Personalised Print on Demand: Screen Customer Uploads and Names

Screen personalised print on demand orders in three passes before anything is sent to print: automatic file and text checks, an AI triage that describes uploads and flags risky text, and a short manual review queue for anything flagged. AI helps sort and explain. It cannot clear copyright, so a person makes every borderline decision.

The goal is modest and measurable. You are not trying to catch every possible problem, which no shop can. You are trying to stop the obvious ones before they cost you a print, a refund and a complaint, and to keep a record showing that you looked. Most personalised orders are harmless names and dates. The workflow protects your time by letting those through quickly.

Follow me on Instagram@sagnikteaches

What actually goes wrong with personalised orders

Personalisation moves part of the design from you to the customer, and with it part of the risk. These are the problems that show up in practice.

Connect on LinkedInSagnik Bhattacharya

• Uploads that belong to someone else: a film still, a sports club crest, a cartoon character, a photographer's watermarked picture.
• Trademarked phrases and names: slogans, brand names and team names typed into a text box.
• Offensive or abusive text: slurs, hate symbols, harassment of a named person, disguised with spaces, numbers or symbols.
• Personal data on products: phone numbers, addresses or full names of a private person on a gift.
• Technical failures: low-resolution uploads, wrong file formats, text that falls off the print area.

The last category is not a legal risk but it produces more refunds than the others, which is why file checks come first in the workflow. Your print provider has its own content policy that sits above everything you do. Read it, because it decides what they will print, and a provider rejection after you have taken payment is the most expensive version of any of these problems.

Subscribe on YouTube@sagnikteaches

The three-pass workflow

Text-only orders that clear pass one go straight to print. Orders with uploads always go through pass two, because you cannot see risk in a file name. Pass three is where responsibility sits.

Pass one: file checks and a blocklist you can build in an afternoon

Start with what a spreadsheet can do. Export the personalised text and upload links from your orders into a sheet, with one row per order. Then add these columns.

• Cleaned text: the customer's text in lower case with spaces and punctuation removed, so a blocked word typed as "W O R D", "w.o.r.d" or "wo-rd" ends up looking the same as the plain word. In Google Sheets a formula such as =LOWER(REGEXREPLACE(B2,"[^a-zA-Z0-9]","")) does this; REGEXREPLACE works on text, so a purely numeric entry needs wrapping in TEXT first.
• Look-alike swap: a second cleaned column that turns common substitutions back into letters, such as 0 to o, 1 to i, 3 to e, 4 to a, 5 to s, using nested SUBSTITUTE functions.
• Blocklist hit: a column that checks each cleaned text against your blocklist sheet, for example =IF(SUMPRODUCT((Blocklist!A2:A150<>"")*ISNUMBER(SEARCH(Blocklist!A2:A150,D2)))>0,"REVIEW","OK"). SEARCH ignores upper and lower case, and the (Blocklist!A2:A150<>"") part skips empty rows, so you can leave spare rows at the bottom of the list for new entries. Store blocklist entries in the same cleaned form, lower case with no spaces or hyphens, or they will never match the cleaned column.

Your blocklist has two parts. The first is language you never want printed: slurs, explicit terms, threats. Build it by asking an AI to generate categories and examples, then pare it to what is relevant to your products and keep a second pair of eyes on it. The second is brand and trademark phrases specific to your niche. A shop selling sports gifts needs club names and slogans, a shop selling kids' party items needs character names, a pet shop needs breed-name brands. List what customers have actually asked for in the past, then add as you learn.

False alarms you will need to tune

A substring check flags innocent words that contain a blocked string, such as an ordinary word with a short blocked word buried inside it. Two fixes help. Match whole words for short entries by padding the text with spaces and searching for the entry with a space on each side. And keep a short allowlist of names and words that trigger repeatedly and have been reviewed as fine. Do not try to fix everything with rules: a false alarm costs you thirty seconds in the queue, which is cheap, and a missed one costs you a refund.

For files, set minimum standards you can enforce at upload: accepted formats, a minimum pixel size for the print area, a maximum file size. Your print provider or storefront app will list what it needs for each product. Reject the file at upload with a clear message rather than finding the problem after payment.

