Sales and customer service
Order status, delivery dates, prices and stock while the customer is on the phone.
AI / ERP Agents
AI agent for ERP: answers on orders, stock and invoices for your staff
Sales, service and finance staff wait for someone who knows the ERP screens to look up an order, a stock level or an unpaid invoice. Agnotic Technologies builds an AI agent that answers those questions in plain language from your live ERP data, read-only by default, with any write-back limited to the actions you approve.
Plain-language questionsLive ERP dataRead-only by defaultAnswers with record links
Trusted by global innovators
What is an AI agent for ERP?
An AI agent for ERP is an assistant that staff ask questions in plain language, such as where an order is, how much stock is free or which invoices are overdue, and that answers by calling your ERP's own APIs and quoting the records it found. Agnotic Technologies builds it for your ERP and your teams: a chat in Microsoft Teams, Slack or your intranet, lookup tools for the questions people really ask, answers limited to what each person may see in the ERP, and links back to the source records. We start read-only, and add record creation only for specific actions where a person approves each change.
Staff questions
Can an AI agent look up orders, stock and invoices in our ERP and answer staff questions?
Yes. The agent turns a question into one or more lookups in your ERP, such as an order search, a stock query by item and warehouse or a list of open invoices for a customer, and replies with the answer and links to the records. Agnotic Technologies builds one lookup tool per kind of question, so the agent runs tested queries instead of writing its own.
Answers come from live data at the moment of the question, and each person only sees what their ERP role allows. When the agent cannot find a record or the question is unclear, it says so and asks, rather than guessing.
Read or write
Should the agent only read from the ERP, or also create records?
Start read-only. Let the agent create records only for a few well-defined actions where a wrong entry is easy to spot and reverse, and keep a person approving each one. This is the decision table we work through with your process owners.
| Staff request | Our usual advice | Why |
|---|---|---|
| Look up orders, stock, invoices or prices | Read-only | Most of the value, with no risk to ERP data |
| Summarize a customer's open orders and balance | Read-only | Combines several lookups; nothing is changed |
| Draft a purchase requisition or a quote | Create a draft for a person to review and post | Saves typing while a person keeps the decision |
| Log a customer call or add a note to an order | Write, limited to that field | Low impact and easy to correct |
| Change a price, a credit limit or a delivery date | Not through the agent at first | High impact and hard to notice when wrong |
| Post invoices, ship orders or delete records | Not through the agent | Financial and legal effects belong in the ERP's own approval flow |
Every write the agent makes is limited to its own tool, runs under an ERP account that can do nothing else, and is logged with the person who asked for it.
Who asks
Who uses an ERP agent day to day?
The teams that ask the ERP the most questions without being ERP experts.
Warehouse and logistics
Stock by location, open picks and inbound purchase orders.
Finance
Open invoices, payment status and customer balances.
Purchasing
Supplier history, last prices and late purchase orders.
One question
How does the agent answer a question?
The path a staff question takes, from the chat window to the ERP and back.
Ask
A staff member asks in Teams, Slack or the intranet, signed in as themselves.
Understand
The model works out what is being asked and which lookup fits, and asks back if something is missing.
Look up
A tested lookup tool queries the ERP API with the person's own access rights.
Answer
The agent replies with the figures it found and links to the ERP records.
Log
The question, the lookups run and the records returned are logged.
Improve
Unanswered or corrected questions are reviewed to add or fix lookup tools.
Failure points
Where do ERP agent projects go wrong?
Problems we design out from the first release.
Challenge
The agent writes its own database queries.
Agnotic approach
Tested lookup tools, one per kind of question, with fixed fields and limits.
Challenge
Everyone sees every customer and price.
Agnotic approach
Answers follow each person's own ERP access rights.
Challenge
Answers sound right but cannot be checked.
Agnotic approach
Every answer links to the ERP records it came from.
Challenge
Write access is added to the first release.
Agnotic approach
Read-only first; writes added one action at a time, with approval.
Challenge
Stock and order numbers are stale.
Agnotic approach
Lookups query live data at the moment of the question.
What do we build into an ERP agent?
The assistant, the lookup tools and the controls, handed over as code you own.
