Decide bid or no-bid before anyone opens a takeoff, by running every package through the same short list of disqualifiers. AI is genuinely good at that reading job — pulling the mandatory requirements, the contract form, the closing date and the scope boundaries out of a few hundred pages — and the estimator still decides.
Key takeaways
Most estimating capacity in a small Canadian contractor is not lost on the jobs it loses. It is lost on the jobs it should never have started — the tender that required a performance bond larger than the company’s surety facility, the one that closed at 2pm on the day the takeoff was half done, the one where the payment terms run so far out that the shop cannot carry the receivable.
Screening is a reading problem, and reading is the thing current AI tools do well. Feed the tender documents into a model and ask it to return a fixed set of fields — closing date and time, bonding and insurance requirements, the contract form named, the scope sections that apply to your trade, the schedule dates, the payment terms, the site location. Then a human reads the flags. That division of labour is the whole method.
A bid/no-bid decision has two stages, and only the second one is judgment. The first is eligibility, which is binary. If you cannot meet a mandatory requirement, nothing else about the job matters.
Bonding is the usual one. The Canadian Construction Documents Committee publishes the standard bond forms — CCDC 220, 221 and 222 — along with CCDC 23, A Guide to Calling Bids and Awarding Contracts, which is what most institutional owners follow when they structure a call for tenders. If a package names a CCDC bond form, that is a surety question you answer before you estimate, not after.
Workers’ compensation status is the second. In Ontario a clearance is, in the WSIB’s own words, “a unique number issued by the WSIB to registered businesses, which shows that a business, contractor or subcontractor is registered and up-to-date with us”, and it is valid for up to 90 days. If your clearance has lapsed, or a sub you intend to carry cannot produce one, that is a problem to solve during the bid period, not during mobilisation.
Trade certification is the third. Skilled Trades Ontario lists 23 compulsory trades, and to work in one a person must hold a Certificate of Qualification, a Provisional Certificate of Qualification or a Registered Training Agreement, with journeypersons renewing annually. For electrical work there is a further gate: the Electrical Safety Authority states that only a Licensed Electrical Contractor can be hired to do electrical work on a home in Ontario. A screening prompt that checks whether the scope contains work you are not licensed to self-perform is worth more than any takeoff feature.
What the model extracts, and what a person decides
Extracted: closing date and time, submission format, bonding and insurance limits, the named contract form, prequalification requirements, the trade sections that apply to you, milestone and completion dates, liquidated damages language, payment terms, site address, mandatory site meeting dates.
Decided by a person: whether the crew is free, whether you want this client, whether the schedule is achievable, whether the price will hold, and whether you are prepared to carry the receivable.
Never accepted on the model’s word alone: any pass/fail requirement. Ask for the page or section reference with every extracted field, then open that page.
Before pricing, look at which standard document the package is built on. The CCDC catalogue distinguishes CCDC 2 (Stipulated Price Contract), CCDC 3 (Cost Plus Contract), CCDC 4 (Unit Price Contract), CCDC 18 (Civil Works Contract) and CCDC 14 (Design-Build Stipulated Price Contract). Those are different risk allocations, and a lump sum against an incomplete design is a different business decision from a unit price contract against measured quantities.
This is a good use for a model with a long context window: ask it to summarise the supplementary conditions against the standard form and list every clause that modifies the standard allocation of risk. It will miss things, so a person still reads the supplementaries. But knowing in ten minutes that the owner has rewritten the changes and delay provisions changes whether you spend the week on it. CCDC also publishes CCDC 16, a guide to changes in the contract, which is worth reading once so you recognise what “normal” looks like.
A profitable job you cannot finance is a bad job. Ontario’s prompt payment regime sets deadlines for paying down the contractual chain once a proper invoice is delivered, and our sister firm sets out the prompt payment rules and timelines under the Construction Act. The same regime gives you construction adjudication as a fast-track route for payment disputes rather than a lawsuit. Read a tender’s payment terms alongside those defaults: where a package tries to push payment further out than the statutory scheme contemplates, that is a flag your screening prompt should catch and your lawyer should read.
A 12-person mechanical contractor in Ottawa receives roughly nine invitations a week — some from the kind of municipal portal the City of Toronto runs as TO Bids, some from generals it has worked for, some forwarded by a supplier. Historically the estimator skimmed them all on Friday afternoon and picked three on instinct.
The new routine: every package that arrives is dropped in a shared folder. An extraction step returns a one-page summary per tender with the fields above and a page reference for each. Three rules run automatically. Bonding above the company’s surety facility — flag as no-bid. Site more than 150 kilometres from the shop with no accommodation allowance — flag for the owner. Scope containing work requiring a licence the company does not hold — flag as sub-required, with a note to price the sub before committing.
The estimator now starts Monday with nine summaries instead of nine packages, and spends the week estimating two or three jobs properly instead of four badly. Nothing about the estimate itself was automated. The filter was.
Two failure modes are worth naming. The first is a model that confidently reports a requirement that is not in the document, or misses one that is. The mitigation is structural: require a citation for every extracted field, and treat any field without one as unread. The second is drift — a screening rule written for institutional work quietly rejecting the private jobs that pay your overhead. Review the no-bid list monthly and look at what you turned away.
No, and you should not want it to. The decision depends on crew availability, cash position, appetite for the client and what else is in the pipeline — none of which is in the tender documents. What AI removes is the reading that stands between you and the decision.
Partly. Text extraction from a specification is reliable enough to work with. Drawing interpretation is much weaker, and a screening pass should not depend on it. Use the model for the written package, and have a person scan the drawings for the two or three things that decide fit — ceiling heights, existing services, site access.
Addenda are where screening breaks. A requirement can appear in addendum three that was not in the original package. Whatever you build, build it so that each addendum is re-run through the same extraction and the summary is versioned, not overwritten.
The same triage applies with different fields. For a homeowner job the pass/fail items are permit requirements, the code edition in force — Ontario’s Building Code is enforced by municipal building officials and the province publishes the current Compendium — access, and whether the client’s budget and the drawings are in the same universe. The screening habit matters more than the sophistication of the tool.
A 30-minute call is enough to tell you whether AI pays for itself here.
