job pays back fastest depends on the plant. A converter swamped with enquiries starts with quoting; one with a trail of artwork rejections starts with checks.
1. Quoting: from messy enquiry to a drafted reply
Enquiries arrive as "need 5k boxes like the attached, 350gsm, matt, by 3 weeks", as a photo of a sample, or as a spreadsheet of sizes. An estimator spends 15 to 40 minutes on each, much of it rereading and keying the same details into the estimating sheet. AI is good at that reading and keying. It is not good at pricing, because the price rests on your board costs, press speed and margin, none of which it knows.
So split the work in two. The assistant turns the enquiry into a structured request. Your spreadsheet, which holds your rates, does the sums. The assistant then drafts the reply from the spreadsheet's figures.
The main worked example, with numbers
Say an enquiry asks for 5,000 folding cartons in 350gsm board, four colours plus a matt varnish, a finished size that lays out four to a sheet. A prompt to extract it:
Read this customer enquiry and return a table with these fields only:
product type, finished size (L x W x H, mm), board and weight, number
of colours, finish or coating, quantity, delivery date wanted,
artwork supplied (yes/no), packing requirements.
If a field is not stated, write NOT STATED. Do not guess sizes or board.
Then list the questions we must ask the customer before we can price.Typical output: everything filled except finished size, which was only in the attached sample photo, so the assistant lists "confirm dimensions and whether the sample is the target size" as the first question. That is the useful behaviour; it stops you pricing a guess.
Now the sums, done in your sheet with your own rates. With four-up on a sheet, 5,000 cartons need 1,250 sheets. Add 8 per cent make-ready and spoilage and you buy 1,350 sheets. Illustrative costs:
| Line | Basis | Cost |
|---|---|---|
| Board | 1,350 sheets at $0.55 | $742.50 |
| Press, including make-ready | 3.5 hours at $145 | $507.50 |
| Cutting tool | one-off | $210.00 |
| Gluing | 5,000 at $18 per 1,000 | $90.00 |
| Varnish and inks | allowance | $60.00 |
| Packing and delivery | allowance | $75.00 |
| Total cost | $1,685.00 |
At a 30 per cent margin on selling price, the price is $1,685 divided by 0.70, which is $2,407.14, or about $481.43 per 1,000 cartons. Notice that a 30 per cent margin on price is not the same as adding 30 per cent to cost; adding 30 per cent to cost would give $2,190.50 and a margin of only 23 per cent. Have the sheet use one formula for this and never let the assistant improvise it: chat assistants slip on exactly this kind of percentage, and a quote that confuses markup with margin loses money on every job.
With the number in hand, the assistant writes the email: confirm the specification as you understood it, state the price per thousand and the total, list the assumptions (artwork supplied as print-ready, delivery to one address, tolerances standard), give the validity period, and ask the open questions. You read it, fix anything odd, and send.
The Complete Excel Course with AI Integration
From formulas and pivot tables to Power Query, DAX, VBA and AI with Copilot, ChatGPT and Claude — hands-on from the first lesson.
By Sagnik Bhattacharya and Virginia Thorn, Primetime Emmy Award winner
Buy nowCost and the check
Running cost is a share of one chat plan. The set-up is a day: build the sheet, write the extraction prompt, and test it on 20 old enquiries whose final prices you know. The check that matters is that for the first 10 live quotes, the assistant's extraction is compared with the enquiry by the estimator, and the sheet's total is compared with how they would have priced it. If three of ten differ by more than 3 per cent, find out why before widening the use. A broader view of how AI behaves with custom jobs is in quoting custom jobs faster in a small workshop.
2. Artwork checks: a second reader against your checklist
Artwork problems cost real money: a rejected proof loses a day, and an error found on press loses a run. Your preflight software catches technical faults in the file, such as missing fonts, wrong colour space, low-resolution images and missing bleed. What it does not catch is whether the file matches the job: wrong barcode number, missing recycling symbol, legal text on the wrong panel, a copy line that differs from the approved wording.
That second kind of check is where an assistant helps. Give it your checklist and the customer's approved text, then the text pulled from the artwork. Ask it to compare line by line and report differences only. An example prompt:
Compare the APPROVED COPY with the TEXT FROM ARTWORK below.
List every difference in wording, numbers, symbols or order, quoting
both versions. Then check the artwork text against this list of
required items and say present, missing or unclear for each:
barcode digits, weight or volume statement, batch/lot area, recycling
symbol, manufacturer line.
Do not comment on design. Do not assume anything not shown.An illustrative result: "Approved: 'Net weight 250 g'. Artwork: 'Net weight 205 g'. Difference in digits." Humans reading a page of small print often miss transposed digits like that, and a barcode number that differs from the master is just as easy to miss. The assistant does not tire, but it can misread a number extracted from an image, so the final number comparison still happens with eyes on the proof.