Pass two: AI triage with a prompt that explains itself

For orders with an upload or a flagged text, ask an AI model that accepts images to describe what it sees and rate the risk. The aim is to give your reviewer a head start, so the output should be short and should say why. Always ask it to separate what it can see from what it is guessing.

You are helping a small print shop screen a customer's uploaded artwork before printing
on a [mug / t-shirt / canvas]. Do not identify any real person from their face.

Describe the image in two sentences. Then answer:
1. Does it contain a recognisable logo, brand mark, sports crest, film or cartoon
   character, or a watermark? Name which, if you can see it.
2. Is there text in the image? Transcribe it exactly.
3. Is there anything abusive, hateful, sexual or violent?
4. Does it show personal data such as a phone number or home address?
Rate risk LOW, MEDIUM or HIGH and give a one-line reason. If you are unsure, say so
and rate MEDIUM. You cannot clear copyright; say that in your last line.

An illustrative output for a mug order with a photo of a cartoon dog wearing a football shirt looks like this.

Description: A cartoon dog in a red-and-white striped football shirt holding a ball,
with the words "Best Dad Ever" above.
1. Possible character: the dog resembles a well-known cartoon character; I cannot be certain.
   The shirt shows a crest that looks like a football club badge.
2. Text: "Best Dad Ever".
3. Nothing abusive.
4. No personal data.
Risk: HIGH. Resemblance to a known character and a club crest.
Note: I cannot clear copyright or trademark.

The reviewer opens the order, sees the dog and the crest, and sends the customer a message asking for a different image or a version without the crest. The AI did not make the decision. It told a person where to look, and it saved them from reading a file name and guessing.

What the model will miss

Models are better at obvious brand marks than at partial ones, and they do not know every character, local club or regional brand. They may over-flag because a design looks like something famous when it is an original. They should not be asked to name private individuals from a photograph, and you should not rely on them for it. This is why a clean result is only "nothing flagged". Keep it in your policy in those words.

A privacy point before you upload anything

Customer uploads often include family photographs, sometimes of children. Before you send any image to an AI tool, check what your plan does with content. Consumer chat plans may use conversations to improve models unless you turn that setting off, and business or API plans generally do not train on your content by default, but terms differ, so read the data section for the exact plan you use. If you are not comfortable, restrict AI triage to the text part of the order and review image uploads by eye. Delete exported files when the job is done.

A week of 120 personalised orders through the three passes

Run the numbers for an invented week at a small store selling personalised mugs and tote bags: 120 personalised orders arrive, 84 of them text only, such as a name and a date, and 36 with an uploaded image.

1. Pass one on the 84 text orders. The blocklist flags 5: two for a swear word with spaces in it, one for a football club slogan, one for a brand name inside a longer phrase, and one false alarm on an ordinary word. 79 go straight to print.
2. Pass one on the 36 uploads. Three fail the file check (too small), and the customer is asked for a larger file automatically. 33 go to AI triage along with the 5 flagged text orders, so 38 orders in all.
3. Pass two. The AI rates 27 LOW, 7 MEDIUM and 4 HIGH. The false alarm on the ordinary word comes back LOW.
4. Pass three. The reviewer checks the 4 HIGH and 7 MEDIUM, 11 orders, at about three minutes each, so 33 minutes. A sample of 5 of the LOW ones is also opened as a spot check, adding about 5 minutes.

The outcome: 11 orders reviewed carefully. The 2 orders with the spaced-out swear word are declined, the 2 with a club slogan or brand phrase get a message asking for different wording, 3 upload orders get a message asking for a different image, and the other 4 are approved after a look. That is about 40 minutes of human time across 120 orders. Without the filter, a person would have to open all 120. With a filter and no human, a crest and a dog would have been printed. The spot check of LOW ratings matters, because it is how you learn whether the triage is missing things.

Pass three: the review queue and what reviewers decide

Give the reviewer one place to work: a sheet or helpdesk view with the order number, the customer's text, a thumbnail or link to the upload, the AI note and a status column with three choices.

