Systems and records
Refused, returned or redirected: procurement rejection decision data
By SourceX Editorial · Reviewed by Noah Loul ·
10 sections
Short answer
Procurement rejection decision data should record why a purchase request was declined, not only that it was. Code each refusal to one of seven grounds (policy, authority, budget, need, an existing contract, stock on hand or a defective requirement), attach the record that proved it on the decision date, and keep refusals apart from redirects, returns and withdrawals. In the systems we audited, whole-request refusals get only free text.
System export
Refused
Key takeaways
- A rejected status is not a refusal label: in Dynamics 365 a rejected requisition can be resubmitted, and only approved lines can be canceled.
- Seven grounds cover most refusals, and each needs a record that proves it on the decision date, from a policy version to an inventory snapshot.
- In our audit, 7 of 16 OpenBoxes line-change codes and 4 of 8 Prozorro tender cancellation codes are refusal grounds; authority and stock on hand never appear.
- Whole-request refusals in OpenBoxes and plan cancellations in Prozorro take free text only, the stage where an approval model needs labels most.
- Label refused, redirected, returned, deferred and withdrawn separately, and pair refusals with matched approvals so a model does not learn to over-refuse.
What should a declined purchase request leave behind?
Four things: a coded ground for the refusal, the record that established that ground on the decision date, the outcome type, and the person who decided and the authority they used. Without the ground, a dataset can teach a model that some requests get declined, but not which ones should be.
Requests that should have been refused come in two populations. One is the requests that were refused, correctly or not. The other is requests that were approved and later found to be wrong, which audits catch and refusal logs never see. GAO's purchase card audit guide (an exposure draft from May 2003) sorts the second group into improper purchases, intended for government use but not permitted by law, regulation or organization policy; abusive purchases, authorized items bought at excessive price or quantity or for a questionable government need; and fraudulent purchases, unauthorized items for personal use or gain. Those audit findings are the negative examples most refusal datasets lack.
This page is about saying no. Urgent buys that skip the normal sequence and the validity of a stand-in approver have their own pages, linked at the end; here the question is how a dataset should label a refusal so the reason survives.
Rejected is a status, not a reason
A status records where a document stands, not why, and in common systems it is not even final. Microsoft's Dynamics 365 documentation says a purchase requisition or line that has been rejected can be modified and resubmitted, and that the canceled status applies only to approved lines that are no longer required. A dataset that reads Rejected as a refusal will count corrected and resubmitted requests as refusals and approved-then-canceled ones as approvals.
Some refusals happen before any approver sees the request. In Dynamics 365 budget control, a transaction that exceeds available funds triggers a warning or an error depending on setup. Where over-budget processing is allowed only for a permission group, a user outside that group gets an error and the transaction is blocked, and failed checks are listed on a budget control errors and warnings page. A refusal dataset built only from approver actions misses every request the budget check stopped.
Policy can refuse earlier still. Procurement policies in the same product control which requisition purposes, and for replenishment which product categories, a requester can use at all, so many requests that should be refused never become records. A dataset card should say which of these upstream filters were active, because they shape what the approver ever had to decide.
Where three public systems store the reason
In fixed code lists for trimmed lines and later procurement stages, and in free text for the refusal of a whole request. We read the source files of two open systems and one vendor's documentation at pinned commits, extracted every reason code with a script and assigned each to a class with a written rule, set out in full in the methods section that closes this page.
OpenBoxes, an open-source supply chain system, returns 16 reason codes when a stock request line's quantity is changed. We classed 7 as refusal grounds: 4 policy or stock list rules (such as a non-formulary product with no substitute, or an item not on the stock list), 1 lack of need (insufficient quantity consumed to merit resupply), 1 existing source (supplied by government) and 1 requirement defect (data entry error). The other 9 are not refusal grounds, and 6 of those describe being unable to fill the line. Rejecting a whole request is different: the form asks the approver to provide a reason for rejecting the request and makes the comment field required, but offers no code. A Duplicate Request code exists in the system's code list, though not among the 16 offered for a line change.
