What is agentic ERP, and what would it look like with Fishbowl?
If you run Fishbowl and keep seeing “agentic ERP” in vendor announcements, it means AI agents that work inside your ERP. They read the data, work out the next step, and carry it out. Usually a person signs off on the result. We build Fishbowl integrations, and for the last few months we’ve been building this kind of connection for Fishbowl ourselves.
What it means
Most ERP AI so far has been an assistant. You ask a question, it answers, and you go do the work. An agent can also do the work, such as drafting a purchase order or matching an invoice, for a person to approve.
The large ERP vendors are all shipping this now:
- SAP announced more than 200 agents across its suite at Sapphire 2026.
- Microsoft lets Dynamics 365 agents write data as well as read it, and its Finance agents cover reconciliation, variance analysis, and collections.
- Oracle claims hundreds of embedded agents across Fusion, and NetSuite has its SuiteAgents framework.
Gartner predicts that 40% of enterprise applications will include task-specific AI agents by the end of 2026, up from less than 5% in 2025.
Most of the connections are being built on MCP (Model Context Protocol), an open standard that lets AI assistants like ChatGPT and Claude use another system’s tools. Microsoft documents MCP for Dynamics 365, and smaller ERPs such as Xentral and bexio ship MCP servers too.
Where it goes wrong
Gartner also predicts that more than 40% of agentic AI projects will be canceled by the end of 2027. The reasons it gives are cost, unclear business value, and weak risk controls. It also estimates that only about 130 of the thousands of vendors selling “agentic AI” offer real agent capabilities.
The risk controls are what we think about most. If an agent learns your business rules from a written instruction document, it can misread one or skip it. Approval limits, preferred vendors, and price floors should be written into the tool the agent calls, so the code checks them every time.
On Fishbowl today
We built ILC MCP for Fishbowl to connect ChatGPT or Claude to a live Fishbowl database. Today it is read-only:
- Each person signs in with their own Fishbowl login.
- Admins approve the queries other users can run, and can grant them from the chat.
- Every query is logged with what the person was trying to find out.
You can watch real questions answered on our demo Fishbowl, like “Can we build 40 batches by Friday, and what’s short?”
Reading is as far as it goes today. The next step is letting the same connection save transactions in Fishbowl, starting with the jobs nobody builds a tool for.
One-off jobs as conversations
If your team gets the same file every week, such as a retailer’s order CSV, you end up retyping or reformatting it every time. Once that gets old, a converter that turns it into a Fishbowl import is worth building, and that is a normal integration project.
The jobs that never get a converter are the one-offs: an order list from a trade show, a new customer’s order in their own spreadsheet format, a vendor’s list of price corrections. Each one comes up once, so it isn’t worth building a tool. Someone on your team reworks the data in Excel or types it into Fishbowl line by line. An agent can turn that data into transactions by conversation, and someone on your team approves each one.
Spreadsheet orders
A new customer emails their first order as a spreadsheet in their own format. You paste it into the chat and ask for a sales order. The agent matches every customer, part, and unit of measure to Fishbowl, flags the one part number it can’t find, and shows a preview. You fix the part and say yes, and the order is saved as an estimate under your own Fishbowl login. This runs on our demo system now and is not yet available to customers.
Shortage to PO
You’ve asked whether you can build 40 batches by Friday, and the agent has found that the Main warehouse has no butter. You ask it to create a bid request to your usual butter vendor. It drafts the purchase order with the vendor’s last cost and waits for you. This is planned.
Price list check
A vendor sends a new price list. You upload the file and ask which costs went up more than 5% against what you last paid. This is read-only, so it works today: the agent reads the file and queries your purchase history.
What has to be true
For any of this to be safe on a live ERP, a few things have to hold:
- A person approves a preview before anything goes into Fishbowl.
- The agent works under the person’s own login and permissions.
- Business rules are enforced by the tool’s code.
- Every step is logged, so you can see why the agent did what it did.
We don’t know yet which of these jobs businesses running Fishbowl will trust to an agent first. Our guess is the spreadsheet order. If you’ve had a one-off job like this that ended up in Excel or typed in by hand, we’d like to hear what it was.
Related: