How to Use AI for Purchasing: Comparing Supplier Quotes Line by Line

To compare supplier quotes with AI, extract each quote's lines into one table, convert everything to the same unit, add freight, set-up charges and payment terms, then calculate the landed cost per unit you actually need. Have the AI do the extraction and the first pass at the sums, and then check every total by hand before you decide.

The lowest unit price is often not the cheapest order. Minimum order quantities force you to buy more than you need, freight can be quoted as "estimated", and a quote priced per hundred looks nothing like one priced each. The sections below walk through one realistic case, three quotes for printed cake boxes, and the places where the AI needs checking.

Follow me on Instagram@sagnikteaches

The purchase: 6,000 printed cake boxes for a year

A small bakery uses about 6,000 cake boxes a year, around 500 a month, printed with its logo in one colour. It has asked three suppliers for quotes. The quotes arrive in three shapes, which is normal:

Connect on LinkedInSagnik Bhattacharya

• Quote A, a PDF: $0.42 per box, minimum order 5,000, one-off print plate $90, freight $210 flat, payment net 30 days, price valid 14 days, delivery in 3 weeks.
• Quote B, an email: $0.37 per box, minimum order 10,000, no plate charge, freight "estimated at $380, charged at cost", 50% payment up front and the balance on delivery, delivery in 5 weeks.
• Quote C, a spreadsheet: $39.00 per 100 boxes, sold in multiples of 500 from a minimum of 2,000, free freight on orders over $400, plain unprinted boxes, with a separate adhesive logo label at $0.04 each, payment net 14 days, delivery in 2 weeks.

(These figures are illustrative.) Look at what a quick glance suggests: Quote B has the lowest price per box at $0.37. Whether it is the best buy depends on what you want to hold, how much cash you can tie up, and what the printing really involves. The line-by-line table settles it.

Subscribe on YouTube@sagnikteaches

Step one: get every quote into the same shape

Paste or upload the three quotes into an AI assistant and ask for extraction only, with no comparison yet. Separate the two jobs, because mistakes in the first poison the second. This prompt asks for blanks rather than guesses:

I will paste three supplier quotes for the same product. Extract each into a table with these fields, one row per supplier:
supplier label (A, B, C), item description, unit price as written, unit the price is quoted per, minimum order quantity, order multiple, set-up or tooling charge, freight as written, freight status (fixed / estimated / free over threshold), payment terms, price validity, lead time, what is NOT included.
Use only what the quote says. If a field is missing, write "not stated". Do not calculate anything yet. Quote the exact words for freight and payment terms.

Quote A: {{paste}}
Quote B: {{paste}}
Quote C: {{paste}}

Illustrative output for Quote C: "Unit price as written: $39.00; unit: per 100; MOQ: 2,000; multiple: 500; set-up: not stated; freight: free over $400 (free if order is over $400); freight status: free over threshold; payment: net 14; validity: not stated; lead time: 2 weeks; not included: logo (label sold separately at $0.04 each)."

The most valuable column is "what is NOT included". It is where hidden costs live. Quote C's boxes are not printed, so its low headline price omits the logo. The AI caught that only because the prompt asked for it.

Read the extraction against the originals

Take five minutes to check the table against each quote, and fix any cell that is wrong. Typical faults: a MOQ taken from a different line of a multi-product quote, a price per hundred read as per unit, "net 30" understood as "30% deposit", and a missing mention of a surcharge in a footnote. Models are good at reading, but a footnote in small grey print is the sort of thing both a person and a model skip. If the quote is a scan, expect more digit errors; read the number, not the summary.

Step two: put everything on one unit and ask what you actually need

Before any sums, decide the quantity you are pricing. This is a buyer's decision, not an AI's. The bakery needs 6,000 boxes over twelve months. That is more than A's minimum of 5,000, less than B's 10,000, and a clean multiple for C's 500s.

Then a second prompt, with the extracted table pasted in:

Using the table above, calculate for each supplier the cost of buying enough to cover 6,000 boxes in a year.
For each, show: units I must buy (respecting MOQ and multiples), boxes left over, product cost, set-up charges, freight, any extras needed to match the printed spec (for the unprinted option, add the label at its price), total cash out for the order, and cost per box USED (total cost divided by the 6,000 I need).
Show every multiplication and addition on its own line. Flag any assumption you had to make.

