We’re putting $5.5M behind research and development of smaller, specialized AI models for real-world deployment — Read our manifesto →
    All posts

    [Case Study] 500,000 Work Orders, 17 Checks Each: Balfour Beatty's Auditable GenAI

    How a facilities operator used retrieval, rules, citations, and dashboarding to review maintenance records against policy with more than 98% reported accuracy.

    Conscious Engines

    Industry: Facilities management
    Organization: Balfour Beatty Investments
    Use case: Work-order compliance review and operational assurance
    Evidence basis: NTT DATA customer case using Balfour Beatty-reported results
    Disclosure: This is an independent analysis by Conscious Engines. The source is vendor published, and some portfolio-scale figures describe planned coverage after a production pilot.

    1. Outcome at a Glance

    Balfour Beatty Investments manages more than 500,000 work orders per year and applies 17 checks to each record. Its generative AI review reportedly achieved more than 98% accuracy compared with human review in the evaluated workflow.

    Key Outcomes

    More than 500,000

    Annual work orders

    Portfolio volume reported in the cited case.

    17

    Checks per work order

    Review workload reported in the cited case.

    More than 98%

    Accuracy versus human review

    Reported pilot performance.

    100% of work orders

    Planned review coverage

    Target state reported in the cited case.

    MeasureReported resultEvidence note
    Annual work ordersMore than 500,000Portfolio volume
    Checks per work order17Review workload
    Accuracy versus human reviewMore than 98%Reported pilot performance
    Planned review coverage100% of work ordersTarget state
    Previous review teamAbout 70 coordinatorsHistorical process context
    Reporting availabilityPower BI output by 8 a.m.Workflow service level

    The distinction between current and target state matters. The public story describes a pilot in production and a goal of reviewing all annual work orders. It should not be written as proof that every one of the 500,000 orders had already been processed at the stated accuracy.

    2. The Operational Problem

    Facilities work orders contain operational and contractual evidence: asset, fault, response time, actions, completion, notes, materials, and approvals. Large portfolios need to verify that each record complies with service policy and supports reporting.

    Manual sampling can miss errors. Full manual review is expensive and slow. Pure rules struggle with unstructured technician notes, while a free-form language model can reach the right-sounding conclusion for the wrong reason.

    The task is ideal for a hybrid system. Structured fields can be checked deterministically. Narrative notes can be interpreted by a language model. Policy can be retrieved from controlled documents. Every conclusion can cite the rule that supports it.

    3. What Was Built

    The solution used Microsoft Azure and OpenAI capabilities, integrated with Power BI reporting. The public account emphasizes traceability to policies.

    System at a Glance

    Work-order ingestion

    Read structured fields and technician notes.

    Rule checks

    Validate dates, statuses, required values, and thresholds.

    Policy retrieval

    Find the relevant contractual or operating requirement.

    Language model

    Interpret notes and explain potential noncompliance.

    Citation layer

    Link each finding to policy evidence.

    Dashboard

    Deliver prioritized findings to reviewers by 8 a.m.

    LayerFunction
    Work-order ingestionRead structured fields and technician notes
    Rule checksValidate dates, statuses, required values, and thresholds
    Policy retrievalFind the relevant contractual or operating requirement
    Language modelInterpret notes and explain potential noncompliance
    Citation layerLink each finding to policy evidence
    DashboardDeliver prioritized findings to reviewers by 8 a.m.
    Human assuranceReview exceptions and correct the source process

    Seventeen explicit checks make the use case bounded and testable. Each check can have its own precision, recall, severity, and escalation threshold. This is preferable to one opaque “compliance score.”

    A smaller specialized model can handle classification and extraction at volume. A stronger model can review ambiguous notes. Deterministic logic should retain numeric, date, and state validation.

    4. How It Reached Production

    The program turned policy review into a repeatable overnight workflow.

    Define each check precisely. State the policy source, required inputs, pass condition, severity, and human action.

    Create a labeled evaluation set. Human-reviewed work orders should cover normal, ambiguous, incomplete, and high-risk examples.

    Optimize for exception review. The system should prioritize likely material failures, not overwhelm staff with low-value flags.

    Make evidence auditable. Every finding should show the work-order fields, relevant note text, policy citation, model version, and confidence.

    Measure coverage and reviewer agreement separately. A 98% accuracy rate can hide false negatives if the positive class is rare. Report precision and recall per check.

    The daily dashboard creates an operational rhythm. Findings can be corrected while the work remains recent, and recurring causes can feed training or supplier management.

    5. What Real-Estate and Facilities Leaders Should Take Away

    Balfour Beatty's case shows how generative AI becomes safer when surrounded by explicit checks, retrieval, citations, and human exception review. It also shows why high-volume back-office assurance is attractive: the denominator is clear and every item can be tested.

    A production-ready work-order intelligence layer combines field speech-to-text, structured extraction, enterprise RAG, task-specific classification, deterministic rules, and an auditable review interface. The system should improve the maintenance process, not merely generate more flags.

    The primary metric is cost per correctly reviewed work order, with false-negative rate on critical checks, reviewer time, remediation speed, and repeat noncompliance as controls.

    Sources