Working copy
Draft text, rebuilt from the manuscript on every change. Stubs and source passages are still here, and none of it is final. Back to the working copy contents
Drafted The chapter is written and still being revised.
The designers and consultants: a second reading
A design firm makes its fee on the documents and loses it after them. The drawings and the specification go out to tender, and the questions start. Bidders ask what a detail means. Addenda go out. During construction come requests for information, submittal reviews, change pricing and the occasional claim. Most fees pay for all of this with a fixed slice, and a good share of the work traces back to gaps and conflicts in the documents themselves.
The offer to designers and consultants is less administration, earned by checking the documents before they are issued. The check is a second reading, by machines that have no stake in the answer.
Who they are
Three groups hear three versions of the offer.
The architects. On a building, the architect usually leads the design team and administers the construction contract. The architect's set is where structure, mechanical, electrical and finishes have to agree. Most building contracts are for one fixed price, so the owner issues no list of quantities to check against.
The engineers. On roads, bridges and water works, the consulting engineer leads. The firm draws the plans, writes the list of bid items with their quantities, prepares the estimate and often measures the work for payment. Unit-price contracts pay by measured quantity, so every item is counted and agreed through the job.
The testers. This is every firm paid to check someone else's work: materials testing labs, inspection firms, commissioning agents, checking engineers and independent safety assessors. Cost consultants belong here when they act as the owner's independent estimator.
Where the hours go
Administration after issue is where a design fee runs short. The fee guide of the Royal Architectural Institute of Canada gives a typical split of 2.5 percent of the architect's fee for bidding and 22.5 percent for contract administration. A practice note from the American Institute of Architects warns that the American rule of thumb of 20 percent will almost always lose money in that phase. In its example, the budget allows under ten hours a week for site visits, submittals, questions, payment requests and changes, all together.
The questions come in the hundreds. A 2013 study of 1,362 projects, most of them in Australia and New Zealand, found an average of 796 requests for information per project, about ten for every million dollars of construction. Each took about a working day to review and answer. The median reply took almost ten days. A study cited there classed about one request in eight as unnecessary, such as a question the documents already answered.
These hours are not like design hours. Design is often billed by the hour, so a firm that designs faster earns less. Administration usually sits inside a fixed or capped fee, so a firm that administers less keeps the difference. That is why saved hours, which would not move a firm billing by the hour, matter here.
Nobody knows yet how much of the administration a check would remove. The large study lists causes such as coordination and incomplete documents, but gives no shares. Two small studies give shares for a handful of projects, which is too little to promise a percentage. The offer is a direction, not a number.
The test of the documents
Hand the issued documents to several bots, each working alone, and have each build its own model of the project. Where the models agree, the documents say one thing. Where they differ, a bidder would have asked a question or priced a risk.
This copies what a bidder already does. An estimator rebuilds the project from the drawings and the specification, in the head or in measuring software, and every question a bidder sends marks a place where that rebuild failed. One bot that builds a model proves little, because it may have guessed. Several bots that build the same model show that the documents are clear. Where their models differ, the difference points to the sheet or the clause that can be read two ways.
Designers already trust this logic. On British highway structures, the checker's analysis must be independent of the designer's, done without exchanging calculation sheets. The internet's engineers raise a specification to their top grade only when two independent teams have built products from it that work together, which tests the text, not the products. Accredited labs compare their results on the same sample with other labs' results, and when they disagree, either a lab or the method gets fixed. One German engineering firm calls its checking work the four-eyes principle.
The check starts with what a bid prices: counts and quantities, such as how many doors, how much concrete, how many metres of pipe. Positions and shapes come next, and a full model last.
The bots cannot do the full version yet. A 2026 benchmark gave seven leading models 324 small changes to make in existing building models. The best averaged about half, and none fully solved more than one task in thirty. Those tasks start from a finished model and a plain instruction. Building a model from drawings and a specification is harder.
A failed reading has two possible causes, weak documents or a weak bot. So each bot sits a set of reference projects first, where the right model is already known, and only bots that pass count as readers. Agreement among them shows that the documents are clear, not that the design is right. Bots from different makers can also share one blind spot if they run on the same underlying model, so "independent" needs a written meaning, as it already has for the checking engineer.
What they need
The architects need fewer questions during construction. The questions come mostly from the documents. A small Queensland study of four building projects found that most of the questions came from inconsistencies, errors or omissions in the documentation. A study of one public building in Edmonton sorted about 1,400 construction messages and found that more than half concerned design revisions, a quarter of them from coordination between disciplines. The check would catch the duct drawn through a beam, the specification section that disagrees with the drawing, the door schedule that does not match the plans, and the missing dimension that each bot fills in differently. The architect keeps the design, the seal and the last word on what the documents mean.
