Why one governed agent instead of one agent per task?
A list of agents, one for each job, is an easy thing to market and a hard thing to govern. Each one needs its own access, its own idea of the project and its own rules, and the rules end up living in whatever wrapper the agent was built in.
Crolle puts the governance in a different place. The ERP owns the chain end to end, and its rules sit beneath the tools the agent calls, not around the agent. So the payment certificate work, the drawing revision work and the tender work below are the same agent, calling the same four tools, under the rights of the person it is working for. A new workflow does not need a new set of permissions, because it inherits the ones that already apply to everyone.
Which workflows is the agent designed to handle?
Each workflow below has its own page. In every one, the agent reads and prepares, and a named role decides.
| Workflow | What the agent reads | What it prepares and stages | Who commits |
|---|---|---|---|
| Subcontractor IPC review | Measured work, variations, subcontract terms, earlier certificates | A draft certificate with the evidence and the differences from the claim | The quantity surveyor reviews; an authorised person approves |
| Drawing revision impact | The revised drawing, the BOQ item, the rate basis, orders and the programme | The quantity change, the cost movement and the affected packages and dates | The commercial manager, with engineering, procurement and planning |
| Tender BOQ review | The tender BOQ, drawings, specifications, addenda and the rate basis | Missing items, rate gaps, assumptions and draft clarification questions | The estimator and the bid approver |
| Retention release | Retention held on each certificate, completion records, defects and dates | A release schedule and, when the conditions are met, a draft release | The quantity surveyor and an authorised approver |
What can the agent never do?
The limits below do not depend on how the agent is prompted, because they are enforced by the ERP, not requested of the model.
- Commit a change. It stages one change at a time and reports what happened to it. An authorised person commits.
- Reach past the person it works for. It finds only the capabilities that person is authorised to use. What is missing is listed with the right that is needed and who grants it.
- Invent records. In a workspace that is not set up, an operation whose records are missing is refused by name rather than guessed at.
- Act without a record. What was proposed, who decided and on what evidence stays in the audit trail.
The full set of controls is on the security and AI governance page.
How does the agent reach the ERP?
Through four tools published over the Model Context Protocol: one to find what the person may do, one to read the exact contract of a capability, one to run a read, and one to stage a change. Any agent that speaks MCP reaches Crolle the same way and under the same rules. Crolle's ERP over MCP describes the four tools in detail, and what is an agentic ERP? explains why the tools, not a screen, are the primary interface.
Who is the agent designed for?
Main contractors of every size in Egypt and the Gulf. The work it is designed around is the work those contractors run every month: bills of quantities, interim payment certificates, retention, variations and notices under FIDIC-style contracts, with tender documents and correspondence in Arabic and English.
Size changes how much coordination the agent absorbs, not which rules apply. One construction ERP for main contractors of every size covers what changes between a small contractor and a large one, and one ERP or separate tools? covers why the agent needs one record to stand on.
Common questions
Is each workflow a separate AI agent?
No. Crolle is designed around one governed agent that calls the same four ERP tools for every workflow. The payment certificate, drawing revision, tender BOQ and retention work differ in what the agent reads and prepares, not in what it is allowed to do, because the permissions belong to the person it works for.
Can the agent approve or commit anything on its own?
No. The agent can stage one change at a time and report what happened to it. An authorised person reviews the evidence and commits the change. Delegated authority, human approval and an audit trail apply to every workflow on this page.
Which contracts and documents is the agent designed for?
It is designed for the documents main contractors in Egypt and the Gulf work with every month: bills of quantities, interim payment certificates, subcontracts with retention and advance recovery, and variations and notices under FIDIC-style contracts, in Arabic and English.
Can we use our own AI assistant instead of Crolle's?
That is what the MCP interface is for. Any agent that speaks the Model Context Protocol can be given access to Crolle's four tools, and it works under the same rights, the same staging and the same audit trail as Crolle's own agent.
Crolle is opening early access to a small group of contractors and partners. To see it on your own projects, write to info@crolle.ai.