Here is the answer you should arrive at, and the arithmetic to check:

Quote B looked cheapest per box and is the most expensive per box used, because 4,000 boxes sit in storage for about eight months. If the bakery will reorder and use all 10,000 within the quote's life, B's cost per box is $4,080 / 10,000 = $0.408, slightly under C's $0.430, but only if all 10,000 get used, and with $2,040 paid before any box arrives. The AI will often present only the $0.408 figure unless you ask for cost per box used. That is the number to prompt for.

Step three: price what the quote does not say

A quote's silence is information. Ask the AI to list what is missing, then decide which gaps matter enough to question. For this case:

• Quote A: it does not say what happens to the plate. Is it yours to keep, and does reordering in six months trigger a second charge? If the plate is reused, the $90 is a one-off; if not, it recurs.
• Quote B: freight "at cost" has no ceiling. Ask for a not-to-exceed figure. A heavy, bulky pallet quoted at $380 can become $520.
• Quote C: the label is an extra step: someone must stick 6,000 labels. At 10 seconds each that is 16 hours and 40 minutes, around $300 of time at $18 an hour. Is the stuck label as good as a printed box? A person decides that, not the table.

Two further lines belong in every comparison: price validity and currency or tax treatment. A quote valid for 14 days that you act on in 20 is not a quote. If prices exclude taxes or charges that your business pays, the three quotes should be compared on the same footing, with the treatment stated by each supplier.

Turning gaps into questions you can send

Ask the AI to draft one short email per supplier with only the open questions, then edit it so it sounds like you. For Quote B:

Hi [first name],

Thanks for the quote on the printed boxes. Before I decide, could you confirm: (1) the maximum freight charge for the 10,000-box order, (2) whether you can split delivery into two lots of 5,000 on the same price, and (3) whether the 50% deposit is refundable if the proof is not approved. Please also confirm the price holds for 30 days.

Thanks, [your name]

The second question is the useful one. Suppliers with high minimums often agree to split deliveries, which changes the comparison without changing the price. Ask for it; it costs nothing.

Check the totals by hand, and why

Language models handle reading well and multi-step arithmetic less well, especially with many small figures, mixed units and per-hundred prices. Check every line of the final table with a calculator or a spreadsheet, not by asking the same model to check itself. A quick method that catches nearly everything: re-add each supplier's total in a spreadsheet, using cell formulas for quantity x price. If the AI's total and yours differ by even a cent, find out why.

The three faults seen most often in comparisons like this one:

1. Unit slips. $39.00 per 100 treated as $39.00 per box, which would make Quote C $234,000. Obvious when huge, quietly harmful when it is off by a factor of 10 on a smaller item.
2. Applying the MOQ and then forgetting it. Totals computed on 6,000 boxes for a supplier whose minimum is 10,000.
3. Freight "free over $400" ignored. A model that does not read the condition may add freight to Quote C anyway, or waive it on an order that does not qualify.

For the longer story on why a model gets sums wrong and what to test, see why AI is bad at maths and what to check in quotes and invoices.

The handling charge in paragraph three of a tube quote

An illustrative case: a bicycle repair shop needed a restock of inner tubes and brake pads. One wholesaler's quote showed $2.10 per tube in the headline and "$0.95 handling per carton of 24" in the third paragraph of the email. The shop pasted the email into an assistant and asked for the cheapest option, and the model answered from the headline, ignoring the carton line. The shop ordered 480 tubes, paid $1,008 plus $19 handling, and a rival wholesaler's $2.18 with no handling would have cost $1,046. It saved little, but the real damage came from payment terms: the cheap wholesaler wanted payment before shipping, so $1,027 left the account three weeks before the stock sold, in the month rent was due.

The corrections are simple: ask the extraction prompt for "extras and surcharges, anywhere in the document", and put "days between paying and being paid" in the table as a column. Cash timing is often the real difference between two suppliers whose totals are close.