Chat where staff work
Microsoft Teams, Slack, your intranet or a panel inside your own apps.
Lookup tools
Order, stock, invoice, customer and supplier lookups through the ERP's APIs.
Question understanding
A language model that maps questions to lookups and asks back when unclear.
Access by person
Results limited to what each user may see in the ERP.
Approved actions
Drafts and narrow writes, each held for a person to approve.
Record links
Every answer points to the source records in the ERP.
Question log
Questions, lookups and results logged for review and audit.
Answer testing
A set of real staff questions with known answers, rerun on every change.
Related builds
Where does an ERP agent connect?
The agent is the business layer; these pages cover the connection and the controls.
MCP server development services
The protocol layer that exposes ERP lookups to this and other AI clients as scoped tools.
Read-only vs write-back AI agents
The wider decision on when agents should only read and when they may write.
Sales order automation
Customer purchase orders read and entered into the ERP for review.
Custom ERP development
New ERP modules and APIs, where the lookups need data the ERP does not expose.
Facts and sources
What do OWASP and ERP vendor documentation say about agent access?
The LLM security risk behind the read-only default, and how ERP APIs expose read and write, each linked to its source.
The OWASP Top 10 for LLM Applications 2025 lists Excessive Agency (LLM06): the vulnerability that lets damaging actions be performed in response to unexpected, ambiguous or manipulated outputs from an LLM, typically caused by excessive functionality, excessive permissions or excessive autonomy.
Source: OWASP Gen AI Security Project, LLM06:2025 Excessive AgencyOWASP's Excessive Agency guidance gives an example where the damage an LLM extension could do is reduced by giving it a read-only scope, and by requiring the user to review and confirm each message it drafts before it is sent.
Source: OWASP Gen AI Security Project, LLM06:2025 Excessive AgencyMicrosoft's Business Central API v2.0 salesOrder resource supports GET, POST, PATCH and DELETE, plus a bound shipAndInvoice action that ships and invoices the order, so the same API an agent reads through can also create, change and post orders if its credentials allow it.
Source: Microsoft Learn, Business Central API v2.0 salesOrder resource typeMicrosoft describes OData, the Open Data Protocol, as an ISO/IEC approved OASIS standard that defines a set of best practices for building and consuming REST APIs.
Source: Microsoft Learn, OData overview
Standards
Standards and interfaces we build ERP agents on
OData v4OWASP LLM Top 10OpenID ConnectOAuth 2.0MCP
Scope
What decides the size of an ERP agent build?
We do not publish prices. These are the drivers we size on the scoping call.
ERP and APIs
Your ERP, its version and the APIs it offers for the data staff ask about.
Questions
How many kinds of question the first release answers.
Access rules
How ERP roles map to what each person sees.
Chat channel
Teams, Slack, intranet or your own application.
Write actions
Whether any drafts or writes are in scope, and who approves them.
Model and hosting
Which model provider, and where the agent runs.
Engagement
How does an ERP agent project run?
We start from the questions staff ask most.
1.
Scoping call
We collect real staff questions and check which ERP APIs can answer them.
Scope2.
First release: read-only for one team
A fixed-scope first release: lookup tools, chat, access by person and the question log for one team.
Build3.
More teams and approved actions
More questions and teams, then drafts and narrow writes with approval, with support after launch.
ScaleFrequently Asked Questions
ERPs that offer APIs for the data staff ask about, such as SAP, Oracle, Microsoft Dynamics 365, NetSuite and custom ERPs. Where the ERP has no suitable API, a read replica or reporting database can be used for lookups.
Language models can produce wrong text, so the agent answers only from lookup results, links each answer to its records and says when it found nothing. Answers are tested against a set of real questions before each release.
The lookup results needed to answer a question are sent to the model. We choose the provider and its data terms with you, and the ERP itself is only reached through the lookup tools.
This page covers the assistant staff use and the decisions about what it answers and may change. An MCP server is the connection layer that can expose the same ERP lookups to this agent and to other AI clients.
We do not publish prices. The ERP and its APIs, the range of questions, access rules, the chat channel, any write actions and hosting decide the size, and we agree a fixed first release on the scoping call.