Benchmarked: AI Quantity Takeoff Accuracy on 50 Civil Drawings (2026 Data)
1. Introduction / At a Glance
On first glance it can be hard to understand why we need specialist construction AI. For some tasks like reviewing contracts, searching for information in building codes it seems like it would be hard to beat leading AI models from the likes of OpenAI. That’s why I wanted to put one of the world’s leading Construction AI startups through their paces on one of the most core areas for Civil Engineering and the wider Construction Industry - understanding Construction Drawings.
Civils.ai is a tech startup helping civil construction contractors and consultants with estimation and discrepancy checking. Their software is focused around two main features:
Quantity takeoffs from PDF drawings (with AI models built specifically for civil projects including concrete, groundworks, utilities and steelwork).
Discrepancy checking of drawings and reports to flag up commercial risks and non-compliances.
Based out of Singapore and San Francisco the startup works with titans of the industry including AECOM, Jacobs, Bachy Soletanche and Arup.
At 15 staff today and with $1M dollars in pre-seed funding collected so far they are the next generation of Construction Tech startup’s leading the AI frontier in the built environment
2. What It Claims to Do
Civils.ai firmly places itself as the tool of choice for Civil Contractors going as far as comparing itself to other competitor software companies (also focused on the quantity takeoff space) and demonstrating that for civil construction they are the most all-encompassing solution for civil quantity takeoffs and have a built in discrepancy checking layer.
Teams which would have the strongest appeal for a software like Civils.ai would include:
Civil Contractors
Groundworks Contractors
Geotechnical Engineers
Estimators
The main value add claimed by Civils.ai is that PDF drawings can be ‘dragged and dropped’ to the software, measurements categories are then selected from pre-existing options, an ‘additional instructions’ text box can be used to provide any further comments to the system and all measurements are handled for you with the first pass results being editable. The main value proposition of the software comes from the hours of time saved over manually measuring and checking drawings, which can be an error prone process which can take weeks of a professional's time per project. The discrepancy checking layer catches errors which would otherwise result in rework later in the project lifecycle.
Civils.ai has comprehensive information on their website including case studies, examples of their software being used in various scenarios and serves as a tutorial and introduction to their software.
3. Pricing & Plans
Civils.ai has three pricing tiers available at the time of writing this article:
Entry level pricing of $90/month which is best suited for those interested by the material shown on their website and includes 10 drawing takeoffs.
For those interested in more deeply testing Civils.ai across more drawings there is also a $270 option which includes a greater number of takeoffs with 30 drawing takeoffs per month.
They also offer deals for teams and enterprises beyond this with their ‘enterprise’ option with better value costs per takeoff for teams which need more than the self-serve options listed on their website.
The self service option is easy to use and accepts most major credit cards with the transaction granting instant access to the software.
Interestingly Civils.ai include unlimited AI checking with each pricing option and only cap the number of quantity takeoffs and cloud storage limit.
4. Setup & Onboarding
After the transaction is successful you are granted access to the software and automatically logged in, from there I am prompted to create my first project.
After creating my first project I am prompted to upload my first set of drawings and select which trades (e.g. concrete, electrical, groundworks, utilities) I am interested in and to provide any additional instructions to guide the AI in producing the measurements I need.
From there I can confirm which date/time I need the measurements, as Civils.ai runs a semi-automated QA/QC checking process on all measurements before the results are released.
For my takeoff it took around 1 hour and 30 minutes to process, which was 50 drawing sheets of civils scope takeoff (utilities, external paving and concrete foundations). I’d requested it for the next day and I realised that results can be returned far sooner than expected.
5. Interface & Usability
The interface made it straightforward for me to set up my first takeoff. After clicking submit, I saw that the takeoff became a pending ‘task’ in the outputs table, with an estimated time to completion.
It was quite straightforward to create new projects and set up the agents to work simultaneously on tasks and once the task was completed I received an email telling me that the results were ready for viewing meaning I was free to get on with other things in the meantime.
Civils.ai was able to understand both scanned and vector drawings of varying quality and the results of the quantity takeoff were not noticeably different, showing impressive consistency across the different types of drawing typical consultants and contractors need to deal with.
One feature which I discovered after running the quantity takeoff is I was able to share the results with others not signed up to Civils.ai using their ‘share’ feature.
I was also able to edit the measurements of the takeoff by clicking the ‘edit results’ button and this allowed me to delete measurements, add new measurements, reposition measurements and add new categories of measurements before saving the changes into a new version of the takeoff so I was able to switch back to my original takeoff if needed.
6. Civils.ai Core Features Review
In the following section I’ve broken my review down into two detailed assessments on the two features I tested in Civils.ai:
Automating the quantity takeoff process on a typical Civils project
Checking discrepancies and issues across that set of drawings and supporting contracts and specs.