Payment terms as money, not as small print

Terms such as "net 30" mean you pay 30 days after the invoice. A supplier offering net 30 is lending you the price for a month. A supplier wanting 50% up front is borrowing from you. If your cash is tight, that matters more than a 3% price gap. A rough way to compare: put the cash out and its date in two columns, then ask which order leaves the bank balance healthiest at the lowest point in the next 60 days.

Early-payment discounts such as "2% off if paid in 10 days, otherwise net 30" are worth working out. Paying 20 days early to save 2% is roughly a 37% annual return, which beats most things a small business can do with cash, but only if the cash is free. Ask the AI to calculate it and then verify with a calculator, because annualised-rate maths is a place where models improvise.

Seeing the table across a second kind of purchase

The same table works for different goods. An HVAC contractor buying twenty identical condenser units will weigh warranty length and pallet delivery against a 4% price difference: add warranty years as a column and ask who pays labour on a warranty claim. A solar installer comparing three inverter quotes needs lead time (a six-week delay can push back an install schedule), and whether quoted prices are firm or tied to a currency or a raw-material index. In both cases the extraction prompt gains two fields, warranty terms and price-adjustment clause, and the "not included" column does the heavy lifting again.

For ongoing relationships rather than one-off orders, the table becomes a record. AI supplier management covers tracking prices and lead times over time, and automating purchase orders and supplier emails covers the step after you choose.

What if demand changes? Re-running the table in two minutes

The bakery's 6,000 boxes assumed steady sales. Suppose a new café customer lifts demand to 9,000 boxes a year. With the extraction table saved, you only change one input and re-run the second prompt, which is the real benefit of doing the first step properly. At 9,000, A costs 9,000 x $0.42 = $3,780 plus $90 and $210, so $4,080, or $0.453 a box. B costs the same $4,080 for 10,000 boxes, leaving 1,000 spare, which is $0.453 a box used, level with A and with a 1,000-box buffer thrown in. C costs 90 x $39.00 = $3,510 plus $360 of labels, $3,870, or $0.430 a box. The gap between A and B closes as the volume grows, and C stays cheapest in cash because its price is linear and has no minimum. Add the label time, though (9,000 labels at 10 seconds is 25 hours, about $450 at $18 an hour), and C rises to $4,320, or $0.480 a box, behind both. The more boxes you need, the more the sticking job matters.

That flip is the sort of thing a one-line comment from a salesperson ("the more you buy, the better it gets") hides. It also shows why you should keep the quantity as a labelled input, not buried inside a sentence of the prompt. In a spreadsheet, put the quantity in one cell and make every total refer to it, then ask the AI only to write the formulas, not to produce the numbers.

Setting the comparison up as a reusable sheet

Once you have done this twice, a sheet beats a chat. Ask the assistant to build a template with one input block (quantity needed, hourly cost of your time, number of months of storage you can bear) and one column per supplier. Rows: unit price, unit conversion factor, MOQ, order multiple, units bought, product cost, set-up, freight, extras, total cash out, left over, cost per unit used, payment due date, cash out date, validity end date. Use formulas that look up the MOQ with MAX, for example units bought = MAX(MOQ, CEILING(needed, multiple)), so changing the quantity redoes the table. Have the assistant explain each formula in plain words in a notes column, and test it against the bakery figures above: you should see $2,820, $4,080 and $2,580 come out.

The sheet also gives you a record. Six months later, when a supplier says "our prices have not changed", you can open the old table and check. Keep the original quotes in a folder with the same names as the column headings, so anyone covering for you can find them.

A short checklist before you place the order

• Every unit the same: per box, not per hundred or per carton.
• Minimum order and order multiple applied to the quantity you really use.
• Set-up, tooling and sample charges included, and whether they recur.
• Freight: fixed, estimated or free over a threshold, and whether you reach it.
• Anything the spec needs that the quote leaves out (the label, the pallet, the adapter).
• Payment terms turned into a date and an amount.
• Price validity date on the calendar.
• Totals recomputed by you, to the cent.
• One line on what you will do if the delivery is late.

Print or save that list. It matters more than the choice of AI tool, and it fits equally on a clipboard.

