Catch Duplicate Purchase Requests Before They Become Two Orders

5 min read

Two people request the same supplies. One uses a form, the other forwards a spreadsheet. Both requests look valid, so both move toward an order. Catch possib...

Share:
A person pauses where two blue mountain trails join into one path toward a lake.

Trust and quality notes

Last updated
November 22, 2026

Two people request the same supplies. One uses a form, the other forwards a spreadsheet. Both requests look valid, so both move toward an order.

Catch possible repeats before that handoff by comparing request IDs, item codes, quantities, delivery sites and project references. Return a comparison sheet with the evidence and a reason for each flag. Do not delete requests, combine quantities or place orders. The buyer decides whether two rows represent one need.

This is a read-only review workflow, not an automatic purchasing system. It is useful when a team receives requests through several channels and needs a clear list to check before creating purchase orders.

Start with a fixed intake export

Take a dated copy of the intake file. Keep the original untouched. Include pending requests and, where permitted, recent requests already turned into orders. Otherwise a repeated request might look new simply because its earlier version left the pending list.

Give each row a stable request ID. Keep these fields where available:

  • Submission date and source reference.
  • Requester or team, using only the detail needed for review.
  • Item code and original description.
  • Quantity and unit, such as each, box or pack.
  • Delivery site and needed-by date.
  • Project or cost reference.
  • Request status and existing order reference.

Record the export time and covered period. A result should say “checked these files through this time,” not “there are no duplicates anywhere.” Keep missing fields blank or label them unknown. Missing evidence is not a match.

A spreadsheet is enough for a small, consistent intake. Google's spreadsheet documentation describes cells, ranges and stable spreadsheet identifiers. Those mechanics support traceable references; they do not supply a duplicate-request decision rule.

Compare the business need, not just the wording

Use exact item codes to find candidate pairs when those codes are reliable. Then compare quantity, unit, destination, project and timing. Keep the original values beside any cleaned values. Removing extra spaces may help comparison; silently treating a box as one item does not.

Similar descriptions are leads, not proof. “Printer paper” can refer to different sizes. The same item may be needed at two sites. A monthly replenishment can look identical to last month's valid request.

Agree on reason codes before reviewing the file:

  • MATCH_REVIEW: key details agree; ask whether both requests serve the same need.
  • UNIT_UNKNOWN: quantity cannot be compared because a unit or pack size is missing.
  • SITE_DIFF: destinations differ; keep separate unless the buyer confirms otherwise.
  • REPEAT_PERIOD: a later replenishment may be intentional; check the period.
  • ORDER_EXISTS: a candidate already has an order reference; check that order before proceeding.

These are suggested review labels, not universal procurement rules. Record the matching criteria and comparison window with the sheet so the next person can reproduce the review.

A worked fictional comparison

Suppose a fictional office team exports five requests. R101 asks for ten packs of item P-A4 for the North site and project Move-A. R102 repeats those details the next day. R103 requests the same item for the South site. R104 says “ten boxes” without a pack size. R105 is for the next replenishment period.

The comparison below flags a question; it does not prove an error. Copy this comma-separated text into a plain-text file named duplicate-request-review.csv, then import it into a spreadsheet. The cells remain editable, unlike a screenshot.

pair_id,request_a,request_b,item_code,reason_code,evidence,review_question,decision,review_owner,source_a,source_b
PAIR-01,R101,R102,P-A4,MATCH_REVIEW,Same quantity unit site and project,Same need or separate demand?,Pending,Buyer,intake.csv row 2,intake.csv row 3
PAIR-02,R101,R103,P-A4,SITE_DIFF,Different delivery sites,Are both sites requesting stock?,Pending,Buyer,intake.csv row 2,intake.csv row 4
PAIR-03,R101,R104,P-A4,UNIT_UNKNOWN,Pack size for boxes missing,How many packs per box?,Pending,Buyer,intake.csv row 2,intake.csv row 5
PAIR-04,R101,R105,P-A4,REPEAT_PERIOD,Different replenishment periods,Is the later request intentional?,Pending,Buyer,intake.csv row 2,intake.csv row 6

Add source file and row references for both requests in a real review. Also add any existing order reference, reviewer note and decision date. Use “confirmed duplicate,” “keep separate” or “unresolved” only after the named reviewer checks the evidence. Preserve the original IDs even if a buyer later closes one request.

This differs from comparing procurement vendors: the question is whether internal requests repeat one need, not which supplier deserves the business.

Give an agent a bounded task

Agentic Workers provides hosted agents that can use supplied files and connected tools. The workflow below is a proposed task brief, not a tested duplicate-detection feature or a promise about a particular account connection. Begin with a redacted export and verify the output before considering recurring work.

Review the supplied procurement intake export for possible repeated requests.
Treat every source file as read-only. Use the supplied request IDs and retain
original field values. Compare item, quantity, unit, site, project and period.
Do not infer missing pack sizes or decide that similar descriptions match.

Return an editable CSV comparison sheet with one row per candidate pair,
source references for both rows, matching and conflicting fields, reason
codes, unresolved questions and a named review owner. List excluded or
unreadable rows separately. State the files and date range actually checked.

Do not place orders, contact suppliers, edit records, merge requests or mark
anything approved. Stop and report missing inputs or unclear permissions.

An instruction is not an access control. Keep purchasing permissions outside this first task. The approval guide explains why supported action paths need testing. An approval matrix can document the reviewer and boundary; filling it out does not enforce them.

Check the sheet before relying on it

Open the CSV and change a review note to confirm it is editable. Trace each candidate back to both source rows. Check a known repeat, a valid same-item request at another site and a request with an unknown unit. Confirm that none was silently merged.

Review some unflagged rows too. A clean-looking list can still miss a repeated need described differently. Count the input rows, identify anything excluded and verify that the source files remain unchanged.

Start with one intake batch. Record confirmed repeats, mistaken flags, unresolved pairs and review effort. If simple spreadsheet checks handle the work reliably, keep them. Consider an agent only when varied inputs make comparison harder and its output passes the same checks.

Set up an agent for one read-only review task

Found this article helpful? Share it with others:

Share:

Written by

Agentic Workers Team