Concerns and Musings about Ai

Kinetic ERP
1

Anything in ERP that is not deterministic, without at least a big flashing warning, is an issue.

2

Completely agree.

For ERP, anything transactional needs to be deterministic, auditable, and explainable. If AI is involved, there needs to be a very clear distinction between:

  1. retrieving/summarizing information

  2. recommending an action

  3. actually executing an action

I’m much more comfortable with AI in the first two categories today.

For example, I’m fine with an AI layer saying, “Here are the late jobs,” “Here are the open POs for this supplier,” or “This part appears short based on current demand and on-hand quantity.”

But once it starts creating jobs, adjusting inventory, changing costing, releasing shipments, modifying orders, or touching financial transactions, that should require deterministic logic, strict guardrails, permissions, audit trails, and probably human approval unless the workflow is extremely well-controlled.

3

The other thing that may be concerning about AI is when we build an agent and assume everyone follows best practices.

While I was working on a Check Register Dashboard, I found out there are like 5+ ways people are voiding checks!

  1. Some VOID
  2. Some Create a Negative Invoice and write a Negative Check
  3. Some Cancel the Invoice, not the Check (I cant recall)
  4. Some add a Misc Check with the Negative etc..
  5. Some may just do Journal Adjustment

etc… I have yet to find someone adhering to the ANSI EDI Standard. :smiley: (that means your UOMs should not be more than 2 chars)

4

5

AI is NOT deterministic so it WILL make mistakes. That means there MUST be a person reviewing what the AI has produced and that person is responsible for approving or rejecting it. Always a human in the loop.

AI has completely transformed software development in 2 years because its a good fit, LLM’s produce text (code is text) and there are already well established review process for code changes that the AI can slot into.

I think of AI as an amplifier, it will increase the the amount of work some one can do. If that someone is skilled => more good work is done, if that person is not skilled => more mistakes that need to be fixed.

6

in Hours, how long before people get Lazy.

watch it every day :wink:

7

This. Yeah absolutely. AI can be a force multiplier in the right hands… or it can be a complete waste of time in the wrong hands. haha truer words…

8

I do think there are definitely many use cases for agentic workflows where you would not necessarily need a human in the loop after the workflow has been properly tested, validated, and governed.

For example, let’s say we ship 1,000 parts at $0 because they are consignment. At that point, the parts are no longer in regular inventory, but in order to maintain traceability, we may want to automatically quantity-adjust them into a non-nettable Consignment warehouse.

Once that workflow is built and validated — for example, “when a $0 shipment is made for this product group to this customer, automatically quantity-adjust the shipped quantity into this consignment warehouse/bin/lot” — I do not think that would necessarily require a human approval every time. At that point, it is not really AI making a judgment call. It is deterministic automation triggered by a defined business event.

That does not mean it should be uncontrolled. There should still be an audit log, exception handling, review capability, and probably alerts for failures or unusual conditions.

To me, anything related to an agentic ERP workflow is probably going to flow out of some kind of trigger event: “If this happens, then do this.” I am not sure exactly how Epicor plans to implement this new agentic functionality, but I would assume it will look somewhat similar to what we have seen elsewhere: node-based workflow building on a canvas, similar in concept to n8n, LangChain-style orchestration, or other workflow builders.

I am okay with that, honestly. I think that could be useful.

But I also agree there are many cases where you absolutely do need a human in the loop. In those situations, I would want the workflow to run, gather the facts, propose an action, and then notify the right person for approval before execution.

So in my mind, there are a few different categories:

  1. Data gathering / read-only analysis
    “Go get me this data and display it how I want.”
    This can be autonomous at the prompt level because it is not changing the system.

  2. Deterministic trigger-based automation
    “When this happens, do this.”
    This can potentially run without human approval after it has been tested and validated, because the logic is defined and repeatable.

  3. AI-assisted decision support with approval
    “Keep an eye out for when X happens, analyze the situation, suggest a solution or action, and notify someone for review before anything is executed.”
    This is where the AI can be useful, but the final approval still belongs to a human.

That is the distinction I am trying to make. Not every automated workflow needs a human approving every step. But anything where the AI is interpreting, recommending, or making a judgment call should have a review/approval layer before it affects inventory, costing, accounting, quality, fulfillment, or customer commitments.

I am curious to see what this looks like inside the Epicor ecosystem. Maybe someone here has a better idea of how Epicor is actually planning to separate those different levels of automation and approval.

