Back to AI in Action
Case Study 02 · AI in Action

BIMVeritas

Faster, more consistent BIM assurance for complex programmes.

On a regulated infrastructure programme, manual BIM QA/QC was taking 1–2 days per model and varying with the reviewer. BIMVeritas was built inside Arup's environment to automate repeatable model checks, support AI-assisted rule generation and summarisation, and keep project data inside Arup. The result is a reusable, governed assurance approach for BIM-heavy programmes.

Revit Navisworks IFC Civil3D Azure OpenAI Power BI
~80% ↓
QA/QC time per model · 1–2 days → hours
4 platforms
Revit live · Navisworks · IFC · Civil3D next
3 standards
ISO 19650 · COBie · RIBA
Defra deployed
Discussions in progress with other programmes
01 · The 30-Second Story

Why This Exists

Large volumes of BIM models needed to be checked quickly and consistently against strict information requirements. Manual QA/QC could take one to two days per model and depended heavily on individual reviewer judgement. BIMVeritas changes that — automating repeatable checks, using AI selectively for rule generation and summarisation, and keeping all project data inside Arup's environment.

Team
Erkan Sahin · Ashish Ranjan
Project context
NBC / SCAH — National Biosecurity Centre, Weybridge (Defra · APHA).
Problem
Large volumes of multi-discipline BIM models needed consistent QA/QC against strict standards (ISO 19650, COBie, RIBA, project-specific rules).
Users
BIM managers · information managers · discipline leads · model reviewers · project assurance teams.
AI role
Plain-language rule generation · run summarisation · conversational query of validation results · future clash intelligence.
Non-AI role
Deterministic model checking · rule-based validation · structured reporting · audit trail.
Main value
Faster QA/QC · more consistent checking · stronger governance · reusable Arup IP.
Evidence
~80% reduction in QA/QC time per model — from 1–2 days to hours. Proven · Defra Basis to confirm
Status
Revit add-in live; Navisworks, IFC and Civil3D next. Live · scaling
Reusable pattern
AI-assisted assurance workflow for BIM-heavy programmes — applicable to any repeatable, standards-aligned compliance check.
02 · Project Context

Why This Mattered on NBC / SCAH

NBC / SCAH is one of the UK's most significant biosecurity infrastructure programmes — and the BIM assurance task scaled with it. BIM quality is not an isolated technical activity. It affects design coordination, stage-gate assurance, client confidence, handover readiness, and future asset information.

Lifecycle point
How BIMVeritas helps
Design development
Checks whether models are structured, named and populated correctly before they move through review gates.
Coordination
Supports clash and issue identification across discipline models.
Information assurance
Provides repeatable checks against agreed standards and project-specific rules.
Stage gateways
Gives project teams clearer evidence of model quality before submission or acceptance.
Handover preparation
Improves completeness and consistency of asset-related information.
Lessons & reuse
Rule packs and checking logic can be reused across similar programmes.
Why this is a delivery story, not a tools story
Models had to be validated against ISO 19650, COBie, RIBA-stage expectations and project-specific rules — across architecture, structure, MEP and civil disciplines. This is not "BIM checking" in isolation. Information quality flows through coordination, assurance, handover and asset information — and ultimately governs the confidence the client can place in what's being produced.
03 · The Problem

Manual QA/QC Could Not Scale

A full check could take one to two days per model. Reviewers had to inspect naming conventions, data completeness, model structure, coordination issues and compliance with project requirements — every model, every review. Three problems compounded.

Problem 1 — Time

Manual validation took 1–2 days per model. That created a bottleneck before submissions, design reviews and stage-gate assurance.

Problem 2 — Consistency

Different reviewers could interpret the same checks differently. Outcomes varied by reviewer style, rigour and memory.

Problem 3 — Scale

As the programme grew, model volume increased faster than team capacity. Adding people was the only obvious lever.

The real question
How can we make BIM QA/QC faster, more consistent and more scalable — without weakening governance or exposing sensitive project data?
04 · Why AI Was Appropriate

AI Helps Where Rules Hit a Ceiling

BIMVeritas is not a chatbot bolted onto a BIM model. The core checking is rule-based and repeatable. AI is used selectively where it adds value — generating rules from plain language, summarising long QA/QC outputs, supporting natural-language queries, and (in future) helping reviewers triage clash results.

Rule engines alone can
  • Execute predefined checks
  • Flag basic compliance issues
  • Run consistently at speed
Rule engines alone cannot
  • Help users author new rules from plain-language intent
  • Summarise long QA/QC outputs into critical issues
  • Answer natural-language questions about results
  • Distinguish likely false positives in clashes

Where rules check, where AI assists

Drawing the line is the reason this case study generalises — start with the right tool for the job, not AI for AI's sake.

Activity
What BIMVeritas does
Capability type
Naming conventions
Validates files, elements and layers against agreed rules
Rules / automation
Data completeness
Checks required parameters and properties
Rules / automation
Standards checks
Tests against ISO 19650, COBie, RIBA and project-specific requirements
Rules / automation
Rule generation
User describes a check in plain language; tool generates structured rule
AI-assisted
Run summarisation
Condenses long QA/QC outputs into critical issues and actions
AI-assisted
Conversational query
Ask questions about validation results in natural language
AI-assisted
Clash intelligence
Future capability to help identify likely false positives
AI-assisted · emerging
05 · What Was Built

A Rule Engine With an Optional AI Layer

Modular by design — AI is enabled or disabled per project. The Revit add-in is live; Navisworks (federated models), IFC (open exchange) and Civil3D (civil disciplines) are next. Project data does not leave Arup's environment.

Rule generation

User defines intent in plain language → AI helps create the structured validation logic.

Run summarisation

Long QA/QC outputs condensed into critical issues, owners and recommended actions.

Conversational analysis

Ask the model results: "what's critical?", "which elements are non-compliant?", "are these clashes valid?"

Try Mode

Validate without saving any data — even inside Arup's environment. Useful for sensitive engagements and onboarding.

User workflow

Day-to-day use, end-to-end

Input
Model data (Revit · IFC · Navisworks · Civil3D) + project-specific rule packs
A
Open / upload model User opens a Revit model (or uploads from another source) and selects the relevant project rule pack.
B
Run validation The rule engine performs structured checks (naming, completeness, standards, project rules). Precision and reproducibility live here, not in the AI layer.
C
Review issues + AI summary AI summarises the run, highlights criticals, and answers natural-language questions about results.
D
Assign actions, fix model, re-run Reviewer applies professional judgement to exceptions. Re-run confirms remediation. Results become QA/QC evidence for stage gates.
Output
Structured reports · insight-led summaries · cross-platform export to Power BI · auditable QA/QC trail
06 · How the Process Changes

What Changes for the Project Team

The reviewer's role doesn't disappear — it shifts. Less time searching for issues, more time applying professional judgement to the issues that matter.

Before BIMVeritas
After BIMVeritas
Manual review of each model, line by line.
Automated validation with expert review on exceptions.
One to two days per model.
Hours, not days.
Reviewer-dependent outcomes.
Consistent rules applied across models.
Manual interpretation of long outputs.
AI-supported summary of critical issues.
Static reporting.
Unified, insight-led QA/QC reporting.
Rules slow to create or update.
Plain-language AI-assisted rule generation.
Governance depends on tool choice and user behaviour.
Data governance designed into the platform.

Who feels the change

BIM manager

Less time searching for issues across long output logs, more time deciding what matters and how to fix it.

Information manager

Stronger evidence that information requirements are being checked consistently — a defensible audit trail for assurance.

Project manager

Reduced risk that model quality becomes a bottleneck before submissions or stage gates.

Discipline lead

Clearer issue visibility across the discipline model and fewer late surprises in coordination.

Client / assurance team

More confidence that information quality is repeatable, traceable and aligned to agreed requirements.

