Estimators and project managers spend a staggering amount of their day looking for things. What's the spec section that governs this fixture? Is this finish called out anywhere on the drawings? Does the project manual require galvanized or stainless? Which addendum changed the door schedule?
On a typical commercial or federal project, the answers are in there somewhere, buried across a 600-page project manual, a few hundred drawing sheets, an addenda log, and a stack of RFIs. Finding any single answer takes minutes. Finding ten answers takes the morning.
Specset's natural-language Q&A turns that experience into a conversation. Ask your plans a question in plain English. Get back a cited answer that points to the exact sheet, section, and page.
But the part that matters (the part most "AI search" tools get wrong) happens before you ever type a question.
The problem with searching a raw project manual
Most construction document search tools index PDFs and call it done. You type a query, they run vector similarity across the text they extracted, and they hand you back the closest-matching paragraphs. It feels like search. It is not useful.
A few reasons why:
Specs are layered, not flat. A real project doesn't have one spec. It has standard specifications, special provisions, supplementals, addenda, and amendments, often from different authors, often contradicting each other, with an order-of-precedence clause buried in the front matter that tells you which one wins. A raw search returns all of them. The estimator still has to figure out which version is current and authoritative.
Specs are structured documents pretending to be prose. CSI MasterFormat divisions, UFGS sections, part 1/2/3 organization, submittal requirements, performance criteria: none of this survives naive text extraction. A flat search across a project manual is like searching a database by reading the raw bytes off the disk.
Drawings aren't text at all. OCR-based search misses everything that matters on a sheet: symbols, dimensions, sheet references, schedules embedded in tables. You can't ask "how many duplex receptacles on level 3" of a tool that only sees the title block.
If the data layer is wrong, no amount of model quality fixes the answer.
Specset conforms the project first
Before Q&A is available on a project, Specset runs a conforming pass on the entire document set.
For specifications, that means:
- Identifying every spec source in the package: standard specs, special provisions, supplementals, addenda, amendments
- Reconciling them against the order of precedence declared in the project, so the current authoritative version of every section is known
- Parsing CSI MasterFormat and UFGS structure (divisions, sections, parts, articles, paragraphs) into a queryable hierarchy
- Extracting structured fields like submittal requirements, product data, performance criteria, and reference standards into their own indexes
- Cross-referencing every spec section to the drawings, schedules, and details that implement it
For drawings, that means reading the vector geometry of every sheet, not running OCR on a rasterized image. Specset identifies sheet types, recognizes symbols, parses schedules and title blocks, and ties each drawing element back to the spec section that governs it.
The output is a conformed spec layer: a structured, deduplicated, hierarchical model of your project that knows what the current spec says, where it says it, and which drawing elements it controls.
Q&A runs on that layer.
What it feels like in practice
Once the conformed layer exists, asking a question is just typing.
A few examples of what estimators and PMs ask on a real project:
"What's the spec for the corridor floor finish, and is there an alternate?"
The answer comes back with the governing section (say, 09 65 13 Resilient Base and Accessories), the specified product family, the alternate listed in Section 01 23 00 Alternates, and direct links to the sheets where the finish schedule lives.
"Are there any sections in Division 26 that require UL listing?"
Specset returns the list of sections with UL requirements, the exact language pulled from each, and the page reference for every hit.
"What changed in the door schedule between the base bid and Addendum 2?"
You get a structured diff: door tags added, removed, or modified, with the before/after values and the page in the addendum that introduced the change.
"Is anywhere in the specs telling me to use galvanized vs. stainless for exterior hardware?"
You get the conflict, if there is one, with both spec sections cited and the order-of-precedence resolution noted.
Every answer carries citations. You don't have to trust the model. You can click straight through to the source.
Why this matters for estimating and preconstruction
For estimators, the bid-day question is always "what am I missing?" The answer to that question is almost never in any single place. It's in the gap between what the drawings show and what the spec actually requires, or in the addendum nobody had time to read, or in the special provision that overrides the standard spec for one particular wall assembly.
Plan Q&A is how estimators close those gaps without staying late.
For project managers and PEs in construction admin, the daily question is "does what the sub is proposing match what we specified?" Same answer pattern. The information is in the documents. The bottleneck is the time it takes to find it.
For owners and owner's reps, Q&A is how you do an independent check on the design package without staffing a six-person review team.
Q&A is the surface. Conformed specs are the foundation.
The reason Specset's Q&A works is the same reason Specset's takeoff is precise and our design quality reviews catch what manual review misses: we structure the data first. Drawing vectors, layered spec compilations, addenda reconciliation, CSI structure, submittal extraction: all of it lives in a conformed model of the project before any analysis runs.
Q&A is the most natural way to interact with that model. It's also the most underrated feature in our platform, partly because we built the underlying structuring work first, and partly because "ask your plans a question" sounds like marketing copy until you watch an estimator do it on a live bid set and ask why they don't have this on every project.
Try it on a live project
If you want to see what Q&A feels like on your own documents, the fastest path is to schedule a demo with one of your active bid sets or specifications. We'll conform the project, you'll ask the questions, and you'll see the citations.
You can also read more about the underlying modules: AI Takeoff & Estimating (which surfaces Q&A as part of the estimator workflow) and Design Quality Review, which uses the same conformed spec layer to flag conflicts and gaps before bid.
Co-founder and CEO of Specset. Over 15 years building software products and leading engineering teams.