The State of UK Specifications · 2026 · Avoice Research
Are architects actually good at writing specs?
We graded 50 real architectural specifications, each issued to tender or contract on a live UK project, to find out. All 50 carried at least one coordination defect. This is what we found, what it costs on site, and the AI we built because of it.
The index · Specification Quality Index · first edition
00 / EXECUTIVE SUMMARY
Good at the building. Not at the 300 pages behind it.
Architects are good at everything that needs an architect: design intent, judgement, knowing what the building has to do. What fails is the reading. No one has the hours to check every clause against every drawing after every revision, and all fifty documents told that story, at every practice size, in every sector. The profession does not have a skill problem. It has a proofreading problem at a scale no person can staff.
The software is a false sense of security
Specs written in dedicated platforms carried the same defects as specs written in Word. The tool structures the writing. It never reads the finished document against the drawings. On site that surfaces as RFIs and variations priced against your client, from a document that looked controlled.
Some defects are liabilities, not tidy-ups
One in five documents specified a wrong product, value or standard. Built as written, that is a non-compliant building with the practice's name on the instruction. Under the Building Safety Act, the document trail is the liability trail.
The spec fails on its own terms first
Where one clause says 60 minutes fire resistance and another says 30, the contractor can price the cheaper reading and charge for the change. Højgaard settled what that costs: the more onerous requirement prevails, and a €26m remedial bill followed.
Open decisions ship inside signed contracts
76% of contract-issue documents carried open TBCs. Each one is a design decision handed to the contractor to price at their number, then recovered as a variation. The client pays for the blank, and the practice signed it.
Nobody's job is the whole document
Practices with no final spec-to-drawing pass accounted for most high-severity defects. The first person to read the whole spec against the drawings is usually the contractor, at their price. The gap is a missing step, not a missing skill.
01 / THE SAMPLE
Every practice size, every sector, same defects
The sample spans small studios to 50+ person practices, across residential, commercial, education, healthcare and hospitality work. No practice size or sector was immune. Select any bar to filter the sample; select it again to clear.
Showing all 50 documents
Documents by size of issuing practice.
Primary sector of each project. Hospitality + covers hospitality, leisure and mixed-use.
Defects appeared in both CAWS and Uniclass specs; the sample is not large enough to rank one against the other.
Total coordination defects per document, all classes combined, for the current selection. No document scored zero. Sample median: 12.
02 / THE DEFECT LEDGER
Six defect classes, ranked by frequency
Every defect falls into one of six classes. Select a row for the definition and a redacted example.
Not all defects are equal
Each document was graded by its most severe defect. One in five reached the critical tier: a wrong product, value or standard, stated plainly in a contract document.
Critical: wrong product, value or standard, with legal consequence if built as written. Major: needs an instruction to resolve. Minor: leftover content with no performance consequence. n = 50.
A critical defect is where cost becomes liability. A specification stating the wrong fire performance is an instruction to build the wrong building, signed by the practice. Under the Building Safety Act, the document trail is the liability trail. None of the ten critical documents showed any sign the defect was noticed before issue.
From defects to a grade
Each document's defect record converts into a 0–100 score and a grade from A to E. Weights follow consequence: one critical defect outweighs a dozen tidy-ups and caps the document at grade E. Full scoring model in Appendix B.
Documents by grade. The ten grade-E documents are the ten carrying a critical defect. Not one document reached grade A.
Contract-ready
No critical or major defects. References verified against the issued set, TBCs closed. Zero documents in the sample.
Sound
Minor inherited content or superseded references only. Nothing a contractor can price against.
Typical, the UK median
Multiple major defects: contradictions and open items that will surface as RFIs and priced ambiguity on site.
Exposed
Contradictions plus open TBCs carried into contract. The document actively transfers cost risk to the client.
Liability
At least one critical defect issued in a signed document: wrong product, wrong performance, wrong standard.
03 / WHAT IT COSTS
Every defect has a price, and the client pays it
Get It Right Initiative research puts the cost of avoidable error in UK construction at £10–25 billion a year. The sample shows how a specification contributes: defects cluster where one trade’s scope meets another’s, and open items survive into signed contracts.