Bid / account lead

A differentiated offer for regulated-sector clients — BIM assurance + responsible AI + controlled data governance.

07 · Value Delivered

Six Facets of Value

Each item is labelled by evidence quality so the case study doesn't overclaim — what's proven on Defra, what's targeted, and what's still to confirm.

Time Proven · Defra

~80% reduction in QA/QC time per model — moving from one to two days to hours. Faster validation reduces delay before submissions, design reviews, assurance gates and downstream coordination.

Quality Proven

The same rule packs apply consistently across models, disciplines and reviewers. Variation drops; the QA/QC process is no longer dependent on individual reviewer style or memory.

Cost Indicative · cash saving TBC

Primarily through reduced manual effort, reduced rework, and avoided third-party validator licence costs. A quantified saving can be developed once model volumes, hourly rates and avoided licence costs are confirmed.

Risk Proven · governance pattern

Repeatable, traceable checks aligned to project rules. Reduces the risk that information-quality slippage passes through stage gates unnoticed and that QA/QC becomes a delivery bottleneck as model volume grows.

Governance Proven · architecture

All AI processing inside Arup's environment; no third-party AI dependency; Try Mode allows validation without saving data. Project data does not leave the controlled boundary.

Commercial Targeted · being productised

A differentiated offer for regulated-sector clients (government, health, defence, biosecurity) combining BIM assurance, responsible AI and controlled data governance.

08 · Governance & Responsible Use

Why It's Safe to Deploy on a Live Programme

BIMVeritas isn't presented as an autonomous decision-maker. It's a controlled assurance tool. The governance pattern is the same shape the playbook recommends for every advisory AI use case.

Hosted inside Arup

Azure OpenAI running in Arup's secured environment. Project data does not leave the controlled boundary. No third-party AI dependency.

AI optional per project

Programme leads can enable or disable the AI layer. The rule engine works without it — there's always a non-AI fall-back for sensitive engagements.

Human review retained

AI-generated rules and summaries are reviewed before being relied on for critical decisions. AI proposes; specialists approve. Outputs support rather than replace professional judgement.

Standards-aligned

Validation logic anchored to ISO 19650, COBie and RIBA. Rule packs are project-specific. Compliance isn't a feature — it's the spec.

Try Mode

Validation runs without saving data — useful for onboarding, training, and confidentiality-sensitive engagements where outputs are evaluated but not retained.

Auditable trail

Every rule run, every AI summary, every reviewer decision becomes part of the QA/QC evidence used at stage gates and assurance reviews.

For client conversations
The reason BIMVeritas is defensible on a Defra programme is the same reason it generalises: the AI layer is bounded, optional, hosted inside our environment, and answers to a rule engine that produces an auditable trail. That's the shape the playbook recommends for every advisory AI use case.
09 · The Technical Bit

For Specialists — How It's Built and Validated

A deeper-dive section for technical readers. The case study is readable without it; this section adds depth on architecture, validation activities, and what we still need to confirm with the build team.

For technical readers

"The rule engine checks. AI explains and assists. Humans review and approve."

Architecture layers

Layer
What it does
Why it matters
Model interface
Connects to Revit (live); Navisworks, IFC and Civil3D planned/emerging.
Allows checking across discipline models and platforms.
Rule engine
Runs structured checks against naming, parameters, completeness, standards and project-specific rules.
Provides repeatable and auditable validation.
Optional AI layer
Helps generate rules, summarise outputs and support natural-language querying.
Improves speed and usability without replacing assurance logic.
Reporting layer
Produces issue breakdowns, summaries and Power BI exports.
Turns QA/QC results into project-level insight.