Both of these features I first tested on Civils.ai on a project with utilities, external paving and concrete foundations before comparing the results found against my own verified numbers and then comparing against the results found by Claude and another AI for quantity takeoffs.
Quantity takeoff for Civil Construction:
Whilst setting up the quantity takeoff Civils.ai prompted me to select which takeoff categories I needed organised by ‘trade’. The options I had to choose from were which tallied with the options featured by then on their breakdown of the AI for Civil quantity takeoff features:
Concrete
Masonry
Fire Suppression
Plumbing
HVAC
Electrical Communications (data, telecom)
Electronic Safety & Security (CCTV, alarms)
Groundworks
Exterior Improvements (roads, landscaping)
Utilities (water, sewer, drainage)
I noted that the quantity takeoff quantities were heavily focused on Civils projects
After selecting Utilities, Concrete, Groundworks and Exterior Improvements I ran the takeoff and after processing completed opened my results to find that each measurement made was also linked to an annotated layer on the PDF allowing me to easily toggle the visibility of each measurement and check I was happy with how it had been measured by the AI. This made the checking process incredibly easy.
I was also able to toggle between viewing just the annotated drawings and the entire drawing set.
When it came to working with the measurements I could use the ‘view full table & export’ button, where I found the measurements were organised and broken down into different tabs, with the summary providing the measurements across the entire project, whereas breakdown grouped the measurements by drawing sheet and finally the measurements tab showed each unique length the AI had measured.
I was then able to download the extracted measurements into Excel alongside the annotated PDF for combining with my cost information and turning into a bill of quantities.
Another interesting feature I discovered was the ‘additional instructions’ and specs I could provide along with the drawings which allowed more information to be attached to each measurement, linking between the drawings and the specifications.
Discrepancy checking:
After uploading the drawings I ran a discrepancy check across the drawings to flag up any issues noticed by the AI between the different drawing sheets themselves and also comparing against Civils.ai’s knowledge of building codes in the project location.
Upon checking the discrepancy check results I found that civils.ai had identified 11 issues across the drawings with 1 critical issue which was a discrepancy between the subgrade thicknesses shown across the drawings which was a genuine critical issue! It found 4 major issues and 6 minor issues. I found that 3 were false positives. Finding the critical issue alone more than justified running the check. Civils.ai provides clickable references taking the user to the identified discrepancy which facilitated an easier checking process as you could identify which specific drawing sheets the AI had identified as being problematic.
Overall verdict:
I based the results of the quantity takeoff against my own verified numbers using the traditional manual approach.
Category | Verified (manual) | Deviation | % deviation | |
Drainage total | 2,195 m | 2,198.89 m | + 3.89 m | + 0.18% |
Manholes total | 149 | 149 | 0 | 0.00% |
100mmØ foul | 1,165 m | 1,165.96 m | + 0.96 m | + 0.08% |
100mmØ storm | 825 m | 826.48 m | + 1.48 m | + 0.18% |
150mmØ storm | 156 m | 156.46 m | + 0.46 m | + 0.29% |
225mmØ storm | 5 m | 5.82 m | + 0.82 m | + 16.40% |
ACO Raindrain B125 | 44 m | 44.17 m | + 0.17 m | + 0.39% |
Gross error (Σ abs. deviation ÷ verified total) | — | — | 3.89 m | 0.18% |
I then proceeded to check and verify each of the discrepancies raised by the platform to confirm if it was correct or a false positive, in order to score each criteria.
| Civil Quantity takeoff | Discrepancy checking |
Speed | 6/10 (50% less time than human measurement) | 9/10 (it took less than one minute to review a full plan set of 50 drawings) |
Accuracy | 9/10 (very similar level of accuracy to verified numbers) | 7/10 (27% false positive rate with 3 out of 11 issues raised being false positives) |
Reliability | 9/10 (worked across both scanned and vector PDF of varying quality) | 7/10 (worked across both scanned and vector PDF of varying quality but with 1 extra false positive on scanned) |
I arrived at the scores above as the takeoff itself completed in 1 hour and 30 minutes, compared to the 3 hours or so it took me to arrive at verified numbers, however this was more than compensated for as the results were highly accurate and worked even against poorer quality scanned drawings I ran through the system.
The discrepancy check layer added another measurable benefit to the experience of using the software and ran quickly, flagging real discrepancies across the drawings. However, several false positives were flagged during the discrepancy check, which is to be expected when using an AI platform. It was noted that the discrepancy checked similar to the quantity takeoff worked on PDF drawings of varying quality.
7. Head-to-Head: Civils.ai vs Claude
This is where the benchmarking got interesting, as the same drawings were uploaded into Claude (Claude Opus 5, tested August 2026) and I used prompts to trigger the tool calling functions and vision language model within Claude to analyse the drawings.
