Epicor's AI, priced by nobody, and the five things a manufacturer can do instead.

The first useful thing a buyer can do is separate the two product lines, because they solve different problems, attach to different ERPs, and are at different stages of maturity. Epicor's own marketing tends to present them as one AI story, and they are not one thing.

Prism is the agent layer. Epicor describes it as a network of vertical AI agents that sit inside the ERP and let a user ask a question in natural language or hand off a task, with the agent reading and acting on ERP data rather than sitting beside it in a separate window. That last part is the meaningful design choice. An agent that lives inside Kinetic inherits the permissions, the master data and the workflow context that a bolt-on chatbot has to be told about, which is genuinely hard to replicate from outside and is the strongest argument for buying the vendor's own AI rather than someone else's.

Grow AI is the predictive layer, and it is a different animal. It is machine learning on structured historical data rather than a language interface. Its shipped product, Item Advisor, uses models trained on historical sales data to identify which products tend to be bought together, and it surfaces those recommendations inside the order entry workflow. The detail most Kinetic buyers miss is that Item Advisor is integrated with Prophet 21, on either Public or Enterprise Cloud, and Prophet 21 is the distribution ERP. A specialty manufacturer running Kinetic is reading about a product that does not attach to the system it owns.

The maturity question is where the published record gets interesting. Epicor's newsroom carries a release announcing an agentic AI stack including the Prism Agent Foundry and a set of new agents, and the same materials describe agents that are being developed rather than delivered. On the Grow AI side, Epicor is actively recruiting customers into a Controlled Release program to shape the next set of capabilities. Recruiting a controlled release group is a completely normal and honest thing for a software company to do. It is also a precise statement about where a capability sits, and it belongs in the buyer's notes next to any agent that is being demonstrated as though it were shipping today.

Availability is the third axis and it moves faster than the coverage of it. Epicor has published separate regional announcements, including a release covering the launch of Prism across Australia and New Zealand, and a Prophet 21 2026.1 release note describing continued Prism evolution with natural language business rule creation and expanded workflow automation. What that pattern tells a buyer is that Prism is rolling out region by region and release by release rather than arriving everywhere at once. What it means practically is that the answer to "can I have this, in my country, on my version, this quarter" is not derivable from the marketing site at all. It has to be asked, and the answer has to be written into the order.

None of this is a criticism of the engineering. An ERP vendor building agents into the transaction layer is doing the harder and more valuable version of the work, and the companies shipping wrappers around a chat box are doing the easier one. It is a criticism of the buying process, which currently asks a manufacturer to form a budget and a rollout plan from a set of pages that do not state a price, a region, or a ship date.

We checked on August 14, 2026. The URL where an Epicor pricing page would sit, epicor.com/en-us/pricing/, returns HTTP 404. Not a redirect to a quote form, not a page that says pricing is bespoke. A 404.

Across the AI product pages the call to action is a conversation. The single exception, and it is a genuinely interesting one, is on the Prism product page, which describes Prism Business Communications as the industry's first ERP AI agent with outcomes-based pricing. That is the only named pricing model on any Epicor AI surface we read, it applies to one agent rather than the platform, and no figure, unit, or mechanism accompanies it. Outcomes-based pricing is a real and defensible commercial idea. It is also unbudgetable until someone tells you what an outcome is and what one costs.

It is worth being precise about why this matters more for an ERP-attached product than it would for a standalone tool. When you cannot price the AI, you cannot separate the AI decision from the ERP decision. A manufacturer weighing whether to add agents to Kinetic is not comparing two line items; it is being asked to expand a relationship with the vendor that already holds its system of record, on terms that arrive after the evaluation has consumed a quarter of someone's year. The negotiating position that produces is not a good one, and it does not improve by waiting.

The counter is arithmetic you can do without the vendor. Take the workflow you actually want an agent to handle. Count the hours it consumes per month today, at loaded cost, and write that number down before any call. That figure is your ceiling, and it is the only number in the conversation that is yours. If you want the general version of that calculation, what off-the-shelf AI actually costs at scale works through how per-seat and per-transaction pricing behave as a business grows, and the manufacturer calculator will size the internal hours for you in a few minutes.

One vendor in this comparison does publish real numbers, and it is worth naming as the contrast rather than as the recommendation. Tulip Interfaces publishes its plans openly: Essentials at $100 per month per interface billed annually, and Professional at $250 per month per interface billed annually, both with a ten interface minimum. Whether Tulip solves your problem is a separate question entirely. What it proves is that publishing a price in this category is possible, which removes the industry-norms defence from every vendor that does not.

