AI agents and MCP tools
Governed AI agents that work from your construction data
We connect AI agents to the systems you already run, such as Procore, Primavera P6 and your project files, through MCP servers with least-privilege access. Answers cite their evidence, every tool call is logged, and quality is scored against a written rubric before anything goes live.
The value it creates
Time saved
Less time digging for answers
Agents search project documents and system records for a question and return the answer with citations, instead of someone paging through submittals, specs and logs.
Better decisions
Answers you can check
Every output points back to the file, page or record it came from, so a project manager can verify an answer before acting on it.
Early warning
Schedule problems surfaced while drafting
Schedule agents return drafts with health checks, so missing logic and open questions are flagged for the scheduler while the schedule is still being built.
Visibility
See what every agent did
An MCP gateway records which tools were called, by which agent, with what result, so IT and leadership can see agent activity rather than guess at it.
Standardization
The same process every time
Specialized agents follow a defined workflow and are scored against the same rubric on every change, so quality does not drift with prompts or models.
Cost saved
Visible usage and fewer rework loops
Gateway analytics show requests and success rates per tool and per agent key, so you can see what usage is worth paying for, and cited drafts cut the back-and-forth that comes from unsupported answers.
What we build
MCP servers for construction systems
Servers that expose Procore or Primavera P6 operations to agents behind a clear boundary, including an offline XER path for read-only schedule analysis.
Tool routing for large APIs
Search-then-call routing so an agent finds the right operation for a task instead of loading an entire API surface into the model.
MCP gateway with telemetry
A gateway that injects credentials, enforces which tools each agent may call, and records every call with its inputs and results for review.
Document Q&A with citations
Agents that answer questions from drawings, specs, submittals and meeting notes, and cite the source for each answer.
Schedule agents
Specialized agents that turn project documents and client answers into draft CPM schedules for scheduler review, with health scores and XER export, as in Connect.
Governance and access controls
Least-privilege credentials, read-only defaults, human approval for writes, and an admin view of agent configuration and access.
Evaluation rubrics and golden sets
A written grading standard and a fixed set of test questions, re-run on every change to prompts, models, tools or knowledge sources.
How it works
1.Person asks
A question or task in plain language.
Example: Which RFIs are overdue?
2.Agent
Plans the steps and picks the tools it needs.
3.MCP gateway
Permissions checked per user; every call logged for telemetry.
4.Tools
Procore, P6, accounting, documents — only what's approved.
5.Answer or draft
Answers cite their sources; actions come back as drafts for review.
- 1
Pick one task worth automating
We start with a single, well-defined task, the people who do it today, and the systems and documents it depends on.
- 2
Define access and quality up front
We agree what the agent can read, what it can change, who approves changes, and the rubric and test questions that define a good answer.
- 3
Build the tools and the agent
We build or reuse MCP servers for the systems involved, route them through a gateway, and configure the agent with its instructions and evidence rules.
- 4
Evaluate before launch
We score the agent against the golden set, record a baseline, and fix gaps at the source before users see it.
- 5
Run with telemetry and review
In use, every tool call is logged and quality is re-scored on each change, so you can see what the agent did and whether it is still meeting the bar.
Systems we work with
- Procore
- Oracle Primavera P6
- XER files
- Model Context Protocol (MCP)
- SharePoint and project documents
- Microsoft Azure
- Salesforce
- ArcGIS
How we measure the value
We agree a baseline before we build and measure the same things after go-live. Use the monthly report cost calculator to put your own numbers on it.
| What we measure | How baseline and after are captured |
|---|---|
| Time to answer a project question | We time a sample of real questions answered the current way, then the same kind of questions answered with the agent, including the time to check its citations. |
| Answer quality against the rubric | We score the golden set before launch to set a baseline, then re-score it after every change to prompts, models, tools or knowledge sources. |
| Citation coverage | We track the share of answers that cite a source a reviewer can open and confirm, from the first evaluation run onward. |
| Scheduler review effort on drafts | For schedule agents, we compare the hours and revision rounds to reach an approved schedule before and after the agent drafts the first version. |
| Tool calls, errors and usage | Gateway telemetry records each tool call and its result from day one, so trends are measured against the first weeks of use. |
Use cases
- Built and shown
AI assistant for project drawings, specs and documents
Ask questions of a project's drawings, specs and meeting notes and get answers that cite the file and passage they came from.
- Built and shown
AI schedule builder for CPM schedules with evidence and XER export
Specialized agents draft a CPM schedule from project documents and client answers, show the evidence behind each activity, and export XER for the scheduler to review.
- Built and shown
MCP servers for Procore, Primavera P6 and other construction systems
An MCP server exposes a construction system's API as tools an AI agent can call, with scoped permissions and a log of every call.
- Built and shown
Meeting and job walk transcription tied to the project record
Meetings and job walks are recorded, transcribed, summarized and attached to the project, so what was said and seen becomes searchable project context.
- We design and build this
Invoice and document OCR extraction into Procore and accounting
OCR reads incoming invoices and documents, extracts the fields you need, validates them against your systems and sends anything uncertain to a person.
- Built and shown
AI sales automation: lead capture, research, follow-ups and proposals
Agents handle the research, scheduling and first drafts around each sales step, a person approves anything customer-facing, and the CRM stays up to date.
Proof
Ways to start
Frequently asked questions
What is an MCP server and why does it matter for construction?
The Model Context Protocol (MCP) is an open standard for connecting AI agents to tools and data. An MCP server exposes a system such as Procore or Primavera P6 as a set of tools the agent can call, behind one boundary where access, logging and approvals are enforced. That lets you connect agents to construction systems without giving them open-ended credentials.
Can an AI agent change data in Procore or P6?
Only if you decide it should. We start agents read-only, scope credentials to the least access the task needs, and require human approval for any write. Those rules live in the gateway and server, not just in the prompt.
How do you know the agent's answers are right?
Before launch we write a rubric that defines a good answer and build a golden set of real test questions. We score the agent against that set to record a baseline, and re-run it on every change. Answers also cite their sources, so a reviewer can check them directly.
Do agents replace our schedulers or project managers?
No. In Connect, for example, specialized agents draft the info sheet, the means-and-methods questions and the schedule, and the scheduler reviews every draft. The agent does the searching and drafting; your people make the decisions.
What does a first AI agent project look like?
Most start with the AI readiness review: one task, its data and access, the risks, and an evaluation rubric. If the pilot makes sense, we build it at a fixed scope and price, agreed after discovery.
Where does our data go?
That is agreed during discovery and depends on your environment and model choices. We design for least privilege, keep credentials in the gateway rather than in the agent, and log every tool call so you can review what was accessed.
Next step
Which report or workflow would you like to improve?
Tell us what your team does today, which systems are involved, and what you want to change. We'll discuss whether there is a practical fit.
Prefer email? charley@buildflows.ai