Chris M. / 7 August 2026 / Reviewed 7 August 2026

Construction Processes You Can Automate Today, Ranked by Evidence

An operations director has been shown three demonstrations this quarter. Each one read a drawing on screen and produced a schedule, and each one was the most impressive thing in the room. Meanwhile the delivery tickets are still being typed into a spreadsheet by hand, because nobody has ever demonstrated that at a conference.

The construction processes that automate reliably today are the rule-bound ones with no judgement in them, such as verifying subcontractors under the Construction Industry Scheme (CIS) and holding the fixed field set that health and safety law requires. Extracting fields from paperwork and turning speech into text form a second tier that works well and is built around human review rather than instead of it. Interpreting drawings, which is the capability sold hardest, is the one where no vendor documentation could be found making a verifiable claim at all.

Ranked by what suppliers are prepared to put in their own technical documentation, in other words, the order that gets presented in a sales meeting comes out close to upside down.

This guide uses one test throughout. A capability counts as established when the company selling it documents it, with its limits, in its technical documentation rather than on a marketing page. That is a demanding standard and it disqualifies a good deal of what is currently being sold into this industry.

Tier one: the rule-bound processes, where the rule has no judgement in it

These are not artificial intelligence at all, and that is precisely why they are first. They are deterministic, the rule is published, and the failure mode is a missed deadline rather than a wrong answer.

Construction Industry Scheme returns are the clearest case in UK construction. Monthly returns must reach HMRC by the 19th of every month following the last tax month, and the penalty ladder starts at £100 at one day late, rises to £200 at two months, and reaches £300 or 5 per cent of deductions at six months, with a further penalty at twelve months ¹. There is no interpretation in that. A date, a submission and a reminder are the whole of it.

Subcontractor verification is the same shape and slightly more interesting. The details supplied to HMRC must exactly match the details the subcontractor used to register, verification is required before paying a new subcontractor and again where one has not appeared on a return in the current or last two tax years, and commercial software is required for verifying 50 or more subcontractors ². Exact match against a published key is exactly what software does well. It is also, for the same reason, the process that fails silently when the data going in is sloppy, so the automation worth building is the one that flags a mismatch loudly rather than the one that runs quietly.

The health and safety record is a third. Under the Reporting of Injuries, Diseases and Dangerous Occurrences Regulations (RIDDOR) a record must be kept of any reportable injury, over-seven-day injury, disease or dangerous occurrence, and it must carry the date and method of reporting, the date, time and place of the event, personal details of those involved, and a brief description of the nature of the event ³. That is a small, fixed, entirely structured field set, and capturing it reliably is a forms problem rather than an AI problem.

The boundary in every one of these is the same and it is worth naming. The recording is mechanical; the classification is not. Whether an event is reportable is a judgement. Whether someone is employed or self-employed is a judgement, and the CIS duty set carries an employment status declaration with a penalty of up to £3,000 attached to getting it wrong ¹. Automate the record and the deadline. Do not automate the decision.

Tier two: extraction and transcription, where the review is the design

This is where most of the available value sits, and where most of the disappointment comes from misunderstanding how the technology is meant to be used.

Start with what the vendor documentation will not give you. Microsoft publishes no headline accuracy figure for its prebuilt document extraction models, and states that performance must be evaluated by the customer on their own content ⁵. Its guidance on model accuracy likewise gives no percentage for prebuilt models ⁴. AWS publishes a figure for one narrow product on one document type, discussed below, and none for extraction generally ⁷. No accuracy figure for invoice extraction was found from any of the three vendors during this research, so anyone quoting you a percentage is quoting something their supplier has not published.

What the vendors do document is a workflow. Microsoft describes using confidence to determine whether to accept a prediction automatically or flag it for human review, and recommends targeting a training score of 80 per cent or higher, or close to 100 per cent for sensitive cases such as financial records ⁴. Its worked illustration has results at or above a chosen confidence going to straight-through processing and everything below it going to a person ⁵. AWS describes the same pattern and suggests a threshold as low as 50 per cent for archival purposes but 90 per cent or higher for business processes involving financial decisions ⁶. The human review step is not an admission of failure. It is the product.

Two documented limits matter more in construction than elsewhere. Microsoft explains why a word-level accuracy figure misleads: one wrong character in a document might be insignificant, unless it is in the text representing the paid amount, in which case the whole document is wrong. Its example shows an entity error rate of 20 per cent where the word error rate was 6 per cent ⁵. Measure the fields that matter, not the characters. And AWS documents inconsistent results when extracting from tables with merged cells spanning multiple columns, or rows and columns that differ from other parts of the same table ⁶, which describes a subcontractor's payment application almost exactly.