On the Prism EDI Agent page, in the section headed for Kinetic ERP and Epicor EDI customers, sits this sentence, live and publicly readable on August 14, 2026:

"Confirm final availability, packaging, and regional details with Epicor before publishing."

That is not copy written for a customer. It is an internal note to whoever was assembling the page, telling them the availability, the packaging and the regional details were not settled yet and needed checking before the page went out. The page went out.

We are not raising this to embarrass anyone. Every marketing team on earth has shipped a placeholder, and a stray line of internal copy says nothing about whether the underlying software is any good. We raise it because it is the most honest sentence about Epicor AI available to a buyer, and it is accidentally so. It confirms, from inside the company, that at the time that page was written the availability was not final, the packaging was not final, and the regional details were not final. Everything else on the marketing site is written to sound settled. This one line says it is not.

Treat it as a checklist rather than a gotcha. The three things the note says to confirm are exactly the three things a manufacturer should refuse to sign without: which agents are generally available rather than announced or in controlled release, what is included in the licence you already hold against what is a separate SKU, and whether the capability is released in your region and on your version. Those questions are now sourced. They came from Epicor.

The broader lesson generalises past this vendor. In a category moving this fast, the gap between the roadmap and the release notes is where most disappointment lives, and it is nearly always visible in the vendor's own documents if you read the boring ones. Release notes, controlled release announcements and compliance pages are written by people with less incentive to round up than the product page is. That habit is the first thing we teach in how to choose an AI consultant as a manufacturer, and it applies just as well when the consultant is the ERP vendor.

This is the section where we have to correct our own earlier reading, and it is worth stating plainly because the correction runs in Epicor's favour.

Epicor publishes an AI Code of Conduct on its compliance site. The document describes itself as provided on a voluntary basis, which is an accurate and somewhat deflating description of any self-authored standards document, and it is the reason we had previously filed the page as marketing rather than commitment. Reading it properly changes that. The document contains this, which is a substantive commitment rather than a sentiment: Epicor will not use or retain Customer Data, a term the document defines to include Personal Data, or confidential information belonging to a customer for the purposes of training or inputting into any AI system or model to develop or enhance an Epicor AI Product without the customer's prior written approval.

That sentence is stronger than what most vendors in adjacent categories publish anywhere. It names the data, it names the activity, and it sets the default to no. A manufacturer whose process knowledge, routings, costed BOMs and customer pricing all live in the ERP should care about that clause more than about any feature on any product page, because it is the one thing on the list that cannot be undone later. Credit where it is due, and it is a real reason to weigh Epicor's own AI above a third-party tool that ingests the same data under vaguer terms.

Two caveats keep it honest. Voluntary means the commitment sits in a document Epicor maintains rather than in the contract you sign, so the correct next step is to ask that the clause be reproduced in the order form or the master agreement, where it becomes enforceable. And a code of conduct governs the vendor, not its subprocessors, so the model providers underneath deserve the same question in writing. Neither point undoes the commitment. Both are ordinary diligence, and the questions worth asking are laid out in the security questions that belong in front of any AI build.

On the security certifications, we are reporting a limit rather than a finding. Epicor describes an information security program with the usual attestations, and the supporting documentation is made available to existing customers through its support portal rather than published openly. That is a common and reasonable posture. It also means we cannot check it from outside, and we are not going to characterise a report we have not read. Ask for it under NDA, check the date, check the scope, and check whether the specific AI products you are buying sit inside that scope rather than only the core ERP.

A manufacturer researching Epicor will meet a Forrester Total Economic Impact study, and it will be quoted back at them in a demo. Here is what it is, stated with the attributions that usually get dropped.

The study is a Total Economic Impact analysis of Epicor Kinetic for manufacturing, commissioned by Epicor and delivered by Forrester Consulting. It does not measure a population of Epicor customers. It builds a single composite organisation from interviews with four manufacturing organisations, and that composite is a discrete diversified manufacturer with $150 million in annual revenue in Year 1, expanding by acquisition to $300 million in Year 2. Against that composite, the study reports benefits of $20.39 million over three years, costs of $5.50 million, a net present value of $14.87 million, a 270 percent return on investment, and a 20-month payback period.