Prozorro, Ukraine's electronic public procurement system, runs on open-source code with an open API. A tender cancellation carries a reason type, and across procedure types the documentation lists 8 distinct codes. We classed 4 as refusal grounds: no further need, a cut in expenses, a legal violation that cannot be fixed, and missing localization requirements. The other 4 are force majeure, a code whose description repeats force majeure's, fewer than two bids, and a deadline violation. Canceling a procurement plan, the stage closest to an internal request, records only a required free-text reason, a status and a date.
Across the 24 codes in the two lists, budget appears once and authority and stock on hand appear zero times. The Dynamics 365 requisition overview defines statuses and describes no reason code. Both open systems show the same pattern: the earlier the refusal, the less structure it gets, and that early decision to approve or decline a request is exactly where a model needs labels most. The table below summarizes the audit.
| System and record | How the reason is stored | Codes | Refusal-ground codes |
|---|---|---|---|
| OpenBoxes: request line quantity changed | Coded reason from a fixed list | 16 | 7 (policy 4, need 1, existing source 1, requirement 1) |
| OpenBoxes: whole request rejected | Required free-text comment | 0 | None; text only |
| Prozorro: tender canceled | Coded reason type plus required free text | 8 | 4 (need, budget, policy, requirement) |
| Prozorro: procurement plan canceled | Required free-text reason only | 0 | None; text only |
| Dynamics 365: requisition rejected | Status that can be resubmitted; no code described | None described | None described |
Seven grounds for refusal and the record that proves each
Each ground is a test that can be run on the decision date against a specific record. If the record does not establish the test, the refusal was a judgment call and should be labeled that way rather than forced into a ground.
Public rules show several grounds written down. FAR 8.002 tells agencies to meet supply requirements first from their own inventories, then from excess held by other agencies, before other mandatory sources. FAR 13.003 makes simplified procedures inapplicable when an agency can meet the requirement through required sources, existing indefinite delivery contracts or other established contracts, and forbids breaking a requirement into smaller purchases to stay under a threshold. The Uniform Guidance at 2 CFR 200.318(d) requires a federal award recipient's procedures to avoid acquiring unnecessary or duplicative items. Private companies put equivalent rules into purchasing policy, and the matrix below translates them into checks a dataset can carry.
| Ground | Test on the decision date | Record that proves it | Usual outcome |
|---|---|---|---|
| Policy | The item, category or supplier is barred by a written rule then in force | Policy version, catalog or category rule | Refuse and cite the rule version |
| Authority | The requester or approver lacks the limit, or one need was split to stay under it | Delegation matrix and linked requests for the same need | Refuse or route to the right approver |
| Budget | Funds are not available for the account and period | Budget check result or funds-available report | Defer to a funded period or refuse |
| Need | The same need is already requested or ordered, or usage does not support it | Open requisitions and orders for the item and site; usage history | Refuse as duplicate or not needed |
| Existing contract | A mandatory source, framework or blanket agreement covers the item | Contract or catalog entry valid on the date | Redirect to that source |
| Stock on hand | Usable stock, net of reservations, covers the need here or at another site | Inventory snapshot with reservations | Redirect to a stock issue or transfer |
| Requirement defect | The specification, quantity or account coding is incomplete or wrong | Request version and reviewer note | Return for correction, not a refusal |
Refuse, redirect, return, defer or withdraw: five labels, not one
Give each outcome its own label, because each teaches a model a different behavior. Most rejection rates quoted inside companies mix all five, which is why they say little about whether the right requests were stopped.
- Refused: no purchase will meet this need; the ground and its evidence are attached and the decision is final.
- Redirected: the need is real but is met another way, through stock or an existing contract; the request was right to raise and wrong to buy.
- Returned: the request needs correcting; the corrected version is a new decision and should be linked, not merged.
