What AI Takeoff
Can't Do Yet

It does not decide scope

This is the limit that matters most in practice, and the one vendors talk about least. AI takeoff counts what is drawn on the plan. It does not interpret what is in or out of your scope of work. When you open a commercial electrical bid, the decision about whether temporary power is included, whether conduit in the slab is yours or the GC's, or which alternates you are pricing — none of that is on the drawing. It lives in the contract, the invitation to bid, and in your team's understanding of how this GC writes scopes.

Inclusions, exclusions, and alternates are estimator judgment calls that come from years of reading project documents and understanding construction intent. Industry commentary from Rocket Takeoffs (2026) puts it plainly: AI lacks the construction intent and scope interpretation that experienced estimators carry. The tool gives you quantities; the estimator decides what quantities matter for this bid.

• AI counts what is drawn — it does not interpret what is in or out of scope
• Inclusions, exclusions, and alternates require an estimator judgment call
• Scope interpretation requires construction intent AI does not yet have

It does not do engineering sizing

AI takeoff measures geometry. Pipe sizing, duct sizing, and conductor sizing are engineering problems governed by loads, code, and system performance — not by what the plan shows. When the mechanical engineer draws a 4-inch branch, the AI can count and measure that branch. It cannot tell you whether that size is correct, whether it matches the load on the schedule, or whether you need to flag a potential design issue before pricing it.

Fire protection is the clearest example. Sprinkler branch lines cannot be pre-sized from the architectural plan geometry alone. NFPA 13 requires hydraulic calculations: density, area of application, available water supply, friction loss. The AI reports what the engineer drew; it does not re-engineer the system. For a fire protection sub, this means the AI saves time on fixture counts and pipe footage, but the hydraulic review and sizing verification still sit with your estimator or designer.

The same principle applies in electrical: wire sizing depends on ampacity tables, derating, voltage drop — not on conduit route length alone. AI gives you the route length; your estimator applies the engineering.

• Pipe, duct, and conductor sizing depend on loads and code, not plan geometry
• Sprinkler branch lines are governed by NFPA 13 hydraulics, not drawn dimensions
• AI reports what the engineer drew; it does not re-engineer the system

It cannot see field reality

Every set of construction drawings is an idealized version of a project. The plan assumes clean access, average soil, and no surprises in the existing structure. Field reality is almost always more complicated, and those complications cost money that does not appear on any drawing.

Site access affects labor productivity on every trade, but it is invisible on a plan. Soil conditions govern earthwork cost more than any quantity. Congestion from other trades — especially in mechanical rooms and ceiling spaces — is something experienced foremen account for instinctively but no AI can read from a PDF. On renovation and tenant improvement work, existing conditions routinely diverge from as-built drawings, and the cost of those surprises falls on whoever didn't account for them in the bid.

Earthwork is particularly exposed: preliminary quantities from plan geometry are typically accurate to only +/- 10–15%, according to GX Contractor (2025), because subsurface conditions, compaction requirements, and haul logistics are off-plan variables. Demolition scope on renovation jobs has the same problem — hidden conditions and hazardous materials rarely show fully on drawings. AI gives you a starting number; field judgment sizes the risk allowance.

• Site access, soil, congestion, and existing conditions are all off-plan
• Earthwork quantities from plan geometry are preliminary at +/-10–15% accuracy
• Demolition and renovation surprises rarely show fully on drawings

It struggles with ambiguity

AI takeoff systems are trained on standard symbol sets and clean, well-produced commercial drawings. When drawings deviate from that baseline, auto-count confidence falls and review time rises. The two most common sources of deviation are custom or non-standard symbol sets and poor scan quality.

Some engineering firms and owner's reps use proprietary legends, custom device symbols, or drawing conventions that differ from the standard library the AI was trained on. When the tool can't reliably match a symbol, it flags it for human review — which is the right behavior, but it means more time on your side, not less. Conflicting notes, late addenda, and open RFIs compound the problem: the plan might show one thing, the addendum corrects it, and a pending RFI clouds the picture further. Reconciling those conflicts is a human task, and AI does not read addenda or RFIs automatically.

Low-resolution scans and hand-drawn or hand-sketched drawings reduce measurement reliability significantly, as noted by appintent (2026). A drawing that was scanned at 150 dpi, rotated slightly, and run through a fax machine twice is a hard input. The AI will attempt it, but the confidence intervals widen and your review workload increases accordingly. The quality of your output scales with the quality of your input drawings.

• Non-standard or custom symbol sets lower auto-count confidence
• Conflicting notes, RFIs, and addenda require human reconciliation
• Low-resolution scans and hand sketches reduce measurement reliability

What this means for your workflow

None of the above limits are reasons to avoid AI takeoff. They are reasons to deploy it correctly. The tool delivers the most value on the tasks it handles cleanly: repetitive counting of standard components, to-scale linear and area measurement across large plan sets, and first-pass quantity generation that your estimator then verifies against schedules and specifications. That work can represent 60–70% of the hours in a manual takeoff, and reclaiming it matters.

The decisions that remain human are the ones that require construction knowledge, contract literacy, and risk judgment. Scope, exclusions, site risk allowances, and markup are not tasks you want an algorithm deciding anyway — they are where your estimating team earns its margin. Industry consensus from ConstructConnect (2026) frames it well: AI redefines the estimator role rather than replacing it. The estimator moves from counting to reviewing, from measuring to deciding.

A practical division: let AI run the takeoff on clean commercial sets, then spend your estimator time on schedule verification, scope letter drafting, risk-adjustment, and value engineering. That is where experience compounds. AI handles the volume; you handle the judgment.

• Use AI for repetitive counting and to-scale measurement; verify against schedules
• Keep scope, exclusions, risk allowances, and markup as human decisions
• Industry consensus: AI redefines the estimator role rather than replacing it (ConstructConnect, 2026)

Questions estimators actually ask