• Approve: print as ordered.
• Ask: send a message to the customer explaining what to change.
• Decline: cancel and refund, with a polite explanation.

Write the decision rules in one page so two people reach the same answer. For example: real brand names, sports crests, film and cartoon characters, and watermarked images need a changed design. Abusive text is declined, with no discussion. A name that is also a brand is approved if it is plainly a personal name in a personal context, and queried if it is part of a slogan. Personal data of a private person other than the customer, such as a phone number, is queried.

Ask the model to help by drafting the customer message. Review it before sending, and keep one stored version for each reason.

Draft a short, friendly message from a print shop to a customer. The uploaded design
contains [a sports crest / a cartoon character / a watermark / text we cannot print].
We can print the design without that element, or the customer can upload a different
image. Do not accuse the customer of anything. Offer a full refund if they would
rather cancel. Under 90 words. Use [first name] and [order number].

An edited version: "Hi [first name], thanks for your order [order number]. We can't print the club crest on the design, because it belongs to someone else. We can print the dog and your text without it, or you can send a different picture. If you'd rather cancel, we'll refund you in full. Just reply to this message and let us know which you prefer." It names the problem, offers two solutions and a refund, and does not suggest the customer did anything wrong.

Five more situations and how to handle each

A funny phrase that might be a trademark

Short slogans are the hardest. A phrase is not protected merely because it is catchy, and the answer depends on the goods and the owner. Add the phrase to a "to check" sheet, search it in your trademark office's free search, and look at the goods it is registered for. The trademark check tutorial for slogans shows a fuller routine. If you cannot decide, ask the customer for different wording rather than printing a risk.

A customer asks for a famous person's name or face

Decline, or ask for something original. Product designs using a real person's name or image can raise issues beyond copyright. A short standard reply helps staff say no consistently.

A tribute or memorial design

These carry an emotional weight. A name and dates are normal and should pass quickly. If the upload contains a copyrighted artwork, handle it gently: print the personal elements and offer an alternative to the artwork. Give a human the task and do not send an automated rejection.

An AI-generated image from the customer

Ownership and rules about AI-made images vary by provider and marketplace, and they change. The tutorial on AI images and copyright checks sets out what to check, and your platform's own rules apply on top.

A repeat customer who keeps sending problem files

Record each decision with the order number, reason and date in your log. After the second or third case, add a note to the account and apply stricter review. This is not an accusation. It is the same process, applied with the history in view.

How a hyphenated slogan can slip past the blocklist

Here is an illustrative slip, the kind this set-up invites. A shop adds a club's slogans to its blocklist, but its match column runs on the raw customer text instead of the cleaned column. A customer types one slogan as two words joined by a hyphen, the raw text does not match the entry, and the order prints. The rights owner complains through the marketplace, and the shop refunds the order and loses the listing while the complaint is dealt with. The fixes are small: point the match column at the cleaned text, store every entry in the same cleaned form, and hold a monthly review of declined and approved-with-doubt orders to extend the list. Treat the blocklist as a living document owned by someone, not a file you set up once.

Checkout wording that prevents half the problems

Screening works better when the customer has already been told the rules. Put three short lines beside the upload button and the text box, written by you and tidied by the model: what file types and sizes work, that you cannot print logos, characters, team crests or anyone else's artwork, and that you may contact them if a design needs changing before it prints. Add a tick box confirming that the customer has the right to use the image. A tick box does not make an unlawful print lawful, but it sets expectations and gives the reviewer a plain starting point for the conversation.

Illustrative wording: "Upload a JPG or PNG, at least [size] pixels on the long side. We can't print brand logos, team crests, film or cartoon characters or other people's artwork. If something in your design needs changing, we'll email you before printing, and you can cancel for a full refund." Notice the last sentence. It promises contact and a refund, which means a flagged order is the start of a conversation, not a surprise.

A second habit: show a live preview of the text inside the print area. Many refunds come from names that run off the edge of a mug or tote. Set a character limit per product, and ask the model to propose limits by testing your longest realistic names, such as a double-barrelled surname and a long date format, against the print width. Then test a few by eye on your mock-up before you publish the limits.

