AI for Procurement: Compare Supplier Quotes Before Raising a PO

Three suppliers have replied. One sent a PDF, another an Excel sheet, and the third revised its price in a WhatsApp message.

Your purchasing executive now has to work out whether they quoted the same specification, whether delivery is included, and whether the lowest price requires buying twice as much stock.

This is a useful, specific job for AI procurement automation: prepare a reliable comparison and a draft purchase order, rather than let an agent decide what your business should buy.

Compare the purchase, not just the quoted price

Consider this illustrative example. A distributor in Klang needs 1,000 identical cartons delivered to its warehouse.

• Supplier A quotes RM2.40 per carton, plus RM180 delivery.
• Supplier B quotes RM125 per bundle of 50, with delivery included.
• Supplier C quotes RM2.30 per carton, with delivery included, but requires a minimum order of 2,000 cartons.

Supplier A totals RM2,580. Supplier B requires 20 bundles and totals RM2,500. Supplier C has the lowest unit price but requires a RM4,600 purchase and leaves 1,000 extra cartons in stock.

For this example, assume identical specifications and no additional charges or tax differences. Real comparisons must make those assumptions visible rather than quietly treating them as facts.

A useful comparison should show the required quantity, actual order quantity, total payable, delivery date and unresolved questions. Ranking three unit prices is not enough.

The useful output is not “Supplier C is cheapest”. It is “Here is the comparable cost, here is what differs, and here is what still needs checking”.

Give the agent a defined purchasing job

Start with one repeat-purchase category: packaging, office consumables or standard maintenance supplies, for example. Avoid beginning with custom machinery or materials where an incorrect specification could create a safety risk.

Define the purchasing request before collecting quotations. Include the item code, specification, quantity, delivery location and required date. If “equivalent product accepted” is allowed, say who can approve the substitution.

Then design the workflow to:

• Collect quotations from an agreed mailbox or document folder.
• Extract supplier details, quotation references, items, quantities, prices and terms.
• Map supplier descriptions to your approved item codes and units of measurement.
• Calculate comparable totals using fixed calculation rules.
• Flag missing information and prepare clarification messages for review.
• Send the comparison to the buyer, then prepare a draft PO after selection.

Document extraction is an established capability. Microsoft’s Document Intelligence layout model, for example, extracts text, tables and document structure from supported files. That does not mean it understands whether a quoted component is technically suitable.

Use AI to interpret varied wording and organise the evidence. Use deterministic calculations for quantities and totals. Keep commercial and technical decisions with the people responsible for them.

Make missing terms visible

The hardest part is not reading “RM125”. It is knowing what that amount covers.

For each quotation, capture the pack size, minimum order quantity, delivery charges, applicable tax treatment as quoted, payment terms, lead time and quotation validity. Record whether each value was explicitly stated or confirmed later.

Do not treat an empty delivery-charge field as free delivery. Do not assume “available” means delivery to your Johor branch by Friday. An agent should mark these as questions, not fill the gaps with plausible answers.

Keep a reference back to the original quotation page, spreadsheet cell or message for each important value. The buyer should be able to check a disputed number without searching through the entire conversation.

Revisions need particular care. Suppose a supplier sends a lower price but changes payment from credit to an advance deposit. The workflow should highlight both changes and preserve the earlier version. It should not simply overwrite the price.

Microsoft’s guidance explains how extraction confidence can help route uncertain results for human review. It also notes that not every field returns a confidence score. Test review rules against your own documents; do not treat a confidence score as proof that a quotation is commercially complete.

Keep your accounting system as the purchasing record

Before adding AI, check what your existing software already handles.

AutoCount’s purchase-order documentation describes transferring information from a Request Quotation into a PO, recording units of measurement and quantities, and viewing item price history. If your team is retyping information that an existing function can transfer, fix that first.

The AI layer should handle work around the purchasing record: interpreting incoming quotations, assembling the comparison and collecting missing details. It should not create a second, conflicting list of suppliers, items and purchase orders.

Integration still needs checking. AutoCount’s store lists a VT Full API SDK covering purchase requests, quotations and purchase orders, with a separately purchased AutoCount API Module as a prerequisite. That is an available route, not a promise that every installation is ready to connect.

For AutoCount, SQL Account or QNE, ask your implementation partner to confirm:

• Which installed version, modules and licences support the proposed connection.
• Whether the workflow can prepare a draft without releasing an order.
• How item codes, suppliers, branches and document references will be mapped.
• How retries will avoid creating a second PO for the same approved request.

Where a suitable connection is unavailable, a reviewed import or draft export may be the more sensible starting point. Do not promise fully automated purchasing before testing the actual system.

Budget for the work AI cannot remove

Expect to clean up item descriptions and units before comparisons become dependable. “Box”, “carton” and “pack” need agreed meanings. Supplier aliases need matching to the correct supplier account. Someone must own those mappings when packaging or specifications change.

Your team also needs an agreed place for quotations and revisions. If buyers continue approving one version while a newer price sits in a private chat, the workflow has an information problem, not a model problem.

AI may not be worth adding if purchasing is infrequent, quotations already arrive in a consistent template, or your existing software handles the task adequately. A standard request form and a better comparison spreadsheet may solve the problem more cheaply.

For specialist construction materials, engineered parts or subcontract packages, use automation to assemble documents and highlight differences. Leave technical equivalence, exclusions and suitability to qualified reviewers.

Test whether buying actually becomes easier

Run a controlled pilot on historical quotations first. Include awkward cases: different pack sizes, revised prices, missing freight charges and an unavailable item offered as a substitute. Ask an experienced buyer to establish the correct comparison before evaluating the agent.

Then test live requests with human review before any supplier receives a PO. Measure total staff time from receiving quotations to producing an approved draft, including corrections and follow-up work.

Track incorrect item matches, missed terms, duplicated drafts and how often buyers have to rebuild the comparison. Separate active working time from time spent waiting for suppliers or managers. Faster document preparation will not fix an approval that sits untouched for days.

Solver AI’s approach is to validate the workflow, test it in a controlled environment and then roll it out around the systems the business already uses.

Explore connected procurement workflows

The first goal is straightforward: give the buyer a comparison they can check and a draft order they do not have to retype. Expand only when that narrower job works reliably.