Median open items per document at each stage, for documents observed at multiple stages. Tap a chip to add or remove a defect class; hover a line for the numbers. Most classes barely move between tender and contract. Drawing–spec mismatch is the one class that grows, because the drawings keep changing after the spec was last checked against them.
Median contract-issue document. Every open TBC and every contradiction is a separate opportunity to convert ambiguity into cost: the contractor prices the blank or the cheaper reading, then recovers the difference as a variation.
38 of 50 (76%)
Each dot is one contract-issue document, sorted by how many open TBCs it carried. Drag the slider to see how many contracts were signed with more than a handful still open. The 12 clean ones are the exception, not the rule.
Percentage of documents with at least one open defect of each class at each stage. Darker means more common. Select a row to focus that class in the chart above. Contradictions and orphan references are near-universal from Stage 4 onward and never get cleaned up; mismatches are the only class that gets worse as issue approaches.
Whatever is open at tender is usually still open at signature. The contractor prices the blank, or prices it low and recovers it later as a variation. The client pays either way, and the practice signed the document.
Real-life examples
Two cases where a specification defect became a documented bill, and one where catching errors early paid for itself.
MT Højgaard v E.ON [2017] UKSC 59: the contradiction that cost €26m
A 20-year design life and a named standard that could not deliver it, in the same document. The contractor followed the standard, the foundations failed, and the Supreme Court held it liable anyway: where a specification conflicts with itself, the more onerous requirement generally prevails. Remedial works cost over €26 million.
UK Supreme Court, 2017 · Robin Rigg offshore wind farm
Grenfell Tower: a substitution the document chain did not stop
Tendered with zinc panels, substituted under cost pressure for combustible ACM. The change passed through the documents without any control connecting it to its consequences. Seventy-two people died. This is the event behind the Building Safety Act and its statutory golden thread.
Grenfell Tower Inquiry Phase 2 Report, 2024
Error control pays for itself
Four major UK contractors applied systematic error identification across 25 projects and avoided £92.6 million in lost value. The same logic applies one document earlier, at the specification.
GIRI Productivity Training Commission, reported 2026
04 / AVOICE AI
The world’s first AI model fine-tuned for specification writing
Everything above is why we built Avoice AI: the first AI model trained specifically to write and coordinate construction specifications. Then we tested it the only fair way. We gave it three real project briefs, the same kinds of jobs as the fifty practice documents, and graded its specifications with the exact method we used on theirs.
Average score out of 100 across the three briefs, graded under the same rubric as the sample (Appendix B). Zero critical defects, zero major defects, grade A. n = 3 briefs; a full sample-matched run will be published with the next edition.
Residential refurbishment
Hospital refurbishment
Stair & timber junction package
Why does Avoice AI beat a human process?
A person checks a spec once, at the end, under deadline. The AI checks it up to 100 times before it goes anywhere. Each cycle it reads every clause against the drawings, finds a problem, fixes it, and reads again. It only stops when there is nothing left to find.
There is a second reason. No architect can know every system deeply: roofing one project, sprinklers the next. Manufacturers help, but a rep only writes about their own product. The AI has read the standards and product data for all of it. The architect decides what the building needs. The AI fills the knowledge gaps.
The human process interleaves writing and reviewing under deadline, and the review happens once, if at all. The Avoice AI ingests the project's drawings, models and schedules ahead of tender, drafts, then runs its check-and-repair loop until the checks return zero findings.
The hours, counted
A specification is weeks of senior technical time. The AI drafts a coordinated, tender-ready document in an afternoon, with your time spent reviewing instead of typing. Set the sliders to your numbers.
Next step
Benchmark your practice on a live project
Book a demo and we’ll run the Avoice AI on one of your briefs, then grade both outputs under the published method.
A / METHODOLOGY
How the sample was built and read
Sample
50 specifications issued at tender or contract stage on live UK projects, produced across a range of specification tools and practice masters, and reviewed by Avoice Research in mid-2026. Documents were obtained through Avoice’s specification review and platform work with the practices that produced them, and are analysed here in anonymised, aggregated form with the issuing practices’ permission.