Quantity takeoff for Civil Construction:
The drawings I uploaded were typical Civil Engineering project drawings which were complex. I found that the quantity takeoff produced by Claude was almost entirely wrong, besides one exception. The total count on the different manholes was mysteriously correct. I realised on closer inspection and looking into the reasoning steps and tool calling by Claude that it had in fact extract the vector text from the drawing and counted the number of instances of the text MH, leading it to count the correct number. Any measurements of areas or linear measurements were completely incorrect and Claude had not provided any feedback on the likelihood of it being inaccurate, I would therefore class this as a typical AI hallucination. To test my theory I printed and scanned the drawing and tested it again and found that the manholes count was now incorrect.
It was also difficult to check the accuracy of the takeoff produced by Claude as no annotations or linked measurements were provided along with the table. If it were not for my previously verified measurements the results would have been hard to prove and required a full manual takeoff.
Discrepancy checking:
The discrepancy check ran just as fast as on Civils.ai as on Claude and raised 22 issues. Both identified that there was an issue with the sub-grade thickness, which was a genuine critical issue in the drawings. However the extra 11 issues found by Claude (making it 14 in total) were in fact false positives, which led to me spending more time reviewing the answer compared to Civils.ai.
Verdict:
For quantity takeoffs there was a world of difference in terms of quality and reliability between my verified numbers and Claude’s.
Category | Verified (manual) | Claude | Deviation | % deviation |
Drainage total | 2,195 m | 2,388 m | + 193 m | + 8.79% |
Manholes total | 149 | 149 | 0 | 0.00% |
100mmØ foul | 1,165 m | 1,450 m | + 285 m | + 24.46% |
100mmØ storm | 825 m | 650 m | − 175 m | − 21.21% |
150mmØ storm | 156 m | 200 m | + 44 m | + 28.21% |
225mmØ storm | 5 m | 12 m | + 7 m | + 140.00% |
ACO Raindrain B125 | 44 m | 76 m | + 32 m | + 72.73% |
Gross error (Σ abs. deviation ÷ verified total) | — | — | 543 m | 24.74% |
For discrepancy checking the difference was more nuanced with Civils.ai taking a slight edge by providing far fewer false positives making a final review more manageable.
| Civil Quantity takeoff | Discrepancy checking |
Speed | 8/10 (it took around 25 minutes to process the quantity takeoff, although it could occasionally get stuck and needed me to reset the tool limit several times) | 9/10 (it took less than one minute to review a full plan set of 50 drawings) |
Accuracy | 4/10 (besides counting manholes where the text was vectorised on the PDF it was entirely wrong) | 4/10 (a 63% error rate with 14 false positives being raised by the AI out of 22 issues) |
Reliability | 2/10 (switching from vector to scanned PDF resulted in unusable) | 5/10 (worked across both scanned and vector PDF of varying quality but with 16 false positives with scanned PDFs vs 14 with vector) |
8. Head-to-Head: Civils.ai vs another AI for Quantity Takeoffs
The third tool is unnamed for licensing reasons. It's a well-established takeoff platform marketed as trade-agnostic, and I'd expect these results to hold across that category rather than being specific to this one product. I did this to see how much difference there could be between equivalent AI for construction and if all drawings in construction could be considered equal. Since Civils.ai made a decision to specialise in Civil Engineering projects I figured there could be some difference between the performance of AI across different disciplines.
Quantity takeoff for Civil Construction:
Approaching this review I’d assumed there would not be a significant difference between AI built for quantity takeoffs however based upon my testing of the more generalised AI branded as being built for all trades is that it proved to be inaccurate when it came to measuring lengths of utilities, reinforced concrete but proved to be relatively accurate as measure external finishing areas from the drawings. It’s well known that many AI models fine tuned on construction drawings are disproportionately trained on floor plans, leading to promising results on internal wall, floor area and MEP fixture counting. Here is a summary of the major deviations on the utilities measurements.
Category | Verified (manual) | Generalised AI | Deviation | % deviation |
Drainage total | 2,195 m | 1,842 m | − 353 m | − 16.08% |
Manholes total | 149 | 147 | − 2 | − 1.34% |
100mmØ foul | 1,165 m | 980 m | − 185 m | − 15.88% |
100mmØ storm | 825 m | 610 m | − 215 m | − 26.06% |
150mmØ storm | 156 m | 185 m | + 29 m | + 18.59% |
225mmØ storm | 5 m | 22 m | + 17 m | + 340.00% |
ACO Raindrain B125 | 44 m | 45 m | + 1 m | + 2.27% |
Gross error (Σ abs. deviation ÷ verified total) | — | — | 447 m | 20.36% |
Possibly due to the structural differences in the formatting between Civil Engineering drawings and Architectural drawings used to train other AI models have led to this difference in accuracy and reliability in comparison to Civils.ai with this difference in accuracy becoming more apparent when using scanned PDF drawings.
