Subscription Box Curation: Match Items to Surveys With AI

Turn each survey answer into a tag, tag every item the same way, then let a spreadsheet remove anything excluded, check stock, and cap the box cost. Ask AI to rank only the items that survive those filters, then read every box list yourself before packing. Exact rules go in the sheet; judgement goes to the assistant.

The reason is simple. A chat assistant is good at reading "I like soft colours but hate glitter" and bad at remembering that you have 40 brush pens for 52 people. Stock and margin are arithmetic, so give them to formulas, and keep the assistant on the language.

Follow me on Instagram@sagnikteaches

The box this tutorial builds: 120 subscribers, five items each

Say a monthly craft box has 120 subscribers at $34 a month. Each box holds five items: one main tool, two consumables, one paper product and one extra. The owner buys in one batch per month, packs on a Thursday and ships Friday. Per box, packaging costs $2.10, postage $6.20 and card fees about $1.35, so $9.65 goes out before any items are bought. To keep a contribution of at least 25 per cent ($8.50) after items, the five items may cost no more than $15.85 together. That cap is the margin check in this workflow; everything else bends around it.

Connect on LinkedInSagnik Bhattacharya

If you have not yet set your own cap, building a subscription box margin model shows how to derive it, and the rest of this page assumes you have one number to defend.

Subscribe on YouTube@sagnikteaches

Survey fields that change what goes in the box

A survey earns its place only if each answer can change a box. Draw a line from every field to a rule before you ask it. For the craft box the owner uses seven fields, all with fixed answers except the last.

Notice what is missing: age, job, income, free-text "tell us about yourself". Those cost completions and change nothing in the box. Keep each subscriber on one row with an ID, the seven answers, the date they last answered and the items they have already received. That last column matters later, because repeating an item is the quickest way to disappoint a long-term subscriber.

Tag the items once, the same way as the survey

The table below is an excerpt of eleven lines from a catalogue of about forty.

Matching works when items and answers speak the same language. For each item in your catalogue, record the tags below. If your item list lives in photos and supplier PDFs, cataloguing stock from photos into item records shows a quick way to build it.

Keep the tag vocabulary short and closed: the colour column accepts only warm, cool, neutral or none, never "sunset" or "earthy". A chat assistant can fill in tags from your supplier descriptions, which saves an afternoon, but have it work from the product page text only, and spot-check ten rows. A misspelled tag silently breaks a rule.

Four matching rules, in the order they run

Write the rules down as a numbered list. Order matters: hard rules first, preferences last.

1. Exclude. Remove any item that carries a tag the subscriber avoids, any item they already own, and any item they received in the last six boxes.
2. Fill the slots. From what is left, choose one item per slot (tool, two consumables, paper, extra) that matches their interest and level.
3. Rank by preference. Where more than one item fits a slot, prefer the one that matches their colour leaning, then the cheaper one.
4. Check the caps. Total item cost must be $15.85 or less, and no item may be allocated beyond its stock.

Rules 1, 2 and 4 are exact and live in the sheet as filters and formulas. Rule 3 is where AI earns its place, because "which of these three eligible sets feels right for someone who wrote 'I mostly paint on holiday'" is a judgement the sheet cannot make.

Three subscribers walked through

Subscriber A: beginner, watercolour, warm colours, avoids nothing, owns no brushes. Eligible items after rule 1 include the pan set, the round brush, the paper block and the warm washi tape. The second consumable is a $1.90 water brush pen (not in the excerpt above). Cost: $6.20 + $2.40 + $3.80 + $1.10 + $1.90 = $15.40, under the cap. Contribution: $34.00 - $9.65 - $15.40 = $8.95, or 26.3 per cent.

Subscriber B: beginner, lettering, avoids scented items, owns a ruler. Nothing in the lettering list is scented, so the box is brush pen set $5.60 (tool), lettering practice pad $3.10 (paper), metallic gel pen $2.30 and a $1.60 lettering ink refill (the two consumables) and washi tape $1.10 (extra): total $13.70, contribution $10.65. A person who had listed metallic items under avoid would lose the gel pen and the rule would fall back to the next cheapest lettering consumable.

Subscriber C: improver, watercolour, cool colours. The preference ranking selects the gouache set at $9.80. The box becomes gouache $9.80, round brush $2.40, paper block $3.80, washi tape $1.10, water brush pen $1.90: $19.00. That breaks the cap by $3.15, so rule 4 sends the box back for a swap. The fallback is the pan set at $6.20, bringing the total to $15.40. The assistant should say why it swapped and what it cost the subscriber's preference; the owner may prefer to keep the gouache and absorb the $3.15 on this one box, logged in the override column described further down.

Let the assistant rank, but show its working

Do the filtering in the sheet and paste only the eligible items for a small group of subscribers into a chat assistant (ChatGPT, Claude or Gemini). Use IDs, not names. Ask it to return a table you can check.

