# ERP Agent API Integration Readiness Assessment > Source: https://netray.co/tools/erp-agent-api-integration-readiness-assessment > Publisher: Netray (netray.co), AI and integration engineering for ERP and manufacturing systems. > Section: Free Tools ## Short answer Answer ten questions about APIs, roles, limits and logging to see whether your ERP is ready for an AI agent or MCP integration. ## Category developer ## Audience - CIO - ERP architect - AI engineer - Integration developer ## Tool **Kind:** assessment **Questions:** #### How does your ERP expose data and transactions programmatically? **Options:** ##### No supported API, only direct database or screen access **Score:** 0 ##### Limited or legacy interfaces (flat files, old SOAP) **Score:** 1 ##### Documented REST, OData or SOAP APIs for the main objects **Score:** 3 ##### Documented, versioned APIs with sandbox and change notices **Score:** 4 #### How is a service account for the agent set up? **Options:** ##### Would reuse a shared administrator login **Score:** 0 ##### Would reuse a named person's login **Score:** 1 ##### Dedicated account, but broad permissions **Score:** 2 ##### Dedicated account with least-privilege, read-only role **Score:** 4 #### Can the agent act with the permissions of the asking user, or at least be limited by them? **Options:** ##### No, everyone shares one identity **Score:** 0 ##### Partly, enforced in the agent only **Score:** 2 ##### Yes for reads, using the user's ERP role **Score:** 3 ##### Yes for reads and writes, with ERP-side enforcement **Score:** 4 #### Do you know your API rate limits and capacity? **Options:** ##### No idea **Score:** 0 ##### Heard there are limits, not measured **Score:** 1 ##### Documented limits, no monitoring **Score:** 2 ##### Documented limits with monitoring and alerts **Score:** 4 #### Is there a non-production ERP environment with representative data? **Options:** ##### No **Score:** 0 ##### Yes but stale or with no realistic data **Score:** 1 ##### Yes, refreshed occasionally **Score:** 3 ##### Yes, refreshed regularly with masked data **Score:** 4 #### How will writes from the agent be controlled? **Options:** ##### Not decided **Score:** 0 ##### Agent writes directly **Score:** 0 ##### Reads only for now **Score:** 3 ##### Human approval before any write, with a log **Score:** 4 #### What logging will record agent actions? **Options:** ##### None planned **Score:** 0 ##### Application logs only **Score:** 1 ##### Request and response logs kept centrally **Score:** 3 ##### Tamper-resistant audit log with user, tool, arguments and result **Score:** 4 #### Where may the ERP data be processed by a language model? **Options:** ##### Not yet decided or no policy **Score:** 0 ##### Public hosted models with no review **Score:** 1 ##### Approved vendor with a data agreement **Score:** 3 ##### Private or on-premises model for sensitive data **Score:** 4 #### How are secrets and tokens handled? **Options:** ##### In code or shared documents **Score:** 0 ##### In configuration files on a server **Score:** 1 ##### Environment variables or a secret store **Score:** 3 ##### Secret manager with rotation and scoped tokens **Score:** 4 #### Is there an agreed test set to check the agent answers correctly before launch? **Options:** ##### No **Score:** 0 ##### A few ad hoc questions **Score:** 1 ##### A written list of questions with expected answers **Score:** 3 ##### Automated evaluation re-run on each change **Score:** 4 **Bands:** #### Foundations needed **Min:** 0 **Max:** 30 **Verdict:** Connecting an agent now would carry real risk. Fix access, identity and logging first. **Recommendations:** - Create a dedicated service identity with a read-only role - Identify the supported API for each object you need - Decide the data policy for language models before any pilot - Set up a non-production environment to test against #### Pilot with tight limits **Min:** 31 **Max:** 60 **Verdict:** A small read-only pilot is reasonable if you close the biggest gaps in parallel. **Recommendations:** - Limit the pilot to a few narrow read tools - Measure API usage and agree a request budget - Add central logging of every tool call - Write 20 to 30 test questions with expected answers #### Ready for a controlled rollout **Min:** 61 **Max:** 85 **Verdict:** Most controls exist. Focus on write approval, monitoring and evaluation. **Recommendations:** - Add human approval for any write tool - Alert on error rates and rate limit responses - Automate evaluation runs on each prompt or tool change - Review role scope quarterly #### Well prepared **Min:** 86 **Max:** 100 **Verdict:** Your integration foundations are strong. Keep them current as tools are added. **Recommendations:** - Re-run this assessment whenever you add a write tool - Rotate secrets and review audit logs on a schedule - Test failure behavior with deliberate API outages - Document the tool catalog and its owners ## Content **Intro:** Most failed AI agent projects on ERP data do not fail on the model. They fail on identity, limits, logging and unclear write rules. This ten question assessment scores those foundations so you know what to fix before a pilot. **Sections:** #### How to use it **Body:** Answer for the system and environment you actually plan to connect, not the one in the vendor brochure. Involve the ERP administrator and someone from security, since a few questions need their input. The score is converted to a percentage and mapped to one of four bands, each with concrete next steps. Export the result to share in a project kickoff or a security review. #### What the questions cover **Body:** The questions cluster in four areas. Access: API availability, the service account and whether user permissions carry through. Capacity: known rate limits and a safe test environment. Control: write approval, logging and secret handling. Quality: where models may process data and how answers will be tested. A low score in access or control matters more than a low score in testing, because those gaps cause incidents. **Bullets:** - Access: supported API, least-privilege identity - Capacity: limits known and monitored - Control: approval, audit log, secret store - Quality: evaluation set and data policy #### How to read the result **Body:** A band is a starting point, not a verdict on your team. A low overall score with strong access controls can still support a narrow read-only pilot. A high score with no write approval should not enable write tools. Look at the individual answers marked zero or one and fix those first. Re-run the assessment after each change to see progress. #### Common mistakes **Body:** Teams often score the intended design instead of what is configured today, assume the ERP vendor's AI features solve identity and logging, or skip the evaluation set until after users complain. Some also connect the agent with an administrator account to save time during a demo. Answer honestly and use the gaps as a work list. **Cta:** Netray helps teams close these gaps and run ERP agents on private infrastructure with read-only defaults and approved write-back. Ask us for a short readiness review if you want a second pair of eyes. **Faq:** #### Do I need an API before I can use an AI agent with my ERP? **Answer:** Not strictly, since agents can use read-only database views or reports, but a supported API with a clear role model is easier to secure and maintain. Direct database access needs a read-only login on approved views and careful handling of upgrades. #### Should the agent have write access from day one? **Answer:** We recommend starting read-only. Add write tools one at a time, behind human approval and with a log, once the read path is accurate and monitored. Writes multiply the impact of any wrong answer. #### How many test questions do I need before launch? **Answer:** A written set of 20 to 50 realistic questions with known correct answers is a workable start. Include edge cases such as missing data and permission denials, and re-run the set whenever tools, prompts or the model change. #### Does a private model change the readiness score? **Answer:** It helps the data policy answer, because sensitive ERP data stays inside your environment. It does not replace identity, rate limits, logging or write approval, which matter no matter where the model runs. ## Search terms erp api readiness assessment, ai agent erp integration checklist, is my erp ready for ai agents, mcp erp readiness, erp api maturity, agent integration prerequisites, erp ai integration assessment, erp api security readiness --- Talk to an engineer: https://netray.co/contact?intent=erp_ai | Machine catalog: https://netray.co/agent/catalog.json