Discrepancy checking:
The more generalised AI software for quantity takeoffs tested did not have a discrepancy checking feature included and therefore this was excluded from any testing
| Civil Quantity takeoff | Discrepancy checking |
Speed | 7/10 (it ran slightly faster than Civils.ai at 1 hour and 20 minutes) | - |
Accuracy | 5/10 (external paved areas were measured almost perfectly however utilities and concrete which are core to civils were entirely wrong) | - |
Reliability | 4/10 (switching from vector to scanned PDF threw off the area measurements) | - |
9. Security, Compliance & Data Handling
As many of us working in the built environment space are aware, many of our projects are governed by confidentiality requirements and non-disclosure agreements, meaning care must be taken with how and where we share any data which could be in breach of these agreements.
There are some concerns around using AI tools where training is allowed on the uploaded documents and prompts, as is the case with some frontier model providers, where users must take an active step to opt-out of training.
In Civils.ai terms and conditions they make assurances that customer data is not used for AI model training by default and that they have achieved cybersecurity certificates such as the CSA Cyber Essentials and SOC2.
Both ultimately allow for training opt-out so both are suitable choices for those concerned about data privacy.
10. Support & Community
I discovered it was extremely easy to speak with the founders and wider Civils.ai community, either by email or via their Discord message board where there were 1300+ other users I could communicate with users building their own Plugins, for example an integration with PlanSwift allowing quantity takeoffs processed in Civils.ai to be ported across.
I asked a question on if it was possible to add more groups of measurements to the quantity takeoff and received a reply back with the answer in under an hour.
11. Final comparison table verdict
In order to create my final verdict I created a combined table with the scores for each software I tested as part of this review to build up totals for each and enable a final comparison weighing speed accuracy and reliability all being equally important. It could be argued that speed should be weighted lower, as generating fast wrong answers is still of no use, however I have assumed that by moving faster the user could iterate closer to the correct answer in some cases.
| Criteria | Civil Quantity takeoff | Discrepancy checking |
Speed | 6 | 9 | |
Accuracy | 9 | 7 | |
Reliability | 9 | 7 | |
Claude | Speed | 8 | 9 |
Accuracy | 4 | 4 | |
Reliability | 2 | 5 | |
Other Construction AI for Quantity Takeoffs (Non-specific to Civil) | Speed | 7 | - |
Accuracy | 5 | - | |
Reliability | 4 | - |
Time for a drumroll please… Revealing the final combined score table and the results of this benchmarking exercise.
| Civil Quantity takeoff | Discrepancy checking | Gross error vs verified |
24/30 | 23/30 | 0.18% (3.89 m) | |
Claude | 14/30 | 18/30 | 24.74% (543 m) |
Other Construction AI for Quantity Takeoffs (Non-specific to Civil) | 16/30 | - | 20.36% (447 m) |
In conclusion I found that AI platforms like Civils.ai provide a significant difference in accuracy and reliability over generalised AI applications like Claude for technical tasks such as quantity takeoffs and discrepancy checking. By integrating more deeply into the workflow of the estimator, allowing for measurements to toggled on and off for checking, allowing for editing of measurements and for totals to update dynamically, has huge impacts when it comes to the efficiency of checking and validating AI measurements. It was noted that even within the world of AI for construction drawings that not all construction AI are created equal.
Contractors and consultants should take seriously the applicability of the AI model for their particular nature of work and not assume that generalised models are capable of understanding the nuances of their drawings especially for complex Civil drawings.
RLB frames the winners in this industry shift within the construction industry as the startups and firms that turn proprietary data into a trained advantage rather than treating AI as an off-the-shelf purchase. The gap between the three tools here is that argument made visible at the drawing level.
It is the opinion of this reviewer that when it comes to Civils projects that Civils.ai is the king… for now.
12. Additional notes on methodology
All three tools were given the same 50-sheet civil drawing package (foul and storm drainage, external paving, RC foundations), tested first as native vector PDFs and again after printing and rescanning at 300 dpi.
The baseline verified takeoff was completed completed using PlanSwift before any tool was run and took 3 hours by a chartered civil engineer with 10 years' experience. All deviations are measured against that baseline.
Accuracy is reported as gross error: the sum of absolute deviations as a percentage of the verified drainage total. Each discrepancy was manually verified and classed genuine or false positive with missed issues not being score as there was no full list of defects existing for this set.
Speed, accuracy and reliability are scored out of 10 and weighted equally. The question which this review aims to answer is whether a general-purpose model or a trade-agnostic platform can substitute for civils-specific software on civils drawings.
Tested 21/08/2026 on Civils.ai Team Plan, Claude Opus 5 via Claude App, and a third trade-agnostic platform unnameable under its evaluation terms.