The best published number in this area shows how careful you have to be with published numbers. AWS reports that on an internal dataset of the fronts of United States driving licences from 50 states, the lowest F1 accuracy between states is 95 per cent ⁷. That is a genuinely strong result on flat, standardised, well-lit documents of a single known type, and AWS states in the same document that performance depends on image variation, confidence thresholding, human review, workflow consistency and periodic testing for drift, and will differ between applications ⁷. It is not a number that transfers to a photograph of a crumpled delivery ticket taken in a site cabin at dusk.

Transcription follows the same shape with one finding that deserves attention from anyone employing a UK site workforce. AWS reports a median word error rate of 2.4 per cent on its evaluation set, and within that a statistically significant gap between native English speakers at 0.9 per cent and non-native English speakers at 4.8 per cent ⁹. Those figures come from read speech in controlled conditions, and AWS warns that applications in different acoustic conditions will have differing error rates ⁹. A construction site is close to the worst case: Microsoft documents that background and environmental noise affects accuracy, that very specific accents may lead to error, and that a word which does not exist in the model produces a transcription error, which is what organisation-specific jargon is ⁸. Its conclusion is that users should expect a certain level of inaccuracy when designing an application ⁸.

The practical remedy is unglamorous and effective. Google's guidance is to position the microphone as close as possible to the person speaking, particularly when background noise is present ¹⁰. A supervisor holding a phone to their mouth in a quiet corner produces usable text. The same supervisor with the phone in a pocket beside a breaker does not. Google states that its recogniser is built to ignore background voices without additional noise cancelling and, in the same guidance, still tells users to get the microphone close ¹⁰, and both halves of that are worth taking seriously.

Tier three: what is being oversold

Two things belong here, one on the evidence and one on the absence of it.

Drafting and summarising work, and they are being sold as though they produce finished output. Anthropic states that even the most advanced language models can sometimes generate text that is factually incorrect or inconsistent with the given context, and that while the recommended techniques reduce this they do not eliminate it, so critical information should always be validated, especially for high-stakes decisions ¹¹. That is the vendor's own position. A monthly report drafted from site records is a first version for somebody to correct, which is worth a great deal, and it is not a document to issue unread.

Interpreting drawings is the harder case, and honesty requires stating what kind of finding this is. No vendor documentation could be found making a verifiable capability or accuracy claim for extraction, classification or quantity takeoff from construction drawings. The nearest primary evidence is Microsoft's documentation for its layout model, which describes figure handling as extracting text as words and paragraphs, with embedded images not supported, and bounding regions for figures covering only the core content and excluding captions and footnotes ¹². That is locating where a figure sits on a page and gathering the text near it. It is not interpreting what the figure depicts.

Products in this space exist and some of them may work well. What could not be found is a vendor stating in its own documentation what they achieve and on what basis. Treat that as untested rather than disproved, and treat a supplier's inability to point at their own documentation as the answer to your question.

One further signal is worth recording, because it runs against the direction most people assume. Google's managed "Human in the Loop" product for Document AI is listed in its documentation as deprecated, with a deprecation date of 16 January 2024 ¹³. No rationale is given. The practical consequence for a buyer is that the review step which every vendor's documentation says you need is, at least with that provider, now yours to build rather than something you can purchase.

The pattern underneath the ranking

Read the three tiers together and a single rule emerges, which is more useful than the list itself.

Automation is reliable in inverse proportion to how much judgement the task contains. A deadline has none, so it automates completely. Pulling a total off an invoice has a little, in deciding whether the number was read correctly, so it automates with a confidence threshold and a reviewer. Deciding whether a variation is compensable has a great deal, so it does not automate at all, and the useful question there is whether the evidence can be assembled faster for the person who decides.

That rule also tells you where the cost hides. Every tier two process carries a review cost that must appear in the business case, because a task that saves 20 minutes and needs five minutes of checking saves 15. The method for putting a defensible figure on that is in the guide to calculating return on construction AI.

Where to start, and what this does not tell you

Start in tier one, even though it is the least interesting, because it pays back without needing anyone to trust anything and it builds the credibility that tier two will need. Then take one tier two process with high volume and low consequence, delivery tickets rather than payment certificates, and run it with the confidence threshold and the named reviewer the vendors describe.

Three limits on this guide. It ranks by documented capability rather than by value to your business, and the two are different: a well-evidenced capability applied to something you do twice a year is worth less than a shakier one applied to something you do daily. It is dated, because vendor documentation on this changes quickly and everything here was read on 7 August 2026. And it contains no measured outcome from a UK construction business, because no such published study was found, which is discussed further in the guide to where the time actually comes back.

The ordering question of which of these to attack first, given a sceptical team and limited time, is answered from a different angle in the earlier piece on the five construction workflows to automate first, which ranks by payback speed rather than by evidence.

Sorting a real business's processes into these three tiers is the first stage of any workflow automation engagement, and it is usually the stage that changes what gets built.