If your MES renewal is landing this quarter, there’s a good chance it now includes a line item you didn’t ask for: an AI copilot. Siemens has been building generative AI features into the Opcenter line. SAP is pushing Joule as an assistant layer across its enterprise stack, including manufacturing modules. Rockwell Automation’s Plex and FactoryTalk portfolios have picked up assistant-style features aimed at production and quality workflows. AVEVA has done the same across its MES and operations offerings. The pattern is consistent: a generative or agentic AI layer, sold either as a bundled capability in a new pricing tier or as a metered add-on, showing up on renewal paperwork whether or not the plant asked for it.
None of this is inherently bad. Copilots that can summarize a shift’s downtime events, draft a root-cause narrative from historian tags, or answer a natural-language question against work-order data are genuinely useful when they work. The problem is that “AI-included” has no standard definition across vendors, and in some cases no standard definition within a single vendor’s own price book from one quarter to the next. That ambiguity is exactly where you get overcharged, or worse, sign up for a capability you can’t actually operationalize this year.
What “bundled” usually means, and what it doesn’t
Across this category, the AI layer tends to fall into one of three commercial shapes, and it matters enormously which one you’re being quoted.
- Included in the platform tier. Basic natural-language query or a chat-style interface over existing dashboards, often capped in scope — read-only, limited to certain modules, no write-back to MES records.
- Metered by consumption. Per-seat licensing for named users of the copilot, per-query or per-token charges for generative responses, or per-agent pricing if the vendor is selling autonomous or semi-autonomous “agents” that take actions (creating a work order, flagging a deviation, routing an alert) rather than just answering questions.
- Separately licensed module. A distinct SKU that sits alongside your core MES license, priced and renewed independently, sometimes tied to a separate cloud consumption agreement because the model inference doesn’t run on-prem.
The vendors have a real incentive to blur these categories in a demo and in early-stage commercial conversation, not out of bad faith, but because “AI included” is a stronger renewal narrative than “here’s another SKU.” Your job is to get the quote broken out into which of the three buckets each capability actually lives in, before you sign anything.
The agent question is the one that actually costs money
The distinction between a copilot that answers questions and an agent that takes action is not a marketing nuance — it’s the difference between a reporting tool and something that touches your MES data model and potentially your ISA-95 transaction flows. Per-agent pricing is newer and less standardized than per-seat or per-query pricing, and it’s where vendors have the most room to define terms favorably to themselves. If a proposed capability can write back to genealogy records, trigger a quality hold, or auto-generate a nonconformance report, treat it as a distinct negotiation line, not a feature bullet under the platform tier.
Pressure-test the demo against your own historian, not theirs
Every vendor demo of an MES copilot runs against a curated dataset: clean tag names, complete batch records, no sensor dropouts, no shift-to-shift naming drift. Your plant’s historian looks nothing like that. Before you evaluate any copilot on capability, insist on a proof-of-concept against your own data, and specifically against the parts of your data that are messiest — the line that got re-tagged after an equipment swap, the product code that means two different things in two different work centers, the OPC UA tags with inconsistent naming from a controls upgrade three years ago.
A few things worth checking directly rather than taking on faith from a sales engineer:
- Does the copilot correctly refuse or flag a question it can’t answer confidently, or does it generate a plausible-sounding but wrong number? This is the single most important thing to test — a wrong OEE figure delivered with total confidence is worse than no answer.
- How does it handle unit-of-measure inconsistencies, duplicate equipment IDs, or gaps in historian coverage from network drops?
- Can it explain, in plain terms, which underlying tags or records it pulled to generate an answer? If the answer is a black box with no traceable lineage, that’s a real limitation for anyone in a regulated environment who needs an audit trail.
- What happens with your actual naming conventions and your actual ISA-88 recipe structure, not the vendor’s reference model?
If a vendor resists running the demo against a sanitized export of your own data, that’s worth noting as a data point in itself — reasonable security concerns are understandable, but a copilot you can’t test against your own mess is a copilot you can’t trust in production.
Negotiating AI out of the contract, cleanly
Plenty of plants are not ready to operationalize an MES copilot this year — not because the technology is bad, but because the prerequisite work (clean master data, a governed historian, defined use cases, a change-management plan for operators) hasn’t happened yet. That’s a legitimate position, and it’s negotiable.
- Ask for the AI capability to be quoted and itemized as an optional line, separate from the base platform renewal, with its own start date.
- Confirm in writing whether declining it now affects pricing or tier eligibility later — some vendors structure tiers so that adding AI after the fact costs more than including it upfront, which is a legitimate commercial lever but one you should know about before you decide.
- If a “free” or included copilot feature is tied to a cloud consumption meter, get the metering unit defined precisely (per query, per token, per user-month) and ask for a usage cap or alert threshold so a curious plant floor doesn’t generate a surprise bill.
- Push for a defined pilot period with an exit clause, rather than a multi-year commitment on a feature set that both you and the vendor are still learning how to use well.
Bottom line
In our assessment, the honest way to evaluate any MES vendor’s AI copilot right now is to treat it like any other module: judge it on what it does against your data, not what it does against theirs, and price it as its own line item regardless of how it’s bundled on the quote. Siemens, SAP, Rockwell, and AVEVA are all racing to establish their assistant layer as the default interface to shop-floor data, and there’s real long-term value in that direction. But “included in your renewal” and “ready for your plant” are two different claims, and only one of them is yours to verify. Do the proof-of-concept on your own mess before you decide which capabilities are worth paying for this cycle — and don’t let a genuinely promising technology talk you into operationalizing it before your data and your team are ready to trust it.
This article was written with the assistance of artificial intelligence. While we aim for accuracy, the information may be incomplete, out of date, or incorrect, and should be independently verified before you rely on it for any decision. It is provided for general information only and does not constitute professional advice.
