Case Study · Construction · February 18, 2026 · 10 min read
A project manager who used to run 4 jobs can now run 12. The admin runs itself. The PM runs the work.
A Los Angeles-area general contractor wanted to bid more work without hiring more senior labor. We showed them what AI could take off the PM's plate, then we built it: bid analysis, permit tracking, subcontractor coordination, and invoice review, all handled. The proof follows.
running active jobs
workflows in production
up from 4-5
during deployment
The company wanted to grow without adding senior labor.
The client is a Los-Angeles-area general contractor with 8 project managers running concurrent commercial and residential jobs. On paper the PM runs the work: scheduling, coordinating subs, managing the owner relationship, holding the schedule and budget. In practice each PM was carrying 4 to 5 jobs and spending 40 to 50 percent of the week on admin.
The admin was the same loop every week. Extracting scope and quantities from RFP PDFs by hand. Logging into city portals to check permit status. Emailing subs for RFQs and chasing responses. Matching receipts to budget control forms and purchase orders. Reconciling field changes with the accounting system.
Some of it was outsourced to a Filipino BPO for $8,000 per month. The output was acceptable, but the time-zone gap meant a 24-hour minimum turnaround and uneven quality. When something needed to ship today, the PM did it themselves anyway.
The owner ran the numbers. Eight PMs at $130K loaded, spending half their week on procedural, rule-based, high-volume document work, came to roughly $520K per year of senior project labor. The BPO added another $96K per year on top.
PMs were not management. They were document-processing nodes.
Walking a PM's actual week, the pattern was clear within an hour. The role had been absorbed into document processing. Read the RFP, extract the scope. Open the permit portal, read the status. Send the RFQ, score the responses. Read the invoice, match to the BCF, flag the variance. None of it required construction judgment. All of it required attention.
The judgment work, the part that needs a PM's experience, was the smaller share of the week. Negotiating a change with the owner. Solving a sequencing conflict between trades. Walking the site to verify quality. Knowing when to escalate. It kept getting squeezed because the document work was always more urgent.
The company was already turning down work it could have won, because adding a PM meant a six-month ramp and another six figures of loaded cost. The constraint was never skill. It was admin throughput per PM.
Across 8 PMs at $130K loaded, that was roughly $520K per year of senior project labor on document processing. Plus $96K per year to a BPO. Plus the ceiling on how much work the company could safely bid.
A multi-agent admin layer. Four workflows, one governance fabric.
We built four production agents on top of Power Automate, Azure OpenAI, and the Procore connector the company already had. Each one took a document-heavy workflow and turned it into a PM-supervised pipeline.
Bid analyzer
Extracts scope, materials, and labor estimates from inbound RFP PDFs. Returns a structured bid summary the PM reviews instead of assembles. Cuts bid prep from a half-day to 30 minutes.
Permit tracker
Monitors permit-portal status across the LA-area municipalities the company works in. Alerts the PM on stalled applications, missing items, and aging inspections.
Subcontractor coordinator
Sends RFQs, follows up on non-responses, scores incoming bids against the project scope, and surfaces the top three options for the PM's decision.
Invoice reviewer
Matches receipts to budget control forms and purchase orders in QuickBooks. Flags exceptions, variances, and unauthorized charges before they hit the books.
Security and compliance
Cloudflare Zero Trust at the edge. Every AI action is audit-logged. Aligned to ISO 42001 controls. Every PM sees only the projects they own.
Human in the loop
The agents do not approve, sign, or submit on their own. The PM stays in the decision seat. The agents do the assembly. The PM does the judgment.
The permit tracker caught a missing inspection before it was buried.
About 40 days into deployment, the permit tracker flagged a missing intermediate inspection on an active project. It had been scheduled, missed by the city inspector, and never rescheduled. Under the old workflow, the gap would not have surfaced until the final walkthrough, by which point the work above it would have been covered.
