AI Agents Connected to Your CRM and ERP

What an AI agent does between CRM and ERP

The gap between a CRM and an ERP is where a lot of manual work hides. Sales closes a deal in Salesforce, then someone retypes it as a sales order in NetSuite. A customer emails a purchase order as a PDF, and an inside sales rep keys forty lines into the ERP. A rep asks the warehouse whether a part will ship this week because the CRM cannot see inventory. An address changes in one system and not the other.

An AI agent can handle much of that by reading the request, deciding which steps apply, calling the right system for each step and stopping for approval when the rules say so. Tepia designs these agents for specific jobs with clear endpoints, not as open ended assistants. A purchase order agent reads the document, matches the customer and items, checks stock and credit status, drafts the sales order and places it in a review queue. The rep confirms, and the agent posts it and logs a note on the CRM account.

After thirteen years of integration work, Tepia’s view is that most failures in agent projects are integration failures: wrong item mapping, duplicate records, a posting that half succeeded. That is why agent work starts with the data and the APIs, not the prompt.

Agent tasks, systems and autonomy levels

Tepia sets autonomy per task, based on what a mistake would do to your records.

Autonomy can change over time. Tepia usually launches an agent at the draft and approve level, measures how often reviewers change its output, and only then discusses letting it post certain cases on its own, such as repeat orders of identical items for established accounts.

Tool design, permissions and the audit trail

An agent is only as safe as the tools it can call. Tepia gives each agent a short list of narrow, typed tools rather than general API access: find customer by email domain, get item availability by SKU and warehouse, create draft sales order, add CRM note. Each tool validates its inputs, and write tools return a draft identifier rather than posting directly.

• Scoped credentials: each agent has its own integration user in NetSuite, Dynamics 365 or Salesforce with only the roles its tools need.
• Model Context Protocol: where it fits, Tepia exposes tools through MCP servers so the same tools work with OpenAI or Anthropic models and in other assistants your team uses.
• Approvals: review queues live where staff already work, such as a CRM task, a Teams message or an internal portal.
• Audit trail: every run records the input, each tool call with its parameters and result, the model version and the person who approved.

For the integration patterns underneath, see how to connect your CRM and ERP.

Handling failures without corrupting records

ERP systems do not forgive half finished work. If an agent creates a customer but fails to create the order, someone has to notice. Tepia builds agent runs to behave like transactions:

• Idempotency keys on every write, so a retried step does not create a second order.
• Checkpoints that store progress, so a failed run resumes at the failed step instead of starting over.
• Reconciliation jobs that compare CRM and ERP each night and flag records that disagree.
• Escalation to a named owner when the agent is unsure, such as an item it cannot match with confidence or a customer with two possible accounts.
• Rate limits and queues that respect NetSuite concurrency limits and Salesforce API allocations during busy periods.

These patterns come from Tepia’s ERP integration services and middleware development work, where the same failure modes appear with or without AI involved.

How Tepia approaches AI agents for CRM and ERP

1. Discovery: a walkthrough of the manual steps between systems, sample documents and emails, API and permission reviews for each system, and a count of how often each task occurs. Deliverables are an Investigation Summary with an integration map, an Interview Summary from sales, customer service and finance, and User Stories such as an inside rep converting an emailed purchase order.
2. Design: tool definitions, autonomy levels per task, review queue screens and the run history view, documented as wireframes and tool specifications.
3. Development and Testing: tools, the agent loop, review queues and reconciliation. The Alpha runs against sandbox ERP and CRM instances with replayed real requests; the Beta runs live for one team at the draft and approve level. Test plans cover item matching accuracy, retry behavior, permission boundaries and every failure path.
4. Training: sessions for reviewers on approving and correcting drafts, and for system admins on credentials and run logs.
5. Launch: agents enabled task by task, with support reps watching the run history and reconciliation reports during the first weeks.
6. Support: review of correction rates, new tools and tasks, and model upgrades tested against stored runs before release.

Typically Discovery runs 1 to 3 months, development and QA 2 to 6 months, and launch about a month. One well defined task with sandbox access can move faster. See our process.

Vendor agents or a custom agent across systems

CRM and ERP vendors now offer their own agent features, and they work well inside a single platform’s data. If the task lives entirely in Salesforce or entirely in your ERP, starting there is sensible, and Tepia will say so.

A custom agent is worth building when the work crosses systems from different vendors, when documents arrive by email or EDI in formats the vendor tools do not read, when you need approvals and audit records in one place across every system, or when you want to choose and change models without changing platforms. Read about custom AI agent development, browse our work or talk to Tepia about the task your team retypes most.

Frequently asked questions