9

But that’s not even AI anymore. That’s just an automation. And that’s most of the peoples point here. AI would be great to help you MAKE the automation tool, but once it’s made, I don’t want to rely on AI anymore because AI won’t do it the same way twice, I want to use a traditional computer program that does it the same way every time based on a very specific, pre-determined set of rules. There’s a right tool for the job problem here, and large corporations putting AI into literally EVERYTHING is a waste of time and money.

10

Yeah, I agree with that.

In the consignment example, the final deployed process would absolutely be traditional deterministic automation, not AI making a live decision every time a shipment happens.

That is actually the distinction I am trying to make.

I do not want AI sitting in the middle of every ERP transaction deciding what to do. For something like a $0 consignment shipment adjustment, once the rules are defined, tested, validated, and governed, it should just execute the same way every time based on those rules.

Where I see AI being useful is in helping build, document, troubleshoot, or improve that automation faster.

So the useful pattern to me is:

AI helps create the tool.

The deterministic tool runs the process.

Humans review, test, validate, and govern what gets deployed.

I agree there is absolutely a right-tool-for-the-job problem here. Not everything needs AI in the loop, and forcing AI into places where normal automation is the better answer is mostly marketing noise.

11

@Banderson I think the real question is what Epicor means by “agent” in this context.

In traditional workflow automation, the system follows a predefined path: a trigger occurs, conditions are evaluated, actions execute, exceptions are handled, and the result is logged.

That is deterministic automation.

An agent would imply something more than that. It would imply the system can interpret context, select from available tools, determine the next step, and adapt its behavior based on the situation.

That distinction matters.

If Epicor is using “agent” to describe a governed component that can use approved tools inside a controlled workflow, then I can understand the terminology, even if I still think it needs to be clearly defined.

But if what is being built is primarily a node-based workflow builder with deterministic branches and actions, then I would call that workflow orchestration, not an agent.

So the part I am trying to understand is where the agency actually exists.

Is the AI determining the workflow?

Is it selecting tools?

Is it interpreting exceptions?

Is it only summarizing outcomes?

Or is it simply helping a user build a workflow that later runs deterministically?

Some or even all of these things MAY have been explained at Insights. I hit a few classes on sunday and Monday but by Tuesday I was basically bedridden (I had gotten the flu really bad and missed 90% of the discussion).

So if anyone has a clearer picture of the goal / intention here and how it would work please fill me in. This was a big part of attending this year and of course I got sick :nauseated_face: .

12

After 47,130 orders, 198,030 order lines and 10,989 logged AP invoices done with Gen AI I can safely say human-in-the-loop.

Gen AI when treated more as an aid than fully autonomous gets you the results. More calculator than HAL.

13

100%. The key is basically having enough knowledge to know when AI is full of gas. What are you currently using? External harness? Always interested in hearing what others are doing in this specific space.

14

Custom code hand written code. Started doing this when I found Chat GPT 3.5 could do what I need, the software is that ‘old’, and was pre terminology IE Harness and Agent. I think though Harness is closest.

15

Hey if it works it works. Very clever. You plan on improving it more or are you content with the current state?

@GabeFranco has a really cool harness that he has been developing and experimenting with which I think is really cool. LLM agnostic so you can plug in whatever.

16

Literally patching in AP invoice matching as we speak, well type. :slight_smile:

17

Vision models are continuous AI usage that you can’t exactly escape. We scan job travelers, invoices, and receipts with Qwen3.6 with high accuracy.

On the travelers, traditional OCR fails hard. With a vision and reasoning model, we are able to get much higher accuracy because it can think about what it sees.

18

What’s the most annoying though, when you think about it, is all of that stuff when from 1s and 0s got printed on a piece of paper or as a PDF, and now has to have AI re-read it BACK to 1s and 0s… What a dumb round trip right? lol.

Kind of like the joke (that’s funny because it’s becoming more and more true) where the email was written with AI from a couple of bullet points because the person sending it didn’t want to write the email, then it it was boiled back down to a couple of bullet points using AI because the person on the receiving end didn’t want to read the email, just wanted the bullet points.

If we could stop all of the translating and converting we would be saving sooooo many resources!

19

Lets Go Yes GIF by I Hate Being Single

20

You’re absolutely right! (lol)

Ideally it would be the original digital version I can deterministically extract what I want. Unfortunately, not reality where I am at. For some cases, it is completely inefficient but the alternative is more costly.