A reviewed AI pilot for oilfield RFQ document intake

Topics: oil and gas procurement automation; RFQ document extraction; quotation intake automation; oilfield bid comparison workflow.

A buyer spends the morning copying prices and delivery terms from quotations into a comparison sheet. One PDF is scanned, one includes a revised schedule and another puts freight in an email. The slow part is not typing alone. It is deciding which statement belongs to which offer.

An RFQ document intake pilot should prepare a reviewable record of supplier statements. It should not recommend a technically acceptable spare or select a winning bid. That narrower scope makes both the benefit and the errors easier to measure.

Choose the fields before choosing the model

Start with one repeated quotation family, such as spare assemblies bought against controlled line-item RFQs. Define the initial fields: RFQ reference, supplier identity, quotation number and revision, offered item reference, quantity, unit, unit price, currency, delivery statement, validity statement and document references. Treat technical deviations as text to surface for engineering review, not as classifications the model is authorised to resolve.

For every field, specify what counts as an acceptable answer. A delivery field may need the exact phrase and its trigger, rather than a guessed calendar date. A missing currency should remain missing even if the supplier usually quotes in rupees. If freight appears in a separate attachment, retain that source relationship.

JIP33's procurement-document framework can inform the field dictionary where the purchaser has adopted it. The pilot's extraction schema should follow the actual RFQ, not an assumed sector template. JIP33 FAQ

Show the buyer what changed

Illustrative example. Quotation Q-73 revision 1 offers a unit at ₹85,000 with delivery eight weeks after order acknowledgement. Revision 2 keeps the price but changes delivery to twelve weeks after drawing approval.

A model that returns only “12 weeks” loses the event that starts the clock. If it labels revision 1's delivery condition as part of the current offer, it carries an obsolete term into the comparison. Show both passages and have the buyer confirm the duration and trigger together.

The reviewer packet needs the original file, source anchor, proposed value, missing or conflicting evidence, extraction version and reviewer outcome. The corrected value must not overwrite the machine output; retaining both makes error analysis possible.

Establish a baseline the team can reproduce

Use synthetic or authorised redacted documents first. A proposed evaluation might contain 60 quotation packages split into configuration and held-out groups, with supplier-layout families kept together so near-identical revisions do not appear on both sides. These numbers are illustrative. Increase the sample if it fails to represent the formats and exceptions the buyer actually handles.

Have procurement label the correct field values and a second reviewer resolve disputed labels. Include unreadable scans, missing pages, multiple currencies, duplicate attachments and late revisions. Track which cases cannot be answered from the supplied evidence; guessing is an error, not extra coverage.

Measure the current process's active handling time, correction count and unresolved cases. Then compare structured supplier forms, simple text/OCR extraction with rules and the AI-assisted workflow on the same evaluation task. If a structured return sheet solves most of the problem, that is a legitimate outcome.

Microsoft's Document Intelligence guidance discusses field confidence and human review. A model's confidence value still needs evaluation on the organisation's own document mix; it is not a substitute for measured correctness. Microsoft confidence guidance

Make acceptance field-specific

Report correct proposed values divided by all proposed values, and correctly populated required fields divided by all answerable required fields. The first measures precision; the second exposes excessive blanks. Report unsupported guesses on genuinely absent fields separately.

For example, suppose a held-out set contains 120 answerable fields. The tool proposes 100 values and gets 90 right: precision is 90/100 = 90%, while correct coverage is 90/120 = 75%. Ten wrong proposals and twenty missing answers still need attention. These are illustrative results, not pilot targets. Score each complete delivery statement as one answer so a correct duration cannot conceal a wrong trigger.

Also measure quotation-revision selection, accurate source anchors, reviewer corrections and total review time. A good average must not hide a wrong currency or an obsolete delivery condition. Procurement should agree critical fields and error tolerances before the held-out run. Any proposed numerical target is a pilot decision, not an industry benchmark.

Check line association as well as the field value. A correctly read ₹85,000 assigned to the wrong RFQ line is an error. Keep the supplier's line identifier, the proposed RFQ line and the buyer's confirmation; repeated descriptions and multi-page tables make this a useful challenge case. Ambiguous associations should stay unassigned until reviewed.

During this first pilot, every packet receives human review. There are no automatic award, supplier-qualification or ERP-posting decisions. A useful output is a reviewed comparison draft, not a purchase order.

Test the intake boundary and the recovery path

The staging service can read an approved document folder and authorised ERP RFQ references. It writes only to the pilot review store. Restrict the corpus by RFQ and buyer role, set retention with the data owner and log access without copying full quotations into general diagnostic logs.

A quotation is untrusted source content. Instructions embedded in it must not become commands to the assistant or connected systems; OWASP describes this indirect prompt-injection risk. OWASP prompt injection

Test a duplicate email, an identical file with a new name, a revised file using the old name, a missing page and a document the user cannot access. Compare file identity and quotation identity separately. A revised bid is not merely a duplicate because the quotation number is unchanged.

Stop the batch if access leaks, source anchors disappear or any production write occurs. Quarantine affected outputs, preserve the originals and resume the approved manual intake process. After fixing the cause, rerun the affected packages in staging and have the reviewer confirm that corrections did not alter previously accepted fields.

Summary and next actions

The pilot should establish whether preparing source-linked fields actually reduces the buyer's work. Better extraction helps only if revision checks, line matching and corrections leave the reviewed comparison at least as dependable as the current process.

1. Select one quotation family and agree the field definitions with its buyer.
2. Create the labelled baseline and separate evaluation set, including missing and conflicting evidence.
3. Compare forms, rules and AI using field correctness, revision handling and reviewer effort.
4. Demonstrate access controls and recovery before considering a wider integration.

A pilot discussion can start with the RFQ family, current document repository, ERP reference format and review owner. Synthetic quotations are sufficient for the first design session.