Using AI in Construction Estimating and Bids

AI can extract quantities, classify documents, compare supplier returns and draft tender narrative. It cannot know that a drawing is superseded, a site constraint is missing or a productivity rate is commercially unrealistic unless the estimating system gives it that context. The objective is therefore not an “automatic bid”. It is a faster, more consistent chain of evidence from information received to quantity, rate, allowance, qualification and approval.

This article is current to 31 July 2026 and focuses on UK construction businesses. Health and safety legislation, building-control arrangements and public procurement differ in territorial reach, project type and contracting authority. Procurement Act 2023 guidance mainly concerns covered public procurement; private tenders remain governed by their contracts and applicable law. Obtain project-specific professional and legal advice.

Select one estimating bottleneck

Start with a repeatable task whose output can be checked against source documents. Good candidates include classifying tender files, identifying changed drawing references, extracting a defined quantity type, normalising subcontractor returns or drafting a first qualification schedule from an estimator’s marked assumptions.

Avoid beginning with the final tender price. That number combines design maturity, quantities, market testing, logistics, programme, risk, overhead, cash flow and commercial strategy. A plausible result can conceal several wrong assumptions.

Write a one-sentence decision boundary. For example: “The tool will propose wall-area quantities from the issued drawing set; the quantity surveyor will verify revision, measurement basis, openings and waste before any rate is applied.”

Build an information spine

Create a controlled tender workspace before adding AI. Every input needs a stable identifier, title, revision, issue date, originator, status and received time. Keep tender bulletins and answers linked to the documents they change. Hash or otherwise preserve the file received so a later audit can distinguish a revised document from a model change.

The estimating record should connect:

• employer or client requirement;
• drawing, model, specification or schedule;
• measurement rule and classification;
• extracted or measured quantity;
• rate build-up and quotation evidence;
• programme, logistics and productivity assumption;
• risk, opportunity and contingency treatment;
• qualification or exclusion; and
• named approval and tender submission.

The RICS Cost Prediction professional standard addresses scope, data, skills and systems and stresses consistent classification and transparency. Use its principles even when the pilot performs only one small step. AI output without a recognised cost structure is hard to reconcile with later cost plans or project controls.

Do not silently flatten ambiguity. If two drawings disagree, create an exception with both references. If a specification uses an unfamiliar product description, retain the original text and route it for classification. The estimator needs to see uncertainty before it hardens into a quantity.

Test quantities against measurement rules

For a take-off pilot, define the element, drawing type, scale, units, inclusions, deductions and treatment of waste. Select a test set spanning straightforward, dense, revised and deliberately incomplete information. Keep entire projects together in training or test partitions so near-identical sheets do not inflate performance.

Useful quality measures include:

• absolute and percentage quantity variance against verified measurement;
• missed items and duplicated items per drawing;
• incorrect drawing-revision use;
• classification accuracy by element;
• rate of exceptions correctly declined;
• estimator review minutes per measured item;
• downstream value variance after rates; and
• percentage of output linked to a precise source region.

Report distribution, not only the average. A system with small errors on many low-value items and one large omission can look accurate while destroying the margin. Set a lower tolerance for safety-critical, programme-critical or high-value packages and require full review where information maturity is low.

Keep a simple comparator: the existing template, rule-based extraction or manual sample. AI should beat a credible process, not a blank page. If review takes longer than measurement, improve the interface or stop the use case.

Keep rates, inflation and risk visible

A rate is not a timeless fact. Store currency, location, base date, source, scope, productivity, labour composition, plant, material, waste, delivery, tax treatment and commercial adjustment. Never let a generated rate enter a bid because it “looks about right”.

When using historical projects, remove outcomes that would leak the answer into an earlier forecast and explain changes in specification or procurement route. Record whether an outturn reflects disruption, claim settlement or scope growth. Historical rates may carry obsolete methods, abnormal supplier discounts or regional labour conditions.

For every proposed rate, show:

• comparable jobs and why they are comparable;
• supplier quotations and their validity dates;
• excluded or provisional scope;
• index or market adjustment;
• productivity assumption and crew composition;
• quantity sensitivity;
• risk allowance and correlation with other risks; and
• the estimator’s accepted value and reason.