Read that construction twice, because it is doing a lot of work. A composite built from four interviews is a modelling device, not a sample. It has no distribution, no median, no failure cases and no confidence interval, and it cannot have any of those things, because four organisations were combined into one hypothetical company rather than measured as a group. The numbers that come out are internally consistent with the assumptions that went in. They are not a forecast for your plant.

None of that makes the study dishonest. Commissioned TEI studies are a standard format, Forrester discloses the sponsorship and the composite method inside the document, and the exercise of modelling benefits against costs over three years is a useful discipline. The dishonesty, when it happens, is downstream: the figure gets repeated as though 270 percent were an observed average across Epicor customers. It is a modelled result for one invented company that is larger than most specialty manufacturers, doubling in size by acquisition, and running the full ERP rather than the AI products this page is about. The study is about Kinetic. It is not a measurement of Prism or Grow AI at all.

The useful response is not cynicism, it is substitution. Build the same shape of model with your own numbers. Name the two or three workflows you would actually put an agent on, measure what they cost you today, estimate the fraction a tool could take, and carry your own implementation and change-management costs rather than a composite's. How to measure ROI on a mid-market AI engagement walks through that calculation, and the manufacturing AI benchmark covers what the wider field is reporting. A model you built is worth more in a board meeting than a model somebody built for a vendor, even when yours produces a smaller number. Especially then.

Five vendors a specialty manufacturer might reasonably shortlist, on the four questions that are answerable from outside, each read on its own site on August 14, 2026.

Two notes on things that have moved, because a manufacturer working from a comparison article written a year ago will get them wrong. Infor's CloudSuite Industrial product URL no longer resolves to a product page and now lands on Infor's industrial manufacturing industry page instead, so anything you read about SyteLine positioning deserves rechecking against the current site. And Plex now lives at plex.rockwellautomation.com under Rockwell Automation, which matters because it is an MES lineage rather than an ERP one, and putting it on an ERP shortlist compares two different categories of product.

The pattern across the table is the finding. Four of five vendors will not tell a manufacturer what anything costs until a salesperson is in the room, and the fifth is the smallest commitment on the list. That is not a coincidence and it is not a conspiracy. It is what happens in a category where the buyer's switching costs are enormous and the seller knows it. The defence is to arrive with your own number.

How to tell which one you are.

Three questions decide this, and a vendor cannot answer any of them for you.

The first is whether your problem is a language problem or a numbers problem. Prism is a language interface onto ERP data, and it earns its keep where people currently ask other people questions that the system already knows the answer to. Grow AI is prediction on structured history, and it earns its keep where a recurring judgment call has enough clean past data behind it. If your actual pain is that a job travels through the shop with three different due dates on it, neither product is aimed at you, and buying one will produce a very impressive demo followed by a quiet year.

The second is how particular your process is. This is the question that separates buy from build more reliably than size does. A manufacturer whose scheduling, quoting and routing conventions look broadly like everyone else's in the industry should buy, because a configured product will fit and a build will burn money reinventing a standard. A manufacturer whose margin comes from doing something structurally different, which is most specialty shops, tends to find that the configurable product covers the eighty percent that was never the problem. The general form of that test applies across every vertical we work in.

The third is whether your data can support any of it. This is the one that quietly decides the outcome. If your job history is inconsistent, if machine data never reaches the ERP, or if half the shop's real scheduling logic lives in one person's spreadsheet, then every option on this page inherits that problem and the honest first project is data readiness rather than AI. That is a less exciting sentence than any vendor will say to you, and it is the single most common reason these purchases underdeliver in year one. The maturity assessment is a faster way to find out where you stand than a quarter of trying.

There is a fourth condition that overrides the other three. If nobody in the plant owns the outcome, meaning there is no named person whose job it is to decide within ninety days whether the thing stays or goes, the purchase will drift into a subscription that no one can defend and no one will cancel. That failure has nothing to do with which vendor was chosen, and it is the most common one we see.

What we would build, and what we have not built.

ColabContent commissions custom AI systems for mid-market operators at a fixed fee, with the code owned by the client at handoff, and we say no when an off-the-shelf product is the better instrument. For a specialty manufacturer, the pattern we would build is a narrow layer on top of the ERP rather than a replacement for any part of it: it reads the job history, routings and machine data you already have, does the one thing your plant actually needs done, and writes back into Kinetic through supported interfaces. The differences from a subscription are that it runs on your own data, it does not charge again per agent or per seat, and it belongs to you afterward. What commissioning means and what custom AI means cover the model in full.

