5 ways AI is changing construction takeoffs β and where humans still rule
AI is rewriting the QTO process, but not in the way the marketing decks promise. Five concrete shifts we see in 2026, and the parts of the workflow that still need a senior estimator.
Construction takeoffs are the unglamorous middle of every project. Before any pricing decision, before any tender submission, before any IPC reconciliation, somebody has to count and measure. Until 2024, that somebody was a senior QS with a printout, a scale ruler and a coffee.
In 2026, it is still a senior QS β but the workload has moved. Here are the five concrete shifts we are seeing across the contractors on ORKSTRA, and the parts of the job that genuinely still belong to humans.
1. From counting everything to verifying disagreements
Pre-AI, a QS counted every diffuser on every reflected ceiling plan, every door on every architectural sheet, every junction box on every electrical layout. Roughly 28 hours per BOQ on a typical UAE villa.
Post-AI, the QS spends most of the time on a much smaller surface: the items where the three AI models disagree. ORKSTRA's Triple-AI router agrees on roughly 85-90% of items immediately. The QS reviews the remaining 10-15% β usually fewer than 200 items on a 1,800-line BOQ.
The job is now disagreement triage. It takes about 5 hours instead of 28, and the output is more defensible because every decision has a human stamp on the contentious lines.
2. From single-source rates to multi-source price suggestions
Old workflow: open Excel, copy rates from last project, adjust by gut feel. New workflow: the system suggests a rate with four sources weighted behind it.
- 40% historical (your own past projects)
- 30% government (KAPSARC for KSA, CAPMAS for Egypt, UAE Stats)
- 20% crowd-sourced (anonymised contractor pool, k-anonymity β₯ 5)
- 10% heuristic AI fallback
The QS sees the suggestion, the spread between sources, and the confidence band. The decision is still theirs. What is gone is the hours of supplier-quote chasing for routine lines.
3. From PDF redlining to drawing-aware annotation
Old workflow: print the drawing, redline counts, photo back to office. New workflow: the AI overlays counts directly on the drawing in the browser, the QS clicks the ones to keep and rejects the ones to discard, and the BOQ row is auto-mapped to the cluster of measurements that justified it.
The killer feature here is not the AI vision. It is the link from BOQ line to drawing region. Two years from now, "why does this BOQ line cost what it costs?" will be answerable by clicking the line and seeing the highlighted regions across all relevant drawings.
4. From batch tendering to streaming pricing
When tender deadlines compress, the old approach was to throw more estimators at the file. The new approach is to keep one senior estimator and let the platform stream price suggestions as drawings are processed. By the time the QS opens the BOQ, 90% of it already has a draft rate.
The shift is psychological as much as technical. The QS is no longer the bottleneck. They are the quality gate.
5. From private BOQ archives to learning loops
Every accepted or rejected AI suggestion becomes training data for the per-tenant learning loop. After about 200 BOQ corrections, the system learns your house style β the rate adjustments you always make, the materials you always swap, the units you prefer.
On Aletlala's tenant, after roughly six months and ~340 corrections, the AI suggestion accept rate climbed from 62% to 81%. That is the compounding effect of a per-tenant loop. It does not exist in single-shot AI tools.
Where humans still rule
Five things we have not seen any AI handle reliably.
Contract-clause-aware pricing
Whether a delay-related variation is paid at BOQ rate, dayworks, or lump sum depends on contract clauses that an AI summarises but does not interpret. A senior QS reads the EOT clause and decides.
Buildability judgements
"This pour can be done in one go" versus "you need a construction joint here" is a buildability call. Drawings do not show pour sequence. Models do not understand site logistics. A human walks the site.
Relationship-driven supplier negotiations
The platform suggests a rate. The supplier negotiates. The senior QS extracts the last 4% of margin because they have known the supplier for a decade. That step is not automatable, and probably never should be.
Ambiguous specifications
When the spec says "Class B finish, similar to project X," the AI flags the ambiguity. The human resolves it by phoning the consultant.
IPC defence in disputes
When a consultant rejects a line, the QS argues. The argument blends contract, photo evidence, schedule context and relationship. AI can prepare the brief. The QS delivers it.
The actual time saving
For a contractor running 50 projects with average 1,200 BOQ lines each, the math is straightforward:
- Pre-AI: 50 Γ 28 hours = 1,400 QS-hours per cycle.
- Post-AI: 50 Γ 5 hours = 250 QS-hours per cycle.
That is 1,150 hours saved, or roughly 28 weeks of senior QS time returned to the business per cycle. Re-deployed to better tender quality, better margin protection and better variation management, that is a fundamental change in unit economics β not a 10% productivity nudge.
The 28-hours-to-5-hours number is not a marketing extrapolation. It is the measured number from Aletlala's first six months on ORKSTRA, across 47 BOQs.
Where to start
Two front doors.
- /demo β bring a drawing, we will run Triple-AI live and show you the disagreement view.
- /premium-stack β read the technical map of how the QTO pipeline fits together.
β Eng. Amr Shoieb