Tender Clause Risk Classification

A classification run is useful when each tagged clause carries a risk category, a cite back to the contract text, and a confidence flag, and a lawyer still reads the high-risk items. The tag is a queue, not a finding. It is not legal advice, and it does not approve a bid.

Refuse to ship a category with no quote, a paraphrase that cannot be found in the pack, or a low-confidence guess coded as if it were certain. If the model is unsure, the clause stays uncoded and goes to legal as not classified. Inventing a clause to fill a category is a process failure, not a helpful completion.

The closed list is the point. Typical bid-stage buckets include liquidated damages, unlimited liability, unusual scope, indemnities that sit outside your standard, payment and retention terms that shift cash risk, and a residual bucket for items that need counsel. Freeze that list before the first tender. Do not let the model invent a new label mid-pack.

This sits next to, not instead of, adversarial bid review. Classification tells you which clauses to read first. Adversarial review stress-tests the commercial story those clauses create.

Classification on a long PDF as one blob hides liquidated damages in a schedule and buries unlimited liability in a definitions annex. Extract first.

Pull the invitation to tender, the form of contract, special conditions, preliminaries, pricing documents, and any employer amendments. Construction bid packs often live in document control and project systems from vendors such as Autodesk and Procore. Treat those systems as the source of the files, not as a substitute for extraction. Export or download the contract set so the classifier sees the same text legal will cite.

Segment at clause and sub-clause level, keeping the numbering. A cite that names a clause number without the quoted sentence is not a cite. Keep the surrounding heading so delay damages in a preliminaries item is not merged with delay in a programme narrative.

Mark non-contract text: cover letters, question-and-answer logs, drawing notes, and marketing summaries. If those fragments enter the classifier, you get false unusual-scope tags on ordinary specification language.

One illustrative pass: a bid lead on a school refurbishment invitation to tender splits the pack into the unamended form of contract, the employer's special conditions, and the preliminaries. Unlimited liability language, when it appears at all, usually sits in the special conditions or in an employer amendment, not in the unamended form. If you classify the form of contract alone, that language never enters the queue. Extraction is the quality gate. Classification cannot recover a clause that never made it into the input.

For each extracted clause, the model returns one label from the frozen list, the verbatim span it used, and a confidence band. High and medium confidence may receive a category. Low confidence stays uncoded.

For liquidated damages, the cite must point at the damages or delay-damages provision, including any daily or weekly rate and any cap, or the clear absence of a cap. If the model tags time-is-of-the-essence language as liquidated damages without a damages mechanism, send it back as uncoded.

For unlimited liability, the cite must show the carve-out or the missing cap, not a general indemnity heading. Many construction contracts cap some heads and leave others open. Tag the open head, quote the sentence that leaves it open, and leave capped heads in a different category if your list has one.

Unusual scope is for obligations that sit outside the trade, geography, or procurement model you bid, not for any clause a junior estimator has not seen before. Standard insurance, bonding, and statutory cooperation language is usual for construction tenders. Tagging a normal public-liability or contractors-all-risks requirement as unusual scope burns legal time and trains the team to ignore the tag.

The cite is a quote plus locator: clause number, document name, and page or PDF page if numbering is unreliable. Legal must be able to jump to the same sentence. If they cannot, the tag is void.

Do not auto-map a tag into a commercial position. A liquidated damages tag does not mean you should price the damages or qualify the bid. That decision stays with legal and the bid lead after review.

Where the employer has issued a marked-up form, pair this pass with contract redline deviation analysis so you classify the version you would actually sign, not the unamended template.

Route the queue in this order: unlimited liability and other uncapped heads, liquidated damages and related delay remedies, unusual scope, then the remaining coded categories, then the uncoded set.

Legal's job on each high-risk tag is to confirm the quote, confirm the category, and write the actual advice: qualify, price, decline, or accept with a named residual. The classifier does not close that step.

Uncoded clauses are not cleared. They are unread. Budget time for them after the high-risk queue, or you have only reviewed what the model was confident about.

Bid leads should not wait for a full legal opinion on every preliminaries paragraph before starting takeoff. They should wait for legal on unlimited liability and liquidated damages before they treat the tender as bidable. That is the time saving. It is not a skip.

If win-or-lose scoring is in use, do not feed raw tags into win probability scoring as if a high-risk count were a probability. A single unlimited-liability cite can dominate a clean commercial score. Keep the legal outcome (qualified, declined, or accepted) as the input, not the tag count.

Missing liquidated damages. Those provisions often live in a contract particular, an appendix, or a preliminaries item named damages for delay, not under a heading that says liquidated damages. If extraction dropped schedules, or the model required the words liquidated damages as a keyword, the most expensive time risk in the pack never enters the high-risk queue. Guardrail: a bid is not classified-complete until the damages-for-delay location is either tagged or explicitly logged as absent after a targeted search of particulars and appendices.

Tagging standard insurance as unusual. Public liability minima, professional indemnity for design portions, and joint-names requirements are ordinary in construction invitations to tender. An unusual-scope tag on those clauses floods the legal queue. Counsel then learns that unusual means the model was surprised. Guardrail: maintain an allow-list of insurance and bonding patterns that cannot receive unusual-scope without a second, human confirmation.

Auto-accepting a bid because no high-risk tag fired. Silence is not a legal review. A run can return no high-risk tags because the special conditions file was omitted, because delay damages were phrased in a schedule the parser skipped, or because every difficult clause was low-confidence and therefore uncoded. Guardrail: the bid lead cannot mark legal complete from an empty high-risk list. Completion requires legal sign-off on the high-risk queue and a disposition of the uncoded set.

After legal has confirmed or rejected tags, the surviving high-risk items belong on the project risk conversation, not only in a review email. Feed confirmed categories, cites, and the legal disposition into project risk register generation so delay damages, uncapped liability, and unusual obligations remain visible if you win.

Keep the audit trail: source file, clause locator, quote, model category, confidence, legal category if different, and decision. That trail is what you reuse on the next tender from the same employer. It is also what you need if someone later asks why a qualification was or was not taken.

Classification does not replace reading the contract. It orders the reading so liquidated damages, unlimited liability, and unusual scope are not found on the afternoon the bid is due.