Is AI takeoff accurate? How to check it yourself

Nobody can tell you how close software will come on your drawings, and that includes us. We publish no accuracy figure. What you can do is run a tool on a job you have already bid, then compare it with your own takeoff wall type by wall type, not on the total. Your own numbers are what it gets measured against.

We build and sell one of these tools, so read this knowing that. What follows about where reading a drawing set goes well and where it goes wrong comes from building ours. Other tools are built differently, which is why the test further down is worth more than anything a vendor tells you, us included. By the end you will know what to look for and how to run that test yourself.

What does software read well on a drawing set?

A drawing set is mostly words and numbers printed in expected places, and that is where reading goes well. These are the things ours reads, and they are a fair place to start with any tool:

• Schedules: door and room schedules, finish schedules, and the marks and sizes in their rows.
• Partition types the legend spells out in words: stud size, gauge and spacing, board layers on each face, insulation, and how high the wall runs.
• Typical notes. A line like “ALL INTERIOR WALLS TO BE TYPE P1 U.N.O.” gives the walls with no tag that type, unless a wall is drawn plainly thicker than a P1.
• Heights where the set writes them: a ceiling height on a room tag, a ceiling height column in a finish schedule, a default ceiling height note. A ceiling height is not always a wall height, so you check these too.
• The scale in each title block, read for you to confirm, because every length on the sheet depends on it.

The harder part is everything a set does not print as plain words, starting with the walls themselves.

Where does software that reads a set go wrong?

These are the places we have seen reading go wrong while building ours. They are worth knowing before you test, because a demo on a tidy set will show you none of them:

• A symbol with no text beside it. Software that reads text does not see it at all. Ours reads text and does not count symbols.
• Deck heights. A set rarely prints one where it can be read off the plan. It sits in a section, or nowhere, and if the drawings do not give one, a tool that fills it in is guessing.
• Wall types by thickness alone. Several partition types often share one stud and board depth, so the drawn thickness of an untagged wall cannot tell them apart.
• Legends written as a code, a number for the stud and a letter for the board, instead of words. Ours does not read those yet.
• Exterior and structural walls, drawn thicker than any interior type, get no type in ours unless a tag beside them names one.
• The walls on the plan. Wall suggestions taken from a CAD drawing's lines miss walls and mark things that are not walls: door jambs, casework, furniture, dimension lines. In ours they are a beta, and you check and confirm each one. A scanned sheet has no lines to follow, so ours suggests no walls on it.
• Scans. A scan can turn a 3 into an 8, so every number ours reads from a scan is flagged for you to check. A full-size (36 × 24 in) scanned sheet does not finish reading on our site yet; smaller scanned sheets do.
• Ceilings in open areas. Where one ceiling type stops and the next starts is often not drawn as a wall, so nothing on the plan says where the line is.
• Doors with no swing. Cased openings and pocket, sliding and bifold doors draw nothing that tells them apart from wall, so ours places one only where a schedule's tag sits on the wall. A door's height is never drawn on a plan: it comes from the schedule, or from you.
• A number that reads two ways. A “20” printed as a door width could be feet or inches. Ours leaves it out and points it out.

Why does a published percentage tell you so little?

A single figure answers a question about somebody else's drawings. Before you weigh one, from anyone, ask what it was measured on:

• Which drawings: clean CAD files, or scans, addenda and sheets drawn by three different offices?
• Of what: walls by count, walls by length, or board by area? Is a wall found with the wrong type a hit or a miss?
• Against whose takeoff, and was that takeoff itself checked?
• Before or after a person corrected it?
• For which work? A figure for walls says nothing about ceilings, doors or the counts.

Two tools can print the same number from tests that have nothing in common. In Canada, a claim about how well a product performs has to rest on “an adequate and proper test”, and the proof is on whoever makes the claim (Competition Act, paragraph 74.01(1)(b); this is a summary, not legal advice). So asking how a figure was tested is a fair question, not a rude one.

That is also why we publish no accuracy figure. We have none we have checked well enough to print, and a figure measured on drawings we chose would tell you little about yours. What we publish instead is the working: every formula an interior wall goes through, run on one wall, in how to check our arithmetic.

How do you test a takeoff tool on a job you have already bid?