- Deferred: valid but not fundable now; link it to any later request for the same need.
- Withdrawn: the requester canceled before anyone decided; keep it as its own class or leave it out of refusal training.
The refusal decision packet
One packet per declined request, built only from what the decider could see on the decision date, with later events added in their own fields. Keep the approver's own words: the code adds structure, it does not replace the text.
At a glance
The refusal decision packet
- Request version and timestamp, with the item or category, quantity, value band and cost center.
- Primary ground code, with an optional secondary ground when two tests failed.
- Evidence links valid on the decision date: policy version, budget check ID, open order or requisition number, contract ID, inventory snapshot.
- Decider role and the authority relied on, such as the delegation rule and limit.
- Outcome label from the five above.
- The free-text reason exactly as written.
- Downstream events: resubmission, stock transfer receipt, a purchase of the same item for the same need within a stated window.
- Reviewer agreement: whether an independent reviewer assigned the same ground from the same evidence.
When the quicker yes is still the wrong answer
Approving is usually faster and buying off contract often shows a lower unit price, yet both can defeat what the purchase was for. A refusal dataset earns its value on these cases, where the quick yes passes every field check and still fails.
The opposite error matters as much. A set made only of refusals teaches a model to refuse. Pair every refusal with approved requests matched on category, value band and requester type, so the model learns where the line sits rather than that lines exist.
- A spot buy below the contract price that breaks a volume commitment or a mandatory-source rule.
- Two requests just under an approval limit for one need: each passes alone, the pair fails.
- A purchase approved because budget exists while usable stock sits at another site.
- A budget refusal that sends the requester to a card purchase outside the controls the refusal was protecting.
Worked example (illustrative): forty laptops, three answers
In this invented case one request gets three different answers, and only a coded record shows it. Every figure is an illustrative count, not data from any company.
A department asks for 40 laptops. The checks find an approved requisition from the same cost center, entered a week earlier, for 25 of the same model; a framework agreement with a reseller that covers the model; 9 returned units in usable condition in IT stock; and a budget check that passes. The decision: refuse 25 as a duplicate need, redirect 9 to a stock issue, and redirect the remaining 6 to the framework agreement.
Stored as a status, the history reads Rejected, followed a few days later by a new request for 6. Stored as packets, it shows three grounds, three evidence links and three outcomes, and a model trained on it can learn to check open requisitions and stock before approving, which is the behavior the company wanted.
Testing a model on refusals, and where this design stops
Test refusal judgment on balanced sets: refusals by ground, matched approvals and a held-out group of approvals that audits later flagged. Grade the ground and the evidence cited, not only the yes or no, and score returns separately from refusals so a model is not rewarded for refusing requests that only needed fixing.
The audit has limits. It covers two open systems and one vendor's documentation; commercial products may offer reason codes we did not inspect, and deployments can add their own. The classification of each code is editorial, though fully listed below. Public procurement rules illustrate the grounds; private purchasing policies differ by company. The laptop case is illustrative and is not measured research.
SourceX sources datasets from U.S. businesses on request rather than from stock, and buyers set out scope, volume, language, format, timing and licensing needs on the request form. A request for refusal episodes is most useful when it names the intended use and an acceptance test, such as recomputing the ground from the attached evidence.
How the reason-code audit was counted
Files and commits: OpenBoxes at commit d28bf7c (October 26, 2026), the reason code enum, the reason code API controller, the reject request form and the English message file; Prozorro's openprocurement.api at commit f13a6ba (October 27, 2026), the tender cancellation and plan cancellation documentation and the plan cancellation model; Microsoft's Dynamics 365 documentation at commit 7a30680 (October 27, 2026), the purchase requisition overview and the budget control overview.