Review process
Each document was read in full and cross-referenced against its issued drawing set where available. Defects were recorded against the six-class taxonomy (D/01–D/06) by an automated coordination pass, with human verification of every recorded defect before it entered the log.
What counted as a defect
Only objective, checkable failures: two clauses specifying the same element differently; a reference to a document absent from the issued set; a clause describing scope not present in the project; an explicit open item at issue; a citation of a withdrawn standard. Judgement calls about quality of specification writing were excluded.
Anonymisation
No practice, project, client or individual is identifiable in this report. Example clauses are redacted and altered in non-material details. Aggregate figures are reported only where the underlying group contains five or more documents.
The model test (Section 04)
Avoice specification model output was generated from three briefs with cached project inputs (drawings, schedules, required work sections) and graded under the identical taxonomy, severity rule and overrides as the sample documents. Grading was performed by Avoice; this is disclosed in the section itself. Machine output was not included in any sample figure, distribution or median elsewhere in this report. Independent grading of machine output, and a run across sample-matched briefs, will accompany the final edition.
Scoring
The full scoring model (formula, overrides, severity rule, counting rules and worked examples) is set out in Appendix B below. The defect log schema will be published with the next edition.
References
1. MT Højgaard A/S v E.ON Climate & Renewables UK Robin Rigg East Ltd [2017] UKSC 59.
2. Grenfell Tower Inquiry, Phase 2 Report (2024).
3. Get It Right Initiative, Strategy for Change and research reports (2015–2026).
4. Arcadis, Global Construction Disputes Reports (2015–2021).
5. Building Safety Act 2022, Part 4 (golden thread duties).
B / The scoring model
The Specification Quality Index converts a defect record into a single 0–100 score, so a specification can be graded, compared and tracked over time.
Override 1: any critical defect means grade E, score capped at 39
Override 2: grade A requires 85+ and zero critical or major defects
Weights follow consequence, not count: one critical defect outweighs a dozen tidy-up items. The overrides exist because a legally consequential error is not offset by tidiness elsewhere in the document.
How severity is assigned
Severity is assigned by rule, not impression. Every defect is tested against three questions in order, and takes the highest tier it qualifies for.
Would the building be wrong if built as written?
The defect, taken at face value, changes performance, compliance or safety: a wrong performance value, a product whose performance differs from the requirement, a standard whose provisions materially differ from the correct one, or a contradiction in which one available reading is non-compliant.
Does it take an instruction to resolve?
The defect requires a decision, clarification or instruction before the work can be priced or built as intended: a contradiction between compliant options, a drawing–spec mismatch, an orphan reference to a document the contractor needs, or an open TBC on a priced item at contract issue.
Is it inert?
No performance or price consequence: inherited content plainly inapplicable to the project, a superseded reference whose successor is materially identical, or a broken cross-reference within the document that does not affect scope.
Two counting rules keep the score honest. One defect, one count: a contradiction is a single defect however many clauses it touches. Grade down when in doubt: a borderline defect takes the lower tier, so the score understates the problem rather than inflating it.
Two documents, scored
Worked examples from the sample: one at the median, one showing why the critical override exists.
8 major × 4 = −32
9 minor × 1 = −9
SQI = 100 − 41 = 59 · Grade C
Six contradictions and two drawing–spec mismatches, all between compliant options; nine items of inert inherited content. The median UK specification looks like this.
6 major × 4 = −24
4 minor × 1 = −4
Numeric score 57 · Override 1 applies
SQI = 39 · Grade E
Numerically similar to Document 31, but one defect specified a fire performance below the required rating. A document that instructs the wrong building cannot grade above E.
About Avoice Research
Avoice Research is the research arm of Avoice Lab, the AI-native workspace for architecture and engineering practices. We publish evidence about how project documentation actually performs. The profession deserves better data about its own output.
Cite this report
Avoice Research (2026). The State of UK Specifications 2026. Specification Quality Index, first edition. avoice.co/research/sqi-2026