We have not shipped a commission for an Epicor shop, and we would rather say that here than let a comparison page imply otherwise. Our delivered manufacturing work has not been on Kinetic or Prophet 21. That is a genuine limit on what we can claim, it is the kind of thing worth asking any consultant to state plainly before engaging them, and it is why the honest recommendation on this page is frequently to buy the product rather than to call us. If you want the cost side before any conversation, what a manufacturing commission actually costs lays out the bands, and what a mid-market engagement costs covers the general case.

What is Epicor Prism, and what does it actually do?

Prism is Epicor's agent layer for its ERP products. Epicor describes it as a network of vertical AI agents that work inside Kinetic and Prophet 21, letting a user ask a question in natural language or hand off a task, with the agent reading and acting on ERP data rather than sitting beside the system in a separate application. Named surfaces include a coding assistance agent inside Epicor Application Studio, Prism Business Communications for RFQ workflows in Kinetic, and a Prism EDI Agent for customers running Kinetic ERP together with Epicor EDI, which answers questions about EDI documents, errors and transaction status. Epicor has also published an agentic AI stack release describing a Prism Agent Foundry and further agents, some of which are described as in development rather than delivered, so a buyer should ask which specific agents are generally available on their version and in their region.

How much does Epicor AI cost?

Epicor publishes no price for its AI products. We checked on August 14, 2026 and the URL where an Epicor pricing page would sit returns HTTP 404. The only pricing model named anywhere on the Prism product page is outcomes-based pricing, applied to a single agent called Prism Business Communications and described as an industry first, with no figure, unit or mechanism attached to it. Every other AI product page ends in a conversation with sales. The practical consequence is that a manufacturer cannot build a business case before taking the sales call, which means the vendor controls the order in which you learn things. The counter is to fix your own walk-away number first, based on what the target workflow costs your plant today in loaded hours, and to give that number to whoever attends the call.

Is Epicor Grow AI available for Kinetic?

Grow AI is a different product line from Prism and it is important not to conflate them. Grow AI is predictive machine learning on structured historical data rather than a language interface, and its shipped product, Item Advisor, is integrated with Epicor Prophet 21 on either Public or Enterprise Cloud. Prophet 21 is the distribution ERP, so a specialty manufacturer running Kinetic is reading about a product that does not attach to the system it owns. Epicor is separately recruiting customers into a Controlled Release program to shape the next set of Grow AI capabilities, which is a normal and honest way to develop software and is also a precise statement about maturity. If a demo includes a Grow AI capability, ask whether it is generally available on your ERP or part of that controlled release.

Does Epicor train its AI models on customer data?

Epicor's published AI Code of Conduct commits that it will not. The document states that Epicor will not use or retain Customer Data, a term it defines to include Personal Data, or confidential information belonging to a customer, for the purposes of training or inputting into any AI system or model to develop or enhance an Epicor AI Product without the customer's prior written approval. That is a substantive commitment and it is stronger than what many vendors in adjacent categories publish. Two caveats keep it accurate. The document describes itself as provided on a voluntary basis, meaning it lives in a standards document Epicor maintains rather than in the contract you sign, so ask for the clause to be reproduced in the order form or master agreement where it becomes enforceable. And a code of conduct governs the vendor rather than its subprocessors, so the underlying model providers deserve the same question in writing.

Why does Epicor's own EDI agent page say to confirm availability before publishing?

Because it is an internal instruction that was never removed. On the Prism EDI Agent page, in the section addressed to Kinetic ERP and Epicor EDI customers, the sentence "Confirm final availability, packaging, and regional details with Epicor before publishing" was live and publicly readable on August 14, 2026. It is a note to whoever assembled the page telling them those three things were not yet settled. We do not raise it to embarrass anyone, because every marketing team has shipped a placeholder and a stray line says nothing about the quality of the software. We raise it because it is accidentally the most honest sentence available to a buyer, and because the three items it names are exactly the three a manufacturer should refuse to sign without: which agents are generally available rather than announced, what is included in the licence you already hold versus a separate SKU, and whether the capability ships in your region and on your version.

Is the 270 percent ROI figure for Epicor real?