You are helping me curate a monthly craft box. Each box holds exactly
5 items: 1 tool, 2 consumables, 1 paper, 1 extra. Total item cost must
be $15.85 or less. Use ONLY the eligible items listed for each
subscriber. Do not invent items.

For each subscriber ID: choose the 5 items, show each item's cost, the
total, and one line saying why this suits their notes. If the best set
exceeds $15.85, say which swap you made and what preference it cost.

Subscriber S-0417: improver, watercolour, cool colours.
Note: "I mostly paint on holiday, travel size would be great."
Eligible items: [paste the rows]
Subscriber S-0418: ...

An illustrative reply for S-0417 would offer a travel-friendly set, add a one-line reason ("pan set is compact, suits holiday painting"), total it at $15.40, and mention that gouache was dropped for the cap. What you fix: it may pick an item you only have 3 of, because it cannot see stock unless you paste it; it may total one line wrongly; and it may flatter the subscriber's note by choosing an item that does not exist. The sheet recomputes every total and stock check regardless. Treat the assistant's output as a proposal that the sheet confirms or rejects.

Stock limits: when demand outruns the batch

The 120 subscribers split 40 watercolour, 52 lettering and 28 journalling. You bought 40 brush pen sets, but all 52 lettering subscribers qualify for one. Twelve boxes are short. The sheet shows it in a single check: allocated count by item against stock, via COUNTIF on the allocation column, with a flag where allocated exceeds stock.

Decide the tie-break before you run it, not while subscribers wait. Three sensible orders: longest-standing subscribers first, subscribers who have not yet received a tool in that interest first, or subscribers whose last box scored lowest on the monthly feedback question first. The owner picks the second, so the 12 who miss out are those who have already had a different lettering tool, such as a fineliner set, in an earlier box. They get the single brush pen at $2.80 instead, which also lowers their box cost, and a note card that says honestly the set sold out.

If the shortage is regular, change the buying instead: order against survey counts, not last month's habit. Setting reorder points for ecommerce stock and tracking purchase orders with a supplier chase sheet cover the buying side once you know how many of each tag you need.

Manual overrides the sheet cannot know

Some boxes should never be left to rules. Add an "Override" column and a "Reason" column, and make the rule that a filled override beats the matching result.

• Complaint last month. A subscriber whose item arrived broken gets the owner's pick of the best item in their interest, even if it breaks the cap by a dollar. A single dollar is cheaper than a cancellation.
• Gift subscriptions. The recipient has not answered the survey; the giver has. Send the default box, adjusted only for any hard exclusion the giver mentions (an allergy, a dislike of scented items), until the recipient answers.
• New product test. Slot a new item into every tenth box and ask for feedback in the next survey. Without an override column, new items never get a trial because matching only chooses what it already knows.
• Allergy or safety flags. If a subscriber writes something about a skin reaction, a person reads that note, not the assistant.

The retention link is worth tracking: a mismatched box is a common reason to cancel, which is why reading cancel reasons and save offers pairs naturally with this sheet. Log the override reason so you can spot patterns. If six overrides in a month say "wrong colour", fix the tag or the survey wording rather than the box.

The note card in each box, written from the survey

A personalised box feels personal only if the subscriber can tell. A short card that names one thing they asked for does more than any amount of clever matching. Ask the assistant to write the card from the allocation row and the note, with a strict limit on what it may claim.

Write a 35-word note card for subscriber S-0417's box. Use only
these facts: improver, watercolour, likes cool colours, wrote "I mostly
paint on holiday". Items: pan set (compact), round brush size 6, cotton
paper block, washi tape, water brush pen. Friendly, plain, no
exclamation marks, do not promise anything about the next box.

An illustrative reply: "You said you paint on holiday, so this month's set is small enough for a bag. The pan set and water brush pen need no jar of water if you are short of room. Tell us how it travels." Fix any line that mentions a feature you did not give it. If the assistant writes "perfect for the beach" when the item is not waterproof, cut it. Cards that promise the next box will contain something you have not bought cause more trouble than a plain card.

Keeping the survey fresh without nagging

People change. A beginner is an improver after four months, and a "warm colours" answer from January may not hold in June. Rather than resurvey everyone, put one question in the shipping email each month: "Did last month's box suit you?" with three links (loved it, fine, not for me). Keep the answers in the subscriber row. Two "not for me" replies in a row trigger a person to read their record and perhaps write to them, which is usually worth more than a discount offer. Refresh the full survey once a year, or when someone's last answer is more than nine months old, and tag those rows so matching treats them with a little less certainty, for example by pairing one safe item with one stretch item.