Public-sector teams can use the government’s Should Cost Model guidance to structure whole-life cost thinking. The Sourcing Playbook, updated in 2026, also emphasises robust commercial delivery. Neither converts an AI estimate into an independent cost assurance.

Do not optimise safety out of the tender

The HSE’s guidance on planning construction work states that significant hazards should be considered at tender or estimate stage and that appropriate allowance should be made in the price. Pre-construction information, design risk, access, temporary works, welfare, sequencing and competent resources are cost inputs, not prose added after the total.

Under the Construction (Design and Management) Regulations framework, duties remain with the people and organisations appointed to them. HSE’s Managing health and safety in construction, L153 explains the legal framework, while its current principal designer guidance sets out that role. A summarisation tool may surface a hazard reference, but it cannot certify that duties have been discharged.

Create mandatory blockers when:

• pre-construction information is absent or superseded;
• design responsibility is unclear;
• hazardous material surveys are missing;
• temporary-works assumptions are unpriced;
• access, lifting or edge-protection constraints are unresolved;
• programme compression conflicts with the proposed method;
• welfare or supervision allowances fall below the project plan; or
• a value-engineering suggestion changes safety or compliance.

Any model-generated saving that touches these items should go to competent design, construction and commercial review. Preserve rejected suggestions; recurring unsafe recommendations are a system defect worth measuring.

Govern tender narrative and supplier information

Generative tools are effective at producing fluent method statements and social-value responses. Fluency raises the risk that unsupported commitments pass unnoticed. Limit drafting to an approved evidence library and display source passages beside each claim. Prohibit invented project references, accreditations, staff availability, carbon results and delivery guarantees.

Supplier quotations and tender documents contain commercially sensitive information. Give access by project and role; do not place competitors’ material into a shared model context. Confirm in the supplier contract whether prompts and files are used for training, where data is processed, how long it is retained and which sub-processors receive it. The ICO’s AI contracts and third parties framework provides a practical review route where personal information is involved.

For covered public procurement, use the government’s live Procurement Act 2023 guidance collection. The July 2026 guidance on conflicts of interest is relevant when tools combine supplier data or individuals move between evaluation roles. Keep the authority’s required audit trail; an AI activity log is only one part of it.

Secure the document pipeline

Tender packs are untrusted input. A malicious or accidental instruction embedded in a document can try to redirect a connected AI system, expose other project data or alter an output. Apply the NCSC’s secure AI deployment guidance.

Use a quarantined ingestion step, malware scanning, file-type controls and separate project indexes. Retrieval should return only authorised documents. Give connectors read-only access unless a write is genuinely needed, and require human approval before output enters the estimating system or tender submission.

Test at least these scenarios:

• a late bulletin supersedes a priced drawing;
• a PDF contains hidden or white text;
• a spreadsheet has hidden rows and external links;
• two projects use the same drawing number;
• the supplier service becomes unavailable near submission;
• access is removed from a departing estimator;
• a model or extraction version changes mid-tender; and
• a generated qualification contradicts the priced assumption.

Maintain an offline submission pack and named fallback owner. The deadline does not become safer because a vendor outage caused it.

Prove value over 90 days

Use completed tenders for shadow testing before a limited live package:

Release should require zero unreviewed quantities or commitments entering the bid, no unresolved severe security or confidentiality event, complete revision traceability, and a genuine reduction in estimator touch time without deterioration in omissions or post-tender clarifications. Track false confidence: the number of outputs reviewers initially accept but later correct.

After submission, compare tender assumptions with buyout and delivery. Do not train blindly on won bids; winning may reflect underpricing. Include lost-bid feedback, supplier movement, change control and outturn where contractual permissions allow.

The archive’s guides on AI construction project management and safety, generative design in architecture and predictive maintenance in manufacturing-uk) extend the adjacent operating questions.

A bid remains a professional commitment

Scale the system element by element and project type by project type. Revalidate when measurement rules, document conventions, geography, procurement route or market conditions change. Monthly governance should examine major variances, rejected suggestions, model updates and user workarounds.

The desired outcome is not a tender that appears in minutes. It is a bid team that can move quickly because every material number and statement has an identifiable source, assumption, reviewer and consequence. AI deserves a place in that chain only when it makes the evidence clearer.