It is real as a published study and it does not mean what it is usually used to mean. The figure comes from a Forrester Total Economic Impact analysis of Epicor Kinetic for manufacturing, commissioned by Epicor and delivered by Forrester Consulting. It does not measure a population of Epicor customers. It builds a single composite organisation from interviews with four manufacturing organisations, and that composite is a discrete diversified manufacturer with $150 million in annual revenue in Year 1 expanding by acquisition to $300 million in Year 2. Against that composite the study reports benefits of $20.39 million over three years, costs of $5.50 million, a net present value of $14.87 million, a 270 percent return on investment and a 20-month payback. A composite built from four interviews is a modelling device rather than a sample, so it has no median, no failure cases and no confidence interval. It is also a study about the Kinetic ERP rather than about Prism or Grow AI.

What are the real alternatives to Epicor's AI for a manufacturer?

There are five, and only one of them involves buying something from a new vendor. The first is to establish exactly which AI capabilities the Kinetic or Prophet 21 licence you already pay for includes, because that line moves between releases and the answer is in your release notes rather than on the product pages. The second is to buy a focused point tool for the specific problem, such as Tulip for frontline operations apps or MachineMetrics and Datanomix for machine utilisation and downtime. The third is to move the layer that is actually failing, which for many shops is quoting and estimating rather than the ERP. The fourth is to wait one release cycle on purpose, with a named owner and a review date, because several agents are still announced or in controlled release. The fifth is to commission a layer on the ERP data you already own, which is the right instrument above a certain scale and when the workflow is genuinely particular to how you build.

Which manufacturing AI vendors publish a price at all?

Of the vendors on this page, one does. Tulip Interfaces publishes its plans openly, with Essentials at $100 per month per interface billed annually and Professional at $250 per month per interface billed annually, both carrying a ten interface minimum. MachineMetrics has a pricing page but no dollar figure on it, listing three tiers by name, Core Platform, Intelligent MES and Enterprise, and stating that pricing is volume based so that per-machine cost falls as machine count rises. Datanomix's pricing URL returns a 404. Epicor's pricing URL returns a 404. That Tulip can publish real numbers in this category is the useful fact, because it removes the industry-norms defence from every vendor that will not, even though whether Tulip solves your particular problem is a completely separate question from whether it is priced honestly.

Should a manufacturer buy Epicor's AI or build its own layer?

Buy when your processes look broadly like the industry's and build when your margin comes from doing something structurally different. A manufacturer whose scheduling, quoting and routing conventions are conventional should buy, because a configured product will fit and a build will spend money reinventing a standard. A specialty manufacturer whose advantage is a particular way of building tends to find that a configurable product covers the eighty percent that was never the problem, and that the remaining twenty percent is the entire reason for the project. There is also a strong argument for the vendor's own AI that has nothing to do with features: an agent inside the ERP inherits permissions, master data and workflow context that an external tool has to be told about. Weigh that against a commissioned layer's advantages, which are that you can price it, you own it at handoff, and it does not depend on which region the vendor ships to next.

What should we ask Epicor before signing anything?

Six questions, and the first three come from Epicor's own page. Which specific agents are generally available today on our ERP, our version and our region, as opposed to announced, in development, or in controlled release. What is included in the licence we already hold and what is a separate SKU. What are the final packaging and regional details, which is the exact language Epicor's own EDI agent page instructs its staff to confirm. Then three of your own. What does it cost, expressed in a unit we can forecast, and what specifically does outcomes-based pricing mean for the agent it applies to. Will the training and retention commitment from your AI Code of Conduct be reproduced in our contract, and does it flow down to your subprocessors and model providers. And can we see the current security report under NDA, with the AI products in scope rather than only the core ERP. Get the answers in writing before the evaluation consumes a quarter of someone's year.

Related reading.

More on AI in specialty and discrete manufacturing: what we would build for a specialty manufacturer, what AI consulting for manufacturing covers, and the consultants manufacturers actually shortlist. On the ERP layer specifically, the Epicor Kinetic playbook covers building an AI layer on the Kinetic you already own, and manufacturing automation consulting covers the wider engagement. For the back office rather than the floor, who automates back-office workflows in manufacturing. The neighbouring comparisons cover different layers of the same stack: Paperless Parts and the quoting layer, renting AI against owning it, and generic SaaS AI against a commission. On ownership, why code handoff matters. To size your own numbers before a call, the manufacturer calculator takes a few minutes, and how we work covers what happens next. The rest of this series sits on the comparisons hub.