Governance
The rules our AI runs under.
AI on a construction project is a liability if it acts without accountability. So we wrote the rules down. Six pillars, a kill switch, an audit log, and a no-training commitment in writing. This page summarizes the operational Constitution our agents are bound to.
Why this exists.
Most "AI for construction" products bolt an LLM onto an existing tool and call it intelligence. The result is software that does things in your name without telling you what it did or why.
Construction doesn't tolerate that. Contracts get signed. Money moves. Subs walk off. Disputes end up in litigation. If an AI on the project gets the change-order wrong — or sends an email nobody asked it to send — somebody real is left holding the consequences.
The Agent Governance Constitution names what our AI is allowed to do, what it must escalate to a human, and how every automated decision is logged. It is binding on us — meaning we cannot ship an agent that violates it, and amendments require version-bump and public notice.
The six pillars.
Plain-English summary of the operational pillars in Constitution v1.2.
Bounded scope
Every action traces back to an explicit grant.
Each AI agent operates inside a permission matrix that names — in writing — which data it can read and which it can write. Anything outside that grant is denied by default. Agents cannot escalate their own privileges or chain through other agents to reach data they weren't granted.
Schema-first reasoning
Deterministic parsing before AI inference.
Before any agent uses an LLM to guess what something is, it must try to resolve the task through structured, deterministic parsing. AI is the fallback. The first answer is the answer that doesn't hallucinate.
Grounded output
Every claim shows its source.
When an agent drafts a change order, summarizes a meeting, or flags risk, the output points back to the exact inputs it used. You can audit any AI claim down to the document, message, or record it came from. No black-box outputs.
External quality gates
A separate validator checks every agent output.
Agent outputs run through a separate validator chain before they can be committed or sent. The validator is a different system, with different code paths, that checks each output against the rules of the destination (file naming, schema conformance, scope adherence, escalation thresholds). One agent failing does not contaminate the rest.
Single-responsibility agents
One agent, one job.
We don't build one mega-agent that 'handles construction.' Each agent has a narrow scope — RFI routing, daily reports, change-order drafting, safety monitoring. When something breaks, you know which agent owns it. When something needs to change, you can change one agent without rewriting the system.
Cheapest viable path
Use the smallest tool that solves the problem.
If a regex resolves it, we use a regex. If a deterministic parser resolves it, we use a parser. We reach for an LLM only when it's the right tool — never because it's the trendy one. This keeps costs predictable, latency low, and behavior easier to audit.
Public commitments.
The pillars are how the system is built. The commitments below are what you, as a user, can hold us to.
Drafts, not actions.
No AI-generated change order, email, or schedule update is binding or sent until a human with the appropriate authority approves it. The agent proposes. A person decides.
Audit log on every decision.
Every flag raised, every draft produced, every automated message: logged with who, when, which inputs, and which rule fired. Audit logs are immutable and exportable.
Kill switch on every agent.
Workspace owners can disable any individual AI watcher — or all of them — at any time. The product continues to work without AI. Removing the AI does not lock you out of your own data.
No training on your data.
We contract with our LLM providers under zero-data-retention terms. Your project data is not used to train models — ours or theirs. This is in writing, with both Resend (email) and our LLM provider(s).
Versioned in public.
Material changes to the Constitution come with a version bump, a dated change log, and at least thirty days' notice to active users before they take effect. Past versions stay public.
Emergency kill switch.
Every project workspace has an emergency kill switch that disables all AI agents on the project, immediately. The button is on the workspace settings page, behind a single confirmation prompt. Workspace owners can hit it without contacting support.
When the kill switch is engaged: no new AI drafts, no new flags, no automated emails, no background analysis. Everything else — budget tracking, document storage, project chat, QuickBooks sync — continues to work. The kill switch is reversible by the same workspace owner.
We also operate a global kill switch on our end. If we detect an active incident — model provider outage, a bug producing wrong outputs at scale, a security event — we can disable all AI agents across all workspaces in under sixty seconds. We will notify affected users within fifteen minutes.
Read the Constitution.
The full Constitution is ~6,000 words and includes the Permission Matrix, the Escalation Matrix, and the per-agent rules. During the Atlanta cohort we share the full document directly. After general availability it will be published openly at this URL.
Current version
v1.2 · April 20, 2026
Scope: binding for all production agents — RFI Router, Daily Report, Change Order, Safety Monitor, Bid Assist, and the CEO governance agent.
Status: binding. All agent system prompts must reference or embed this document.
Validator: a separate validator chain enforces these rules at runtime. Validator source is open to cohort participants on request.
Change log.
Material changes to the Constitution come with a version bump and at least thirty days' notice to active users.
Initial published version. Six pillars, Emergency Kill Switch, Permission Matrix, Escalation Matrix.
Adds rules for external tool invocations via Composio. Under review.
Questions about governance?
We treat this as a serious part of the product. Email governance@monoboard.ai and the founder will respond directly. Lawyers, security teams, and risk officers especially welcome.