Agentic Quantification of a Federated Healthcare BIM

AEC firms typically face three options when the federated BIM model needs to be quantified, reconciled against contract documents and made auditable: do the work manually with internal teams, license an off-the-shelf 5D BIM or quantity take-off product, or commission a custom engagement. Each alternative has clear failure modes. Manual reconciliation consumes a quantity surveyor's week per month and produces a snapshot that is obsolete within days. Off-the-shelf products are general-purpose, demand standardised input data, and cannot adapt to the bespoke procurement schedules, contract programme formats and specification conventions of any given project. Generic AI tools — whether assistants or large language models — fabricate plausible-sounding answers when the underlying data is missing or contradictory; the result is confident classifications that cannot be defended in a project-board meeting. Clients choose Leveret Systems for the hybrid architecture that sits between those alternatives. The service combines deterministic rules — the cheap, fast, predictable layer that classifies the bulk of elements before any AI is asked to reason — with an agentic layer that handles the genuinely ambiguous tail. Every commit is backed by a citation to the property field, the specification page, the procurement schedule row or the prior decision on a similar element that justified it. Items the agent cannot confidently resolve route to a tightly-scoped human review queue rather than being silently committed. In a domain where every line of a Bill of Quantities must be defended to the client and to the project board, that abstention behaviour is the safety feature. The architecture is integration-first. Revit and IFC models, Primavera P6 and revision-tracking programmes, NEC and JCT procurement schedules, and specification PDFs all flow through the same pipeline without forcing the client to migrate to a new platform. The cross-project knowledge layer accumulates across engagements, so subsequent projects start where prior projects ended — an architectural property no off-the-shelf BIM authoring tool can replicate. Leveret Systems is not a platform vendor. We do not replace the quantity surveyor, the planner or the BIM manager. We leave the architecture, the code and the institutional knowledge with the client team — and we step aside once the service is running inside their delivery workflow.

Before engaging Leveret Systems, the federated BIM model carried 23,000+ elements across multiple disciplines, each requiring classification into procurement packages and cost categories. The project folder contained 4,500+ files growing organically across teams and consultants, with the authoritative version of any given document hard to identify. The contract programme had drifted from the model after twelve months of variations and supplier slips. Bill of Quantities reconciliation was running roughly a week per quantity surveyor per month. The priced delta on a change order took three days to compute and document. Monte Carlo risk analysis on the programme had been run once at tender stage and never refreshed — the inputs had gone stale and no one had the capacity to rebuild them. Specification compliance was tracked informally and could not be defended to the project board with audit citations. Off-the-shelf 5D BIM products were a poor fit because they demanded standardised input data the project did not have; generic AI tools were dismissed because they fabricated plausible-sounding classifications that the QS could not stand behind.

Before engaging Leveret Systems, quantity surveyors and BIM coordinators traditionally extracted element schedules from Revit manually, exported to spreadsheets, mapped line items to NRM2 categories by hand, reconciled against the incoming Bill of Quantities row by row, and refreshed the entire exercise every time a variation landed. The work consumed a full week per QS per month, and the result was obsolete within days. Parallel teams in the supply chain — main contractor QS, subcontractor estimators, cost consultants — rebuilt the same view from scratch, producing different numbers from the same source model. Specification PDFs were read by hand to confirm compliance against standards; the cross-references lived in informal spreadsheets and tribal knowledge that walked out of the door when staff left.

The Leveret Systems agentic quantification service delivered measurable impact across cost planning, procurement coordination, programme risk and compliance audit. On the validation engagement, the full federated model — 23,466 elements across 4,500 project files — was processed end-to-end in under four minutes. Quantity surveyor reconciliation cycles that previously consumed a full week per month dropped to minutes for the equivalent dataset. The priced delta on a variation, previously a three-day exercise, now resolves in single-digit minutes with full audit citations. The service's deterministic-first design means roughly half of all elements classify without AI reasoning at all — the cheap, fast, predictable layer carries the bulk of the work, and the more expensive reasoning layer handles only the ambiguous tail. This keeps unit costs low (per-element processing typically below £0.002 in compute) and makes nightly re-runs economically viable in production, rather than the once-at-tender exercise quantification has historically been. Beyond raw speed, the larger return on investment comes from the auditability of every output. Each classification carries a citation back to the property field, the specification page, the procurement schedule row, or the prior decision on a similar element that justified it. The project board can defend the cost plan, the risk register and the procurement linkage with traceable evidence rather than tribal knowledge. Items the agent cannot confidently resolve route to a tightly-scoped human review queue rather than being silently committed — so the quantity surveyor reviews a few thousand ambiguous items rather than starting from twenty-three thousand. The cross-project knowledge layer that accumulates with every engagement is the structural compounding effect. On rerun within a single project, roughly a third of elements resolved from the knowledge layer alone, without firing the agentic stack at all. The same effect compounds across engagements — making each successive project cheaper to deliver and more accurate than the last. This is the architectural property that an off-the-shelf BIM authoring tool cannot replicate. Programme risk reporting transitions from a quarterly artefact to a nightly instrument. Monte Carlo inputs — activity duration distributions, supplier lead-time distributions, element-to-activity links — are structured as live data, so a critical-path kernel reads them as a function rather than a static export. P50, P80 and P95 completion windows refresh against the latest model state on every run, with a tornado chart attributing slip to specific suppliers. Leveret Systems engagements deliver in two modes: per-project (one to three weeks, indicative £8-30k) and strategic embedded (four to six months, indicative £80-180k).

-

-