Counting rules, exactly as the script applies them: OpenBoxes line codes are every constant inside the list method the API returns for a requisition line change, counted once. Prozorro codes are every distinct code listed under reason type in the tender cancellation documentation, counted once across procedure groups. Each code receives exactly one class, and the script stops if any extracted code has none. Policy: NON_FORMULARY_NO_SUBSTITUTION, NOT_STOCKED, SUPPLY_MAX_QUANTITY, NOT_ON_STOCK_LIST, unFixable. Need: INSUFFICIENT_CONSUMPTION, noDemand. Budget: expensesCut. Existing source: SUPPLIED_BY_GOVERNMENT. Requirement defect: DATA_ENTRY_ERROR, noLocalization. Non-ground classes: could not fill (STOCKOUT, LOW_STOCK, DAMAGED, EXPIRED, AVAILABLE_STOCK_RESERVED, INSUFFICIENT_QUANTITY_RECONDITIONED), pack-size rounding (PACKAGE_SIZE), withdrawn (CANCELED_BY_REQUESTER), unspecified (OTHER), external event (forceMajeure, noObjectiveness), market response (noOffer) and procedure timing (dateViolation).
Not measured: how often any code is used in practice, any commercial ERP beyond the documentation named, and configuration a deployment adds. The request-level findings are checks of whether a file contains a required free-text field or a reason type field, not counts.
Frequently asked questions
What is the difference between a rejected and a canceled purchase requisition?
In Microsoft's Dynamics 365, a rejected requisition or line can be modified and resubmitted, while canceled applies to approved lines that are no longer required. Neither status says why. A refusal dataset needs a separate ground code and an outcome label for that reason.
Which reasons justify refusing a purchase request?
Seven grounds cover most refusals: a policy rule, missing authority, unavailable budget, no genuine or a duplicate need, an existing contract that covers the item, usable stock on hand and a defective requirement. The last three usually end in a redirect or a return for correction rather than a final refusal.
Do ERP systems record why a purchase request was rejected?
Often only as free text. In the systems we read, OpenBoxes requires a comment when a whole request is rejected and Prozorro requires a free-text reason when a procurement plan is canceled, while both use fixed codes for later or line-level changes. Check what your own system stores before assuming a reason field exists.
How do you train a model not to over-refuse purchase requests?
Pair each refusal with approved requests matched on category, value band and requester type, and grade the ground and evidence the model cites. Score returns for correction separately, so refusing a fixable request does not count as a correct refusal.
Are approved purchases that auditors later flag useful training data?
Yes, they are the negative examples refusal logs never contain. GAO's purchase card audit guide groups them as improper, abusive or fraudulent purchases; labeled with the ground that should have stopped them, they test whether a model would have caught what the approver missed.
Sources
- GAO's 2003 purchase card audit guide (exposure draft) defines improper, abusive and fraudulent purchases. Source
- Microsoft documents purchase requisition statuses: rejected requisitions can be resubmitted and only approved lines can be canceled. Source
- Microsoft documents budget control checks that warn or block over-budget transactions and list failed checks. Source
- OpenBoxes defines its reason codes and the list offered when a requisition line quantity changes. Source
- OpenBoxes returns the quantity change reason list for the requisition line modification activity. Source
- OpenBoxes's reject request form requires a free-text comment. Source
- OpenBoxes's English message file holds the rejection prompt and the labels of each reason code. Source
- Prozorro's API documentation lists the reason type codes for tender cancellations by procedure group. Source
- Prozorro describes itself as a fully electronic public procurement platform for Ukraine with open source code and an open API. Source
- Prozorro's plan cancellation schema holds a required free-text reason, a status and a date. Source
- FAR 8.002 sets the priority order of mandatory government sources, starting with the requiring agency's inventories. Source
- FAR 13.003 points requirements to existing sources and contracts and forbids splitting requirements to avoid thresholds. Source
- 2 CFR 200.318(d) requires recipients' procedures to avoid acquiring unnecessary or duplicative items. Source
Tell us what data you need
Share scope, volume, format, timing and licensing needs. Datasets are sourced on request.