The engineers need their quantities counted twice. On civil works, the same Queensland study found that only about a third of the questions came from missing, unclear or wrong information, so "fewer questions" is the weaker line for them. Before tender, bots count the bid items alone and the counts are set beside the engineer's list. Bots that agree with each other and not with the list point to an error in the list. Bots that disagree with each other point to an unclear drawing. After award, the same counting can support measurement for payment. Ontario's engineering regulator has accepted checking software output by hand-calculating a random sample of it since 2011, so the method is familiar. Highway owners are also moving toward the model as the contract document. Once the model governs, the bid item list can be counted from it, and the list retires when two bots agree on the counts and the model governs.
The testers need a method to be accredited against. Independence is what they sell already, so nobody has to persuade them that a second party should check. Labs work under one international standard and inspection bodies under another, and labs already compare results with each other as a condition of accreditation. Canada has a precedent for judging software this way: the Standards Council of Canada accredits independent safety assessors for railway systems as inspection bodies. For architects and engineers the draw is less administration. For testers it is revenue, and it is better to say so plainly.
What they fear
A list of known gaps. A check before tender leaves a record of what was flagged, and without a rule for closing each flag, the report is evidence against the firm that received it. Each flag needs a recorded outcome: fixed before issue, answered by addendum, or accepted as intended, with a reason.
A higher bar. Once such a check exists, skipping it may start to look careless. Some firms will read that as new exposure.
The fee. An owner who sees fewer questions may trim the administration fee, and rival firms may lower theirs. The first hours saved only stop a loss. After that, the lasting gain is fewer disputes, not a wider margin.
The seal. Every code lets a professional sign only what they prepared or checked, and nobody has written what "checked" means when a machine did the counting. Engineers carry liability for their quantities and estimates, and Canadian courts have let a contractor sue the engineer behind a tender's documents despite the owner's exclusion clause. A regulator could also treat vouching for a counting tool as engineering.
The cost. Small firms cannot carry the check alone. It spreads only if the owner pays for it as a named lump-sum item, which matters most where transport departments pay for design at cost plus a fee under a ceiling.
The testers have their own fears. No method exists yet to be accredited against. A missed flag needs a report that states what it covers, as an inspection report does: an opinion on whether the documents are clear, with stated limits. And no firm can check documents it helped to produce, or a bot it helped to build.
What a standard does for them
Administration shrinks in stages, and each stage needs the one before it.
- Read. Bots read the tender set and list the conflicts and gaps they find. That removes some bidders' questions and some addenda, and nothing new has to exist first.
- Count. Several tested bots measure the bid items, each alone. That removes quantity errors, measurement disputes and some change orders. It needs reference projects, and a tester to qualify the bots.
- Rebuild. Several bots each build a model from the documents, and the models are compared. That removes coordination questions during construction. It needs bots far better than today's.
- The model governs. The checked model becomes the contract document. That ends arguments over the drawing against the model, and the separate upkeep of the bid item list. It needs the owner's rule that the model governs, and someone who signs the model.
- Administer. Bots answer routine questions from the model and show the source, check submittals and measure progress for payment. That removes most routine administration. It needs a written meaning of "checked" for each task. The professional signs and takes the exceptions.
The first stage can be tried now. The second is near wherever a good model exists. The third is far. The second stage is also where this chapter meets the rest of the book: counted quantities are what a standard bot prices, and a price is only as good as the count under it.
A standard is what turns each stage from one firm's habit into something an owner can ask for. It gives "independent" and "checked" their written meanings. It gives the testers a method to be accredited against, the owner a clause to put in the terms of reference, and the designer a report in a common format with a rule for closing every flag. Without it, each check is a private opinion. With it, the check is a second reading anyone can repeat.
Some things will not shrink. Ground nobody foresaw, an owner's change of mind, judgement on site and arguments about money stay with people. The offer is less routine administration, never none.
There is a longer road past this one. A bot that cannot build the model from the documents could not build the building from them either. One day the same check could tell a builder's machines whether a set is fit to build from. That is a story for the end of the book. The ask for now is smaller, and different for each group. For the architects: which questions on your last project would a coordination check before tender have caught? For the engineers: how large a difference between the bots' count and yours would make you look again? For the testers: what would you need, in writing, to run this check for an owner and put your name on the report?