Cost: half a day to write the checklist as a reusable prompt, then a couple of minutes per job. Limits: it cannot measure bleed, check overprint or tell you a spot colour will not hold, so it never replaces preflight. For the file-side half of this job, see checking artwork against a dieline template, and for tracking the back and forth of proofs, the tutorial on artwork approvals and sign-offs.
A before and after
Before: a customer's file arrives, the operator preflights it, it passes, and the job goes to proof. The customer approves it, and on press it turns out the recycling instruction says "check local rules" where the approved copy said "widely recycled". After: the same file goes through the copy comparison before the proof is sent, and the difference is flagged to the customer in a sentence. Nobody is faster at the first step, but the rework disappears.
3. Specification sheets from messy briefs
A specification sheet is the document buyers and your own floor use to agree what is being made: dimensions, board grade, print method, finish, tolerances, packing, test requirements. Briefs arrive in emails, call notes and old job bags. An assistant turns those into a consistent sheet in your template.
Give it the template headings once, then paste the sources. Tell it to write NOT AGREED in any field the sources do not settle, which forces open questions into view instead of burying them. Example output for a field that was ambiguous: "Board: 350gsm folding box board (customer email) / 'something stiffer' (call note) – NOT AGREED, confirm grade". A line like that is what stops a plant running the wrong board.
Two limits. Technical tolerances, such as how far a score can drift or what compression strength a corrugated box must hold, come from your own standards and the customer's tests, not from the assistant. And a spec you send out becomes a commitment, so someone who knows the plant reads it first. Set-up is half a day; in this illustration, each sheet then takes five to ten minutes instead of forty. A deeper treatment, including how to word specs so a buyer can order from them, is in the packaging specification tutorial.
Wording claims safely
Specs and labels often carry claims such as "recyclable" or "made with recycled content". Ask the assistant to list any claim it sees in the brief and mark each one as needing evidence; do not ask it whether the claim is allowed. Whether it is allowed depends on the rules where the product is sold and on your supplier's certificates. The tutorial on checking recyclable claims goes through that properly.
4. Scheduling: a draft week plan for the planner
Scheduling a small plant is a puzzle: due dates, board arriving on certain days, tools that must be ready before a die-cut job, changeovers that are cheaper when similar colours run together, and a press that is down every second Friday for maintenance. A planner solves it with a spreadsheet and experience. AI can propose a first draft in seconds, which the planner then corrects.
Prepare a table of open jobs: job number, customer code, due date, board needed and whether it is in stock, press hours, tool ready (yes or no), colours, and finishing steps. Then give the assistant the rules in plain words:
Plan next week for one press, two shifts of 7.5 hours, Monday to Friday.
Rules: a job cannot start until its board and tool are ready.
Group jobs with the same colour set to cut washes (each wash costs
45 minutes). The press is down Friday afternoon. Never schedule a job
after its due date; if impossible, list it as LATE with the reason.
Return a day-by-day table with start and end times and the total wash
time. Show your working for any LATE job.The output is a table the planner reads critically. A typical weakness: the assistant happily groups colours but ignores a changeover rule you forgot to state, such as "foil jobs need the cool-down overnight". Each time you catch a missing rule, add it to the prompt. After a month the prompt holds the plant's real constraints, which is a useful document in its own right.
Illustrative saving: if a planner spends three hours every Thursday building the next week's schedule, and a draft cuts that to one hour of correcting, that is about eight hours a month. The check is simple: after each week, record how many jobs ran late and compare with the manually planned weeks. Do not let the assistant move jobs once the week has started; changes to a running schedule belong to a person who can see the floor. Keep the rejected drafts as well: a planner who writes one line on why each draft was wrong builds the rule list far faster.
5. Customer emails: status, proofs and delays
Customer service in a converter is mostly the same dozen emails: where is my order, here is your proof, the board is late, can you change the quantity, here is the delivery slot, we need approval by Friday to hold the date. They are cheap to write once and tedious to write fifty times a week.
Create a short library of templates with the assistant, one per situation, using your own tone. Then each time, give it the facts and ask for the email. A delay example, as a prompt: "Draft a short email to a customer. Facts: job 4418, 10,000 sleeves, board delivery slipped from Tuesday to Thursday, new dispatch Monday not Friday, we can split the delivery if they need part by Friday. Polite, no excuses, offer the split." An illustrative result: "Hi [first name], the board for job 4418 now reaches us on Thursday, which moves dispatch from Friday to Monday. If you need part of the order by Friday, we can split the delivery and send what is ready. Let me know which you prefer. Thanks, [your name]".
Two rules keep it safe. The assistant never supplies a date; you do. And a promise in an email is binding, so for a delay that could trigger a penalty clause, a person reads the text first. Keep the library in one shared document, so whoever is on the desk on a Friday uses the same wording as the person who wrote it.
