AI Adoption GuideProcurementEvaluate
Internal contradiction detection
AI cross-checks each vendor's executive summary, technical specification, and pricing for logical inconsistencies and flags them before scoring.
Procurement processRequestApproveSourceEvaluateSelectOrderReceiveReview
By Don, DoneThat’s AI coach · updated
Flag self-contradictions before anyone scores
The check is simple: does this vendor's packet agree with itself? Run it before anyone opens a score sheet. A mismatch inside one bid is a clarification, not a fail, and not a reason to mark another vendor higher.
Do not compare vendors here. Cross-bid ranking belongs in proposal scoring against RFP criteria. Do not test the packet against your shalls here either. That is compliance requirement extraction: your issued RFP versus what they answered. This page is only the vendor talking to itself.
If two passages disagree, quote both, name the field, and send the question. If they agree, leave them alone. Do not invent a count of typical inconsistencies. Some packets are clean. Some are messy because three authors wrote them in a week.
Sourcing and bid-evaluation suites in the Inventive AI, Zycus, Jaggaer, and Coupa class can ingest the files and surface candidate mismatches. Treat that output as a queue. The suite does not know which mismatch is a rounding difference, which is a last-year exhibit, and which would change the award. You do.
Cross-read the summary, spec, price workbook, and exhibits as one packet
Load four files for the same bid number and the same revision:
- Executive summary (the narrative the panel will remember)
- Technical specification (the method, BOM, SLA matrix, and schedule)
- Price workbook (quantities, unit prices, Incoterms, freight, options)
- Exhibits (site surveys, drawings, certificates, service matrices, delivery calendars)
Build a field map with one row per fact: promised lead time, quantity by site or SKU, Incoterm and named place, SLA (response, restore, exclusions), and the sites actually in scope. Columns are summary, spec, workbook, exhibits, then same / conflict / missing. A fact asserted in one file and absent in another is a flag if the missing file is where you would pay or contract it. Run the map per vendor. Never mix one vendor's exhibit with another's workbook.
If an exhibit is labelled "typical" or "previous project", park it. It is not this bid until the vendor says so in a clarification.
The usual ingestion failure is treating last year's PDF as this event. Vendors reuse a technical annex that still has last year's plant list, FCA term, and SLA. The new summary was rewritten for this RFP. If the model indexes the stale annex, you will flag a leftover file as a live contradiction, or miss the live workbook because the old spec "agrees" with a price sheet that was not submitted. Bind the check to the files listed on this submission receipt. If the vendor uploaded both, ask which revision governs before you score.
If RFP/RFQ auto-drafting produced your event, still bind the vendor check to the issued pack, including addenda. The packet has to be consistent against the revision they bid, not a draft they never saw.
Keep optional lots visible. A workbook that prices Plant A firm and Plants B and C as options is not a contradiction with a summary that mentions three plants, unless the summary states a firm three-site price. Mark "firm versus option" as a field, not as a fail.
Flag lead time, quantity, Incoterm, SLA, and site
Those five fields change what you think you bought.
Lead time. Summary "12 weeks to first delivery." Spec "14 weeks manufacture plus 2 weeks commissioning." Workbook "delivery week 16 from PO." Ask which clock starts (PO, kickoff, drawing approval) and whether commissioning sits inside first delivery.
Quantity. Summary "40 units." Spec BOM 40 plus 8 commissioning spares. Workbook 40 at Plant A, zero firm at B and C. Ask which quantity is priced, which is optional, and whether spares sit in the unit price.
Incoterm. Summary DAP to named plants. Workbook unit prices EXW, freight optional or omitted. A freight-excluded unit price is not a DAP price. Evaluators miss this when they score the story and never open the sheet. Ask the Incoterm, named place, and whether freight, insurance, and unloading sit in the unit price or in a named line.
SLA. Summary "4-hour on-site." Spec "remote first, on-site next business day for P1, excluding planned maintenance." Exhibit matrix "P1 remote 1 hour, on-site 2 business days." Ask which matrix is contractual, which severities exist, and which exclusions apply.
Site. Summary "all three plants." Spec and exhibits only survey Plant A. Workbook only Plant A firm. Ask which sites are in the firm price, which need a visit before a firm date, and whether access constraints change the lead time you just read.
Quantity, Incoterm, and site mismatches usually move landed cost. Feed the reconciled figures into total cost of ownership modeling. Do not TCO the summary's DAP story while the workbook is still EXW.
One vendor packet, read across four files
Picture a packaging-line upgrade bid covering Plants A, B, and C. You are not scoring yet. You open one vendor's packet.
The executive summary says 12-week delivery DAP to all three plants, a 4-hour on-site SLA, and 40 units in a single firm price.
The technical specification clocks 14 weeks of manufacture after drawing approval, plus 2 weeks of commissioning per site. The SLA chapter is remote-first, next-business-day on-site for P1, excluding planned maintenance. The BOM is 40 production units plus 8 commissioning spares. Only Plant A has a completed site survey; B and C are "to be confirmed after award."
The price workbook lists 40 units EXW, a blank or optional freight line, no named DAP place, and firm quantity only for Plant A. Plants B and C sit on an options tab with no lead time.
An exhibit labelled with last year's event code still shows two plants and FCA. Another exhibit is this year's Plant A survey.
You do not fail the bid. You also do not score "12 weeks, DAP, three sites, 4-hour SLA." You write one clarification that cites each passage: which clock starts lead time; whether commissioning is inside first delivery; whether the 40 units include spares; whether unit prices are EXW or DAP and who pays freight; which SLA matrix is contractual; which sites are firm; and which exhibit is this bid. Until those answers land, the packet is not ready for a score sheet, and it is not ready to feed a reverse auction on a unit price that still might exclude freight.
Send a cited question. Do not auto-fail.
Every flag needs two citations (file, section or cell, quoted phrase) and one question. "The summary and the workbook disagree" is not enough. The vendor cannot answer a vibe.
Route the question through the clarification round your procedure already uses. Same clock, same record, same panel visibility. Do not run a side channel with the salesperson. Do not drop a fail on the scoring sheet "pending clarification." Leave the criterion unscored or marked "awaiting response" until the answer is in. Using an internal mismatch to disqualify without a chance to clarify invites a challenge you will struggle to defend.
Do not fail a bid on a rounding difference. A workbook total a few cents off unit-price times quantity, 12 weeks versus 84 days, or 99.5% versus 99.50% is not a commercial fork. Log it as arithmetic and move on. If you treat rounding as a fail, you will throw out packets a human would have corrected in a minute.
The expensive miss is a freight-excluded unit price versus a DAP summary. That pair is not rounding. It is two different deals. If the panel scores DAP from the narrative, the EXW workbook will look cheaper than a rival who priced DAP honestly, or more expensive once freight is added later. Cite both, ask which Incoterm governs, and only then compare prices.
After the vendor answers, patch the packet (or log an accepted variance with the citations) and then score. If they confirm EXW, score EXW and put freight into TCO. If they confirm DAP, require the workbook to match. If they withdraw B and C, score the firm scope only. A flag that never becomes a written answer is still open. Do not let an evaluator interpret it into a score.
Is this worth automating for you?
Whether this pays back depends on how much time it takes your team today. Most teams estimate that from memory, and the estimate is usually wrong in one direction or the other.
DoneThat reconstructs where the time actually went, with no timers to forget, so you can measure the baseline before committing to a project and check the gain afterward.
Measure the baseline first