Using the result in a negotiation, without bluffing

A comparison table is a useful thing to hold when you phone a supplier. If Quote A is $0.47 a box all in and Quote C $0.43 before label time, you can say plainly that you have a landed price of $0.43 elsewhere and ask whether A can include the plate or the freight. Do not invent a competing quote you do not have, and do not paste a competitor's confidential quote into an email to another supplier. Describe your own numbers only. Ask the AI to draft the call notes or the email, and take out any phrase that sounds scripted.

Often the best result is not a lower price but a better term: a split delivery, a longer price validity, a plate kept for you, or net 45. Those cost the supplier little and show up in your cash.

When the decision is not the cheapest line

With the table filled in, C is cheapest in cash at $0.43 a box, but with the $300 of label time from step three it comes to $2,880, or $0.48 a box, a penny above A. That is close enough for the other columns to decide. This bakery chooses C, because its counter staff can stick labels in quiet afternoons they are already paid for, so the $300 is not new spending; about 20 minutes a week, every week, goes on the job. It also asks A to requote after six months with the plate kept on file. A bakery whose staff are busy all day should choose A and its printed boxes, and a different bakery with a long storage room and a steady 12,000 boxes a year would reasonably choose B with split deliveries. None of these answers comes from the AI. It saved an afternoon of copying and caught the label, and the owner made the call, with figures she had checked herself.

If you would rather set this routine up with someone, that is the kind of job my AI implementation consultation covers; you can reach me through the contact page. For quotes that arrive as photos or scans, what ChatGPT can and cannot read shows where extraction goes wrong first.

Questions about using AI on supplier quotes

Is it safe to paste supplier quotes into ChatGPT or Claude?

Quotes are commercial information, so use a business plan or a setting that excludes your chats from training, and strip account numbers, bank details and your contact names first. The comparison needs only line items, quantities, terms and prices. If a supplier's quote carries a confidentiality clause, read it before sharing.

Should I let AI pick the winning supplier?

No. It can rank on cost, but the decision involves reliability, relationship, cash flow and risks that are not in the quote. Have it show the table and name the trade-offs, then decide yourself. Recheck every total on a calculator before anything is ordered.

What if the quotes arrive as PDFs or photos?

Current chat assistants read most PDFs and photos, but scanned or low-resolution pages produce misread digits. Ask for the extracted lines first, read them against the original, and only then run the comparison. Type any figure it flags as unclear.

How many quotes are worth collecting?

Three is a practical number for most small purchases: enough to see a spread, few enough to read closely. For a regular, high-value item, add a fourth supplier now and then to keep the others honest. For a one-off $50 item, two is plenty.

Further reads

• How to Compare Supplier Quotes Side by Side With AI — A general version of side-by-side quote comparison.
• How to Automate Buyer and Vendor Follow-Up With AI — Chase suppliers for answers without nagging.
• How to Match Supplier Invoices to Purchase Orders Automatically — Check the invoice against the order afterwards.
• AI Invoice Processing: Stop Typing Supplier Bills by Hand — Stop typing supplier bills by hand.
• How to Evaluate an AI Implementation Proposal or Quote — Apply the same line-by-line reading to a software quote.
• Can AI Fill In Forms and Supplier Portals for You? — What AI can and cannot do inside supplier portals.
• Construction Estimating With AI: Turn Supplier Quotes Into a Rate Sheet — Turn supplier quote PDFs into checked rates for your estimates: the extraction prompt, unit traps, five tests, an addenda log and a scope-gap review.
• Alibaba Sourcing With AI: Vet Suppliers and Plan a Sample Order — A supplier scorecard, a red-flag check for messages and a sample-order plan, with the document checks left to a person.
• How to Use AI in a Small Business: Perplexity for Research Jobs — A competitor research brief for a bicycle shop, run in Perplexity, with the source checks that decide which findings you keep.
• AI Tools and AI Development: The Complete 2026 Guide — the AI hub, including every tutorial in the AI-for-business series.

Sources: general purchasing practice; ChatGPT pricing page (training opt-out by plan) and Anthropic's privacy centre on model training, checked 5 November 2026.