Validation activities (how we prove the tool works)

Activity
What is demonstrated
Baseline comparison
BIMVeritas results compared against manual QA/QC outputs.
Rule testing
Each rule pack tested against known compliant and non-compliant model examples.
False positive review
Where the tool flags issues that are not genuine failures.
False negative review
Whether the tool misses issues that expert reviewers would catch.
Reviewer sign-off
BIM / IM specialists approve rules and outputs before use.
Repeatability test
Same model + rule pack re-run to confirm consistent results.
Cross-discipline test
Validated across architecture, structure, MEP and civil models.
Reporting test
Exported summaries and dashboard outputs match underlying results.
To confirm with build team — before publication
  • Exact architecture and hosting model — to avoid overclaiming how data is processed and stored.
  • Which Azure OpenAI or AI services are used — for governance, security and licence implications.
  • Whether RAG is used, or only structured validation outputs are sent to AI.
  • Rule representation and versioning method — for auditability and repeatability claims.
  • Number and type of models tested — to support evidence claims.
  • Basis of the ~80% time-reduction figure — to make the quantified value robust.
  • Security / cyber / digital review status — to support regulated-sector positioning.
10 · What Colleagues Can Reuse

Five Patterns to Lift

The reusable lesson is not "use BIMVeritas". It's the underlying delivery patterns — applicable to any repeatable, standards-aligned compliance problem.

Assurance pattern

Convert repeatable manual checks into structured rules. Find the work that's deterministic and codify it.

AI-assist pattern

Use AI selectively — to generate rules, summarise outputs, and interrogate results. Not as the spine.

Governance pattern

Keep sensitive data inside approved environments, retain human review, give the AI a non-AI fall-back.

Product pattern

Build reusable Arup IP from a real project problem — start narrow, scale through rule packs and platform extensions.

Client proposition pattern

Offer AI-supported assurance as part of BIM, IM, digital delivery and PMO services — not as a standalone tool sale.

Reusable assets — now available across Arup
Rule packsProject-specific validation libraries.
QA/QC workflowThe 4-step Rules → AI summarise → Query → Re-run pattern.
AI interaction patternsRule generation, summarisation, conversational query.
11 · Client-Facing Proposition

How This Becomes an Advisory Offer

BIMVeritas underpins a productisable client offer at the intersection of BIM, information assurance, regulated-sector data governance and responsible AI.

Proposition statement
Arup helps clients improve BIM and information assurance by combining domain expertise, structured rule-based checking and responsible AI. Project teams validate models faster, improve consistency, reduce manual effort and maintain stronger control over sensitive project data.

Offer components

Where this lands in the playbook
Maps to Service Line 03 — Programmes & Project Management (P&PM) as model assurance / QA-QC at scale, and supports BIA / MC engagements where BIM and information quality govern decision confidence.
12 · What's Next

Planned Enhancements

Each future capability inherits the same governance baseline as the current build. Outstanding items to confirm with the build team are listed in the technical section above.

01

Advanced clash intelligence

Automatic classification of false vs real clashes — moving the AI layer from descriptive to evaluative.

02

Continuous learning

Feed previous QA/QC runs and user feedback back into rule generation and summarisation.

03

Deeper platform integration

Tighter coupling with Revit, Navisworks, and Civil3D — reducing the export/import friction.

04

Toward bounded autonomous QA/QC

Predictive issue detection and bounded autonomous agents — only as the audit trail and approval flows mature.

Built by Erkan Sahin Programme APHA Weybridge · SCAH / NBC · Defra
Final Takeaway

BIMVeritas shows how AI can support project delivery when applied to a clear, repeatable assurance problem — checking BIM models faster, more consistently, and within a controlled governance environment.

Rule precision + AI interpretation + Bounded governance

The shape that lets a tool live safely on a live Defra programme is the same shape that lets it scale across Arup.

Continue

← All Case Studies

Back to AI in Action — National Highways, Custom Power BI Visual, Multi-Agent Project Intelligence.

See all

Multi-Agent Project Intelligence

The next evolution of the same governance pattern — five domain agents on Azure OpenAI, with Teams approval before write-back.

See the case

AI for Advisory Services

How BIMVeritas underpins Service Line 03 — Programmes & Project Management — as a productisable client offer.

See strategy