A box that went wrong because of one bad tag

Picture a month in which 14 subscribers who ticked "avoid scented items" receive the scented marker set. The cause is not the AI. A bulk tag fill from supplier descriptions has labelled the marker set "Scented: No" because the product page says "fragrance-free ink" in one line and "cherry, lime and berry scents" in another. The matching sheet trusts the tag, the assistant trusts the sheet, and nobody checks the exclusion list against the item table.

The repair has three parts: the sheet counts, before packing, how many boxes contain any item whose tag matches a subscriber's avoid list (the answer should be zero, and the check takes one formula); a person reads the description of every item added to the catalogue for the avoid tags; and the 14 subscribers get a replacement set and a note. The point of the illustration is that a final check which does not depend on the matching logic catches what the logic cannot.

The pre-pack audit: six checks before the packing list prints

Run these on the allocation sheet every month. Each is a formula or a filter and takes a few minutes.

1. Every box has exactly five items, one per slot.
2. No box contains an item tagged with something the subscriber avoids.
3. No box repeats an item from the subscriber's last six boxes.
4. For each item, allocated quantity is at or below stock.
5. No box exceeds $15.85 in item cost, and the average box cost is shown, so you see the month's real contribution.
6. A person has read the five boxes with the highest cost and the five with a filled override.

Keep the count of failures in a small log. If check 2 ever fails, stop and find out why before anything is packed. After a few months, the pattern of what fails shows where to automate next. If the shelf stops matching the allocation sheet, rank the gaps by value before you recount anything, because the most expensive item that is short is the one that breaks a box.

What to hand to AI and what to keep

The split that works for a small box: formulas for exclusion, slot filling, stock and cost; AI for reading notes, ranking close calls, writing the card in each box and drafting the subscriber message when an item changes; the owner for overrides, new items and the final read. For 120 subscribers the monthly work can come down from a day or more of hand-picking to something nearer two or three hours: one to refresh tags and run the sheet, one to ask the assistant to rank the contested boxes, one to audit and fix. Your own numbers will differ with the size of the catalogue and the number of overrides.

The one thing not to automate is the decision that the box is good. The sheet proves the box is allowed; only a person who has held the items can say whether the five of them belong together. If you want to map where matching sits among your other first jobs, five first jobs for a small subscription box puts curation in context, and a 1:1 call can sort which of them is worth doing now.

Questions owners ask before they automate curation

How many survey questions should a subscription box ask?

Only ask what changes what goes in the box. Six to eight fields cover most hobby and consumer boxes: main interest, experience level, colour or flavour leanings, things to avoid, things they already own, and one free-text note. Every extra question lowers completion, and an answer you never use is wasted goodwill.

Should AI choose the whole box on its own?

No. Let rules in a spreadsheet handle exclusions, stock and margin, because those are exact. Use AI for the fuzzy parts, such as reading free-text notes or ranking a handful of eligible items, and have a person look over each box list before it goes to the packing bench.

What do I do for subscribers who skip the survey?

Give them a default box built from your best-selling items that has no common exclusions, and tag them 'no survey'. Send a short reminder with the survey link in the shipping email, and move them into matching the month after they answer. Never guess preferences from a name or address.

Is it safe to paste subscriber survey data into a chat assistant?

Only if you remove names, addresses and emails first and use an ID instead. Check that the plan you use does not train on your content by default, and keep the ID-to-person list in your own sheet. Free-text answers can contain personal details, so skim them or have the assistant flag them.

Further reads

• Farmers Market Stock Planning: How Much to Pack Each Week With AI — Plan quantities per week from demand you can count.
• How to Use AI to Set Reorder Points and Prevent Stockouts — Turn survey counts into monthly buying quantities.
• AI Supplier Negotiation Emails for Importers: Price, MOQ, Terms — Negotiate unit prices once you know your volumes.
• Backorder Notification Email: A Template AI Fills From Stock Data — Tell subscribers about a delayed item without panic.
• How to Find Dead Stock and Slow Movers With AI — Use up leftover items that never match anyone.
• Tea Shop Marketing Ideas: Tasting Notes and Gift Sets Written With AI — Turn your own cupping notes into shelf cards and product pages, then price subscription tiers and gift sets by margin before AI writes a word of copy.
• Veg Box Scheme Planning: Weekly Contents, Swaps and Costs With AI — Turn a harvest list into a weekly veg box mix that hits its value band, with substitution rules, swap messages and a quick cost check.
• Customer Subscription Model: Test a Subscription Offer With AI First — A two-week demand test for a subscription offer, with a margin model, a churn guess and honest sample sizes for a small customer list.
• AI Tools and AI Development: The Complete 2026 Guide — the AI hub, including every tutorial in the AI-for-business series.

Sources: Microsoft Support and Google Docs Editors Help pages for COUNTIF, checked November 2026. Costs, quantities and subscriber details in the examples are illustrative.