A starter blocklist layout

Keep the blocklist sheet in three columns so it is easy to maintain: the phrase as typed by customers, the category (language, club, brand, character, personal data), and the action (always decline, send to review). Add a fourth column with the date you added it and why, such as "customer order 1182". After six months this sheet becomes the most useful document in the shop, because it records what your own customers actually ask for rather than what a general list assumes.

Checking that the system works

A month in, look at five numbers. The share of orders flagged at pass one, the share of flagged orders that were real problems, the number of orders that reached print and later had a complaint, the time per review, and the number of refunds for technical file failures. If few flags are real problems, your list is too broad. If complaints come from unflagged orders, add what you learn. Keep the log of decisions, because it is the evidence that you screen orders and the training material for any new person.

Finally, set aside time each quarter to read your provider's content policy and any marketplace rules again, and to look at your misprint and refund pattern with the misprint and refund claims tutorial. The pairing of a short human review with a simple automatic filter is the same pattern described in the human review tutorial, applied to artwork. For listing-side rules on AI-made designs, see the Etsy AI art rules tutorial.

None of this is legal advice. If you receive a formal notice from a rights owner, or you are unsure whether a design is allowed, ask a qualified adviser, and respond to platform notices within the time they give you.

Questions about screening personalised orders

Can AI tell me whether an upload is copyrighted?

Not reliably. It can describe an image and flag recognisable characters, logos or watermarks, which is useful triage. It cannot know who owns an image or whether the customer has permission. Treat a clean AI result as no flag raised, not as clearance, and keep a human decision for anything doubtful.

Should I screen every order or only some?

Screen every personalised order with the cheap automatic checks, such as a blocklist and file checks. Send only flagged or uploaded-artwork orders to a person. That keeps review time to a small share of orders while making sure nothing personalised goes to print unseen.

What do I do with a customer's name that matches a brand?

Names are normal on personalised goods, so the blocklist should flag only exact brand phrases and well-known slogans, not common first names. Send a flagged name or phrase to the review queue, where a person decides, and keep a note of the decision so the same case is handled the same way next time.

Is it safe to paste customer photos into an AI tool?

Only if the tool's terms for your plan allow it and you are comfortable with where the images go. Personal photos, especially of children, deserve extra care. Consumer plans may use content for training unless you switch that off, so check your plan's data settings before you upload any customer image.

What if my print provider rejects an order after I accepted it?

Refund or offer a changed design promptly, explain which rule applied, and add the cause to your blocklist or review notes. Read your provider's content policy once a quarter, since their rules decide what they will print, whatever your own checks allow.

Further reads

• Print on Demand AI Generators Compared: Resolution and Licensing — Resolution and licensing for AI image generators.
• Print on Demand Mockups With AI: Realistic Shots Without Samples — Show customers their personalised design before printing.
• AI Trademark Search for Small IP Firms: What It Finds and Misses — What an AI trademark search finds and misses.
• What to Check in an AI Tool's Privacy Policy and Terms — Read the privacy terms before uploading customer images.
• AI-Generated Logos and Trademarks: What Designers Can Promise Clients — What designers can promise about AI-made logos.
• Chargeback Response Template for Online Stores, Drafted With AI — Respond to disputes if a printed order is challenged.
• Web-to-Print Software: Online Storefronts, Instant Quotes and AI — Storefront plans from $49 to $439 a month, how instant-quote rules work, and where AI genuinely helps with quotes and proofs.
• DTF Printing Business: Price Transfers, Gang Sheets and Reorders With AI — Work out cost per foot of a DTF gang sheet, set length-based prices with margins, and automate the quote and reorder emails.
• Sublimation Printing Business: Price Jobs and Proof Designs With AI — A worked 24-mug quote built from blank, ink, paper, press time and spoilage, plus the proof email and AI prompts that keep both honest.
• AI Tools and AI Development: The Complete 2026 Guide — the AI hub, including every tutorial in the AI-for-business series.

Sources: general practice for personalised-product shops; print providers' content policies and marketplace rules differ and change, so read your own providers' current pages. No vendor pricing used.