AI electrical takeoff software, and where the AI belongs

General AI models read drawings well and count symbols badly. In AECV-Bench, published in January 2026, the best multimodal models scored up to 0.95 on text extraction from architectural drawings but often 0.40 to 0.55 on counting doors and windows.¹ AI takeoff tools that work pair a model with purpose-built detection or geometry, and keep an estimator reviewing the result. AmpCount uses AI to read the text on a sheet and counts with geometry.

By Agustín Policano, founder of AmpCount, which makes one of the tools discussed here. Research figures link to the papers; vendor claims link to the vendor's own page, checked November 1, 2026.

What the research measures

0.95

top accuracy of multimodal models on text extraction (OCR) questions about drawings in AECV-Bench.¹

0.51

mean exact-match counting accuracy of the best model tested (Gemini 3 Pro) across doors, windows, bedrooms and toilets on 120 floor plans.¹

82.5%

mean average precision of a detector trained only on electrical service keys (YOLOv8), on 45 scanned electrical layout plans.³

AECV-Bench counting results, exact match (the share of answers with the count exactly right) and mean absolute percentage error¹
Model	Mean	Doors	Windows	Counting error¹
Gemini 3 Pro	0.51	0.39	0.34	16.0%
GPT-5.2	0.49	0.28	0.27	19.4%
Claude Opus 4.5	0.42	0.16	0.16	24.9%

“…the bottleneck lies not in reading text but in understanding the graphical language of AEC plans.”

AECV-Bench, Kondratenko et al.¹

“…current VLMs are not yet suitable for autonomous deployment in blueprint analysis and are best integrated into human-in-the-loop workflows.”

BlueprintSymVL, Shteriyanov et al.²

Two cautions when reading these numbers. AECV-Bench tests general chat models on architectural symbols, not electrical devices, and BlueprintSymVL uses industrial piping symbols; neither is a test of any takeoff product. And the trained detector in SkeySpot shows that models built for one symbol set do much better than general ones, but still miss and misread some symbols. Either way, the count needs a way to be checked.

What AI takeoff vendors claim

Vendor	Claim, quoted	How it's measured
Togal.AI	“98% accuracy on floor plans” and “up to 98% accuracy”⁴	Not stated on the page
Beam AI	“99%+ accurate, QA-checked takeoffs and estimates”⁵	Delivered takeoffs are QA-checked before delivery, so the figure covers AI plus human review
Bobyard	Teams “cut takeoff time by 50-90%, depending on drawing quality, trade, and the estimator review process”⁶	A time range, not an accuracy figure
Drawer AI	“Cut Your Electrical Takeoff Time by up to 90%”⁷	Customer case study (Intermountain Electric)
Kreo	Auto Measure “lands within 94% of the real geometry” on busy plans⁸	Area and outline geometry, not counts
On-Screen Takeoff	“It is your responsibility to verify results to ensure accuracy.”⁹	No accuracy figure published

None of these pages says which plan sets the figures come from or how a miss is scored. That doesn't make them wrong, but it means the only accuracy number that matters is the one you get on a bid you've already counted. Bobyard puts it plainly: the estimator “reviews each detection, using the confidence score on it, before it is locked into the estimate.”⁶ On-Screen Takeoff skips the number altogether and leaves verification to you⁹; the On-Screen Takeoff comparison looks at what that means on an electrical set. For the same claims across every trade, including Kreo and Beam AI on areas and lengths, see the broader AI takeoff software guide.

Where AI helps

Reading sheet numbers and titles from title blocks.
Reading legend descriptions and fixture schedule rows.
Telling plan sheets from details, risers and notes.
Giving a second look at marks the counting isn't sure about.
Answering questions about specs and notes.

Where it needs a check

Being the final count, when a number can't be traced to marks on the sheet.
Dense areas, where models miss instances or invent them.
Symbols that look alike: a GFCI and a duplex, two fixture types with one letter apart.
Accuracy figures measured on someone else's plans.

How AmpCount splits the work

Geometry counts. Every symbol in the legend gets a signature built from its strokes, and every drawing on every sheet is compared against it, in any rotation and mirrored. Every number comes from marks you can see on the plan.
AI reads text. A vision model reads each sheet's number and title, the names in the legend and the rows of the fixture schedule.
Uncertain marks wait for you. A drawing that is close but not exact stays out of the count. AI can add a second opinion on whether it looks like the legend symbol, but that's a hint: it never confirms a mark and never counts.
AI can be turned off. Without it, AmpCount reads sheets with its own rules and counts exactly the same way.
Your plans never train a model, and they're deleted when you cancel.

More on the method in how it works, and which plan sets it reads best.

Six questions for any AI takeoff vendor

How was your accuracy figure measured, on whose plans, and how is a miss counted?
Can I click any count and see every mark behind it on the sheet?
What happens to the marks the model isn't sure about? Do they go into the total?
Can I run a bid I've already counted and compare totals per type before I pay for a year?
What does it do with scanned sheets: count them, or tell me it can't?
Are my plans used to train your models?

Where contractors are with AI

Contractors expect a lot and use little so far. In the 2026 AGC and NCCER workforce survey, estimating and bidding was the area where firms expect AI to have the greatest impact, cited by 67 percent¹⁰. In AGC's outlook survey from late 2025, 23 percent of firms said they use AI for estimating¹¹. Dodge found that “87% expect transformation, yet only 19% have adapted workflows”¹², and in RICS's worldwide survey, 45% of organisations reported no AI use at all¹³. Whatever a firm decides about AI, the estimate still runs from quantities to labor to price, the sequence laid out in how electrical estimating works.

If you're comparing tools, the electrical estimating software guide covers prices for 14 of them, and the Togal.AI and Drawer AI comparisons go deeper on the two AI-first ones. If your team counts with a symbol search today, the Bluebeam VisualSearch comparison and the Countfire Smart Select comparison show how a search from one symbol differs from a legend-based count. Every head-to-head is listed on the comparisons hub.

Keep reading

Sources

Product names are trademarks of their owners. AmpCount is not affiliated with any vendor on this page.