Skip to main content

AI agent architecture

Applications propose. Govern decides. Run carries.

Applications on top. Govern in the middle, deciding every action. Run underneath. One path for every consequential action, whichever platform it starts in.

See a governed action
See the platforms

How do governed AI agents work?

Governed AI agents propose actions instead of taking them. In elan1's design, an application such as elan1 Work or elan1 Decide proposes the action, elan1 Govern checks the agent's identity and your rules, then allows it, asks a named person or refuses it, and elan1 Run carries only what was decided, leaving a record your auditor can verify.

The layer model

Three layers. One path to your records.

Work and Decide propose. Govern and Run assure. Only Govern allows.

Propose

Applications

elan1 WorkAgentic enterprise automation for ERP and CRMelan1 DecideDecision intelligencecommerce1 (opens commerce1.ai, external site)Connected platform · commerce1.ai

Decide

Governanceelan1 GovernAllow · ask a person · refuse

Carry

Runtimeelan1 RunIn your country, on your cloud

Ends here

Your systems of record

ERP · CRM · core banking · claims

Every consequential action ends here.

Applications on top. Govern in the middle. Run underneath.

Every consequential action ends in your systems of record, and only after elan1 Govern has decided. Agents never hold the live keys.

  • Applications

    Where agents do the work, in your ERP, CRM, claims, alerts and storefronts.

  • Govern

    The single path every consequential action takes. Allow, ask a person, or refuse.

  • Run

    The runtime underneath, in your country, on parts you can replace.

Layer by layer

What each layer does, and what it never does.

Hands sort through printed financial forms beside a calculator and pen

Applications

Where agents do the work.

elan1 Work changes your ERP and CRM only when your books still hold. elan1 Decide makes everyday decisions in milliseconds, with a person on the hard ones. Applications propose; none of them decides on its own.

A man in a navy shirt signs off a paper document at a dark desk, seen from above

Governance

The single path every action takes.

Every consequential action passes through elan1 Govern before it runs. Your rules are checked first. Routine actions pass, risky ones wait for a named person, and the rest are refused. Every outcome leaves a record your auditor can check offline.

“A model may propose an action. It may never approve one.”

elan1 Govern · Agent governance and assurance

Network cables patched into a rack panel, close up in neutral light

Runtime

The runtime underneath, in your country.

elan1 Run carries, routes, meters and records. It runs in India, on your cloud or an operator's. Your apps talk to 10 stable capability contracts rather than to any one tool, so parts can be replaced without rebuilding them.

Each team's budget is reserved before the first token is used, so spending stops at the limit, not on the invoice.

elan1 Run · Sovereign agent infrastructure

The path of one action

Ask, check, approve, act, record. Every time.

The same five steps, whichever application proposes the action.

  1. 01

    Ask

    Applicationselan1 Work or elan1 Decide proposes an action instead of taking it.

  2. 02

    Check

    GovernGovern confirms who the agent is and reads your rules.

  3. 03

    Approve

    GovernRoutine actions pass. Risky ones wait for a named person. The rest are refused.

  4. 04

    Act

    GovernOnly what was decided runs, with credentials the agent never sees.

  5. 05

    Record

    Govern and RunEvery outcome is written to a record your auditor can verify.
Follow one purchase bill through all five

Rules of the stack

Four rules the layers never break.

  • Work and Decide propose.

    Applications do the work and make the judgement. Neither decides on its own whether a record changes.

  • Govern and Run assure.

    Govern decides what happens. Run carries it, in your country, on parts you can replace.

  • Only Govern allows.

    No product builds its own approvals or records. One approval model, one identity, one record.

  • Fail toward a human, never toward an action.

    If anything is unsure or unavailable, the work waits for a person. An error refuses.

Works with what you run

Starts on the systems you already use.

Tally and Zoho Books can be read for migration, never written to. We announce dates when they are firm.

Read the claims register
Systems of record and where each fits
System
Frappe/ERPNextERP · elan1 Work starts here
Twenty CRMCRM · elan1 Work starts here
Zoho CRMCRM · next for elan1 Work
SAP and SalesforceERP and CRM · on the roadmap

Questions

How the layers work in practice

Do we need all four platforms?
No. The 4 platforms are built to work alone or together. Start with the platform your first process needs, and add the next when it holds. Whatever you start with, consequential actions pass through elan1 Govern, so the approval model, identity and record stay the same as you grow.
Which agents and systems does elan1 Govern work with?
elan1 Govern works with agents built on Claude, the OpenAI Agents SDK, LangChain, Cursor or your own code, and with tools connected through standard agent connectors. Gmail and Zoho Books are built in today. The agent proposes; Govern decides before anything reaches the connected system.
What happens if elan1 Govern is unavailable?
Actions stop; they do not slip through. elan1 is designed to fail towards a human, never towards an action: an error refuses, and the work waits for a person. Nothing reaches your systems of record until elan1 Govern is available again and has decided.
Where does it all run?
In your cloud account in India, or in an operator's data centre, in the Mumbai region. Customer data is hosted in AWS Mumbai, with no US region. On-premises and air-gapped options, and Gulf and EU regions, are on the roadmap; new regions open only on a signed contract.
Do AI agents hold the live credentials to our systems?
No. Agents never hold the live keys. Only what elan1 Govern has decided runs, with credentials the agent never sees, so an agent cannot widen what was approved: not a different amount, not a different supplier, not another record. Every outcome is written to a record your auditor can verify.

Bring one process. We'll show you the path.

A working session with the people who build the platforms, on a process you choose.

Contact usTake the readiness check

Intelligence you can answer for.