A demo runs on a set chosen to look good. Your own job is the test that counts, because you already know what it bought. Run it in this order:

1. Pick the job. One you bid recently, with your own takeoff, the addenda and the drawings you priced from. A messy one teaches more: a scanned sheet, a coded legend, a floor with two deck heights.
2. Write your numbers down by line first: length and board area by wall type, ceiling area by type, doors by mark and size, and every height you used, deck heights included. A total you remember is no comparison.
3. Run the same drawings through the tool, for the same scope, and note every question it asks you and every value it fills in without asking.
4. Compare, in this order, because each line depends on the ones before it:
   • Scales: every sheet's scale against its title block. One wrong scale moves every length on that sheet.
   • Wall types: is every type on the legend there, with the same stud, gauge, spacing, board layers on each face, insulation and height?
   • Length by wall type, never only the total. The example below shows why.
   • Heights: which came from the drawings, which it asked you for, and which it filled in. A height the drawings never print is a guess unless you gave it.
   • Board area by type, both faces, with the openings taken off.
   • Doors and openings by mark and size, with their heights.
   • Ceilings by type and area, the open areas included.
   • What it left out, and whether it told you it had.
5. Find the reason for every line that differs on the tool's own screen: a missed wall, a wrong type, a height, the waste, an opening. If the screen cannot show you which walls and which arithmetic made a number, that is your answer about the tool.
6. Trace one quantity all the way through by hand, from the sheet to the unit your supplier sells it in, with the tool's working beside you.

A worked example: the same total, the wrong board

Here is why the total is the wrong place to look. Say your takeoff has 400 LF of type A, with one layer of board on each face, and 300 LF of type B, a rated wall with two layers on each face, all 10 ft high. The tool comes back with 460 LF of A and 240 LF of B. These lengths are made up to show the arithmetic; they are not a run of any tool.

The length matches to the foot (213 m either way), so a check on the total passes. The board is 1,200 SF (111 m²) short, which is 38 sheets of 4 × 8 (1,200 ÷ 32 = 37.5, rounded up) before waste and openings. And the order is wrong twice over: 1,200 SF too much of type A's board and 2,400 SF too little of type B's, so if the rated wall takes a different board, both lines on the supplier's quote are off.

One 60 ft run typed A where the legend makes it B did all of that. Compared wall type by wall type, it shows on the first line of the table.

What does “the estimator checks each wall” mean, and why is it the point?

You will hear that a takeoff tool a person has to check is not worth using. That is the right test of a tool that hides its working, and the wrong one for a tool that shows it. Your name is on the bid, and you would check a junior estimator's takeoff before it went out. What matters is what the check involves: with the sheet and the arithmetic beside every number, checking is reading. Without them, it is a second takeoff.

In ours, checking looks like this:

• Before any wall, you approve the job's wall types, each shown beside the legend line it was read from. A wall only ever takes a type you approved.
• Wall suggestions are a beta. They come from a CAD drawing's lines and are outlined on the sheet in each wall type's colour, red where no type was read.
• Each wall is a row with its length, height, depth and type, what it still needs, and why the search doubts it, if it does. You look at it on the drawing, fix what is off, and approve it to the takeoff.
• The walls it missed, you draw.
• What the drawings do not say, such as a deck height, is asked once and never filled in. A height worked out from a section is offered for you to accept, never applied.
• Nothing reaches the takeoff until you approve it, and every quantity read from the drawings shows the sheet it came from and its arithmetic, down to the figure before it was rounded up.

It takes as long as reading the set properly takes. What you get is a takeoff you can trace, not a number you have to redo.

What goes wrong when estimators test a tool?

• Comparing totals. Errors in opposite directions cancel out, as the example shows.
• Testing on the vendor's sample set. It was chosen to work. Ours is a fictional community centre we drew to share, and it proves only that the software runs.
• Grading the arithmetic and not the reading. A formula can be right on a wall with the wrong type.
• Counting a wall as right because it was found, without looking at its type and height.
• Letting the tool's answer change what you remember of your own takeoff. Write yours down before you run it.

A checklist for the test

• A job you bid recently, with your own takeoff by line
• The same drawings, the same addenda, the same scope
• The scale checked on every sheet
• Wall types compared field by field
• Length and board by wall type, never only the total
• Every height marked as printed, asked or filled in
• Doors by mark, ceilings by type
• One quantity traced from the sheet to the order
• What it left out, and whether it said so

Where our own software stands

Ours reads text on the drawings (schedules, legends, notes and title blocks) and does not count symbols. Wall suggestions are a beta: you confirm each one and draw the walls it misses, and scanned sheets get none. A full-size (36 × 24 in) scanned sheet does not finish reading on our site yet; smaller scanned sheets do, so test ours on a job without full-size scans. Every quantity read from the drawings shows the sheet it came from and the arithmetic behind it, and we publish no accuracy figure. Your first bid is free, with no card, so run it on a job you have already bid and hold it to the checklist above. For the rest of what to compare before you buy any tool (units, assemblies and how pricing is counted), see choosing takeoff software as a subcontractor.