A mistake that looked like a time saver
Here is an illustrative slip. A planner asked an assistant to draft all of Friday's "your order has shipped" emails from a list. It wrote a lovely message for each, but on two rows the dispatch date column was blank, and it filled the gap with "today". Two customers were told a delivery had left when it had not. The fix: add "if the dispatch date is blank, write NO DATE and do not draft that email" to the prompt, and read the emails before sending. Nothing in the tool was broken; the instruction was missing.
Customer files and confidentiality when you paste things in
Packaging artwork, dielines and specs are often unreleased products, and many converters sign confidentiality terms that say nothing about chatbots. Read the clause before the first trial. If a customer's contract forbids sharing files with third-party services, that customer's jobs stay on the manual route, however convenient the assistant would be.
For everyone else, a few habits cost nothing. Use a business plan whose terms say your content is not used for training; ChatGPT's personal plans, for example, use your content for training unless you opt out (the "Improve the model for everyone" switch under Settings, then Data controls), so switch that off before the first paste. Replace the customer name with a code (C14), remove any product name that has not launched, and paste text rather than whole artwork files where the text is all the job needs. Keep a one-line log of what went into the tool and when, so that if a customer ever asks, you can answer in a sentence. None of this slows the five jobs; it takes a minute to set up as a standing habit on the shared prompt sheet.
A related point is who is allowed to use the tools. In a small plant it is tempting to let everyone paste whatever they like. Name one person as owner of the prompts and templates, and give the others a short rule list: no customer names, no prices from other customers, no bank details, and nothing sent out unread. Without that, the quality of the output depends on who happened to be on the desk that day.
Which job to start with, and what to measure
Choose by pain. Count, for one week, the hours spent on each of the five jobs and the errors each caused. Pick the one with the highest hours and a clear check, which is usually quoting or customer emails. Run it alongside the human method for two weeks before relying on it, and note three numbers: time per item, errors found by the human check, and how many outputs went out unedited.
Total cost for a small plant is modest. One chat plan at about $20 a month covers the five jobs for one user; ChatGPT's Business plan, for two or more users, was $20 a seat a month billed annually or $25 billed monthly when checked in November 2026, and it is not trained on by default, which matters when you paste in specs. The larger cost is the day or two of your own time at set-up, and a short monthly review of the prompts. Treat any hours-saved figure, including the eight hours a month in the scheduling illustration, as a guess until your own two-week side-by-side run has measured it.
If you would rather map which of the five fits your plant against your real enquiry inbox, that is what my AI implementation consultation is for; you can start from the contact page. Brand owners who buy packaging should read the buyer-side tutorial on going from concept to print-ready artwork, so you know what arrives at your prepress desk.
Converter questions before the first trial
Can AI replace our prepress preflight software?
No. Preflight software reads the actual file and tests it against rules such as fonts, colour space, bleed and overprint. A chat assistant looking at a screenshot or a text description can catch obvious omissions against your checklist, but it cannot measure the file. Use it as a second pair of eyes, not as the gate.
Is it safe to paste customer artwork or specs into a chatbot?
Customer artwork and unreleased product details are often covered by confidentiality terms. Check your contracts first, use a business plan that does not train on your content, and strip customer names and product codes from anything pasted. If a customer forbids it, use a local or in-house tool or leave that job to people.
How accurate are AI-calculated quotes?
Only as accurate as the price sheet and formulas behind them. Have the assistant extract the enquiry details and draft the email, but let a spreadsheet you trust do the sums, and check the first ten quotes against how your estimator would have priced them before sending any without a review.
What does this cost a small converter in total?
Software can be as little as one $20 a month chat plan, plus a spreadsheet you already own. The real cost is a few hours per job to set up templates and checks, and a regular review. Expect a rough total of one to three working days to get all five jobs running at a basic level.
Further reads
- Eco-Friendly Claims on Skincare Packaging: Check AI Copy — Checking green claims on packs for a brand customer.
- Food Product Costing: AI-Built Batch, Yield Loss and Margin Sheet — Batch costing and margin logic you can adapt to runs.
- How to Catch Made-Up Figures in AI-Drafted Proposals — How to catch invented figures in AI-drafted quotes.
- Can AI Produce Accurate Estimates for a Small Building Firm? — How accurate AI estimates really are in another trade.
- Subscription Box Pricing: Build a Margin Model With AI — A margin model for a packaging-heavy buyer.
- AI Package Design Generator: Mock Up Boxes and Labels for a Small Brand — Mock up photorealistic 3D boxes, pouches, and product labels using AI packaging generators to test designs before ordering custom print runs.
- AI in the Fishing Industry: Catch Logs, Buyers and Boat Paperwork — For one boat and a handful of buyers: AI for catch logs, invoice checks, price messages and maintenance plans, with the rules left to the regulator.
- AI Tools and AI Development: The Complete 2026 Guide — the AI hub, including every tutorial in the AI-for-business series.
Sources: OpenAI's ChatGPT pricing page, ChatGPT Business pricing page, business data page and data-controls help article (checked 5 Nov 2026); the worked figures for board, press time and margin are illustrative.