The PM rescheduled, the work was inspected as built, and the project closed without incident. Construction counsel later confirmed that an OSHA review of the missed sequence would likely have triggered a fine in the $50,000 range, plus the cost of opening up finished work for re-inspection. That single catch covered most of the Phase 1 build.
The deployment ran nine months without an OSHA violation across the active portfolio. Causation is hard to claim cleanly, but the permit tracker is now part of how the company operates.
The proof: PMs run far more work, and the BPO is gone.
Phase 1 Math
- PM time on admin: 40 to 50 percent → 10 to 15 percent
- Active jobs per PM: 4 to 5 → 12 to 15
- BPO contract cancelled: $96,000 per year recovered
- Recovered PM admin capacity: ~$390,000 per year at $130K loaded
- OSHA fine avoided on first catch: $50,000 single event
- Phase 1 build: $60,000
- Ongoing infrastructure: $1,800 per month
- Payback: under 60 days
The 3x in jobs per PM is the structural change. The company can now bid more work without adding senior labor. That is the line that matters at the leadership table.
The BPO contract was cancelled cleanly at the end of the first 90 days. The PMs say their week shifted from responding to whatever document just came in to actually running the work. The company took on two new clients the following quarter without adding a single PM.
Microsoft governance, Procore-native, ISO 42001-aligned.
The construction industry runs on Procore and QuickBooks. The architecture wrapped both, plus the city permit portals the team already used, under a single governance fabric.
Power Automate
The orchestration layer. Each of the four agents runs as a Power Automate flow with role-based access against the source systems.
Azure OpenAI
The reasoning layer for bid analysis, scope extraction, and exception flagging. Runs in the company's Azure tenant, not a public endpoint.
Procore connector
Read and write against the project records the company already kept in Procore. Subcontractor data, RFI logs, and budget data all surface in the agent workflows.
Permit-portal scrapers
Custom monitoring against the LA-area city permit portals. Authenticated daily checks with PM-level alerting.
Custom QuickBooks integration
The invoice reviewer matches against QuickBooks receipts, BCF entries, and POs. Variances surface to the PM before they post.
Cloudflare Zero Trust + ISO 42001
Zero Trust gates every dashboard surface. Every AI action is audit-logged. The deployment is aligned to ISO 42001 controls for AI governance.
We have done construction before
A different client, a different scope, the same posture.
For California's largest hillside structural engineering firm we built Project IQ, a custom AI-native operations platform that replaced a $1M Salesforce migration in 30 days, recovered 80 hours per week, and is now retiring Salesforce entirely. Same posture every time: we wrap the systems the team already uses, we keep humans in the decision seat, and we secure everything under Cloudflare Zero Trust with ISO 42001-aligned controls.
The Heed Method
Every engagement starts with a POC against your real data.
These four agents did not start with a discovery deck. They started with a two-week POC against the contractor's actual Procore records, real permit portals, and live QuickBooks data. The POC proved the architecture worked before the company committed a production budget.
Keep reading.
Structural engineering case study
$1M Salesforce migration walkaway, 9.6x ROI in the first year. The flagship build.
Read the case study →Executive intelligence case study
62.5 hours per week recovered for a Southern California fresh produce distributor.
Read the case study →Apps and Dashboards
The offering that wraps the systems your team already uses with AI-native workflow.
See the offering →The ROI math
How we calculate whether a workflow is worth building before we write a line of code.
Read the post →The hidden headcount tax
Why disconnected systems force you to hire 3 people for 1 job, and how to find yours.
Read the post →Talk to Heed
Bring us your tools list, your PM headcount, and your current admin pain. We will scope the build.
Book a call →PMs running 4 jobs when they should be running 12? Let us look at the admin layer.
If your project managers are spending half their week on document work, the constraint is not skill, it is admin throughput. The 30-day diagnostic produces the math, the architecture, and the honest answer on whether a multi-agent layer beats hiring more PMs.