It is genuinely useful for the document half of estimating: reading specifications, comparing addenda, and building compliance checklists. It does not take off quantities and it does not price anything.
Key takeaways
The interesting question is not whether Claude can estimate — it cannot, in the sense that matters, because it does not measure drawings or hold your prices. The interesting question is which part of an estimator's week is document work rather than measurement, and the answer for most tendered projects is a surprisingly large share of it.
Specifications, addenda, instructions to bidders, the form of tender, the subcontract you are being asked to sign, the schedule of prices, the site meeting minutes. Reading, comparing and reconciling those is real work with real risk in it, and it is what this class of tool is actually built for.
Anthropic's PDF support documentation describes the process in three steps: the system converts each page of the document into an image, the text from each page is extracted and provided alongside that image, and the model analyses both together. Because it works from the page image as well as the text, it can be asked about charts, tables and diagrams rather than text alone.
The published constraints are specific enough to plan around. Maximum request size is 32 MB, which varies by platform. Maximum pages per request is 600, dropping to 100 when the request's context window is under one million tokens. The format must be a standard PDF with no password or encryption. And the documentation warns that dense PDFs — many small-font pages, complex tables, heavy graphics — can fill the context window before reaching the page limit, with the suggested remedy being to split the document into sections.
For an estimator that translates into a working practice: feed it the specification divisions relevant to your scope rather than the whole project manual, and split a large set rather than fighting it.
Addendum 3 lands two days before close. Give it the addendum and your list of assumptions and ask which assumptions are now wrong. You still verify each one against the document, but you are verifying a short list instead of re-reading everything.
Put the tender's scope description alongside your quoted inclusions and ask what appears in one and not the other. This is tedious, error-prone, high-consequence work and it is precisely the shape of task a document model is good at.
Instructions to bidders typically bury half a dozen mandatory requirements — bid security, a signed addenda acknowledgement, a completed form, a supplement due at a specific hour. A model can extract them into a checklist. Bids get disqualified on this every year.
The general requirements govern how the whole job is administered. CCDC's own page for CCDC 2, the stipulated price contract notes that users should acquaint themselves with the CCDC Master Specification for Division 01 as a companion document. Getting a plain summary of the submittal, closeout and temporary-facilities obligations before you price general conditions is worth an hour of anyone's time.
Quantities. The PDF pipeline described above contains no calibration between page distance and building distance. That is why takeoff tools have an explicit scale step — Procore documents that the drawing scale is not configured by default and that setting it is imperative.
Prices. There is no connection to your catalogue or to current Canadian pricing. Statistics Canada's building construction price indexes move quarterly; a model is not reading them.
Code compliance. This is the one worth being firm about. The National Building Code of Canada 2020 is published by the National Research Council and sets out technical requirements for design and construction, but the National Research Council itself tells you to check with the municipal, provincial or territorial official to determine which edition applies in your jurisdiction. Ontario, for example, works from its own Building Code, published as the 2024 Building Code Compendium, and enforcement is the responsibility of municipal building officials. A summary from a chat window is not a code opinion and should never be treated as one.
The scale of what a new edition changes is also worth knowing before you ask a model to summarize it. The National Research Council’s own overview of the NBC 2020 notes that over 280 technical changes were incorporated in that edition alone, including introducing encapsulated mass timber construction for wood buildings up to 12 storeys and revised accessibility requirements. A model summarizing “what changed” between code editions is summarizing a document with hundreds of individually consequential amendments in it — a fine starting point for orientation, and not a substitute for checking the specific clause your assembly turns on.
The NRC lists the NBC 2020 itself at 792 pages in Volume 1 and 708 in Volume 2 — a fifteen-hundred-page document a model can search in seconds and no estimator will read cover to cover, which is precisely the kind of lookup this class of tool is built for.You are bidding the mechanical scope on an institutional renovation. Your quote carries 41 inclusions and 18 exclusions. The tender's scope description runs four pages, and there are three addenda.
The pass takes about twenty minutes. Load the scope description and your inclusion list; ask for items present in the tender but absent from your list. Load each addendum in turn; ask what changed against the base documents. Load the instructions to bidders; ask for every mandatory submission and its deadline. You now have three short lists to verify.
On a job like this the check typically surfaces something small and expensive — seismic restraint on a piece of equipment you priced without it, or a commissioning obligation that belongs to the mechanical contractor rather than the general contractor. Finding that on Thursday morning rather than in month four is the entire value.
Quantities come from a takeoff tool. Prices come from your catalogue. Documents — reading, comparing, extracting, drafting — are where a model like this earns its place. Canada's Voluntary Code of Conduct on advanced generative AI systems and the Privacy Commissioner's principles for generative AI both point the same way on the rest: know what you are sending, limit sensitive and confidential material, and keep a person accountable for the output.
Yes, as an image with text extracted alongside. It can find where something is referenced or explain what a note says. It is not measuring.
Split it. The documentation recommends exactly that, and in practice you only need the divisions in your scope plus Division 01.
The document-handling behaviour is well documented, which is the reason it is described here in specifics. Test any tool the same way: on a project manual you know well, checking every extracted requirement against the source.
A 30-minute call is enough to tell you whether AI pays for itself here.
