In my role leading Global Alliances at Aras, I spend a lot of time in rooms with partners and customers who are all wrestling with the same question right now. Everyone has seen an impressive AI demo. Fewer people can tell me what happens six months after that demo, when the pilot has to become something that runs across thousands of users, years of engineering data, and a real budget.
That gap between what AI can show you and what it actually takes to run it is where the next phase of enterprise AI evaluation will play out. It’s also the gap that’s shaped a lot of our own thinking at Aras, including how we’ve approached AI within Aras InnovatorEdge, which I’ll come back to later in this piece.
For the last few years, the industry conversation has been dominated by capability: assistants that find information, summarize documents, recommend next steps, draft content, and take routine work off someone’s plate. Those demonstrations are genuinely impressive, and the productivity upside is real. But capability was never the hard part. Operating that capability reliably, securely, and affordably across an entire enterprise is the hard part, and it’s the part a fifteen-minute demo simply isn’t built to show you.
A demo shows you the porch, not the foundation
I’ve started thinking about this the way I’d think about touring a row of newly constructed homes. From the curb, ten houses can look nearly identical, with the same square footage, finishes, and curb appeal. What you can’t see from the street is the quality of the foundation, what’s behind the drywall, or whether the electrical work was done to code. You’d never buy a house that way, and increasingly, our customers are telling us they don’t want to buy AI that way either.
A clean conversational interface doesn’t tell you whether the system underneath actually understands how your engineering data connects requirements to designs, designs to changes, changes to suppliers, and service history. It doesn’t tell you whether user permissions carry through, whether an answer can be traced back to its source, or whether the same experience that worked beautifully in a demo will hold up across a production workload with real-world complexity. The parts of enterprise AI that matter most long-term, architecture, data foundation, governance, and economics, are almost never the parts a vendor leads with.
Ask where the intelligence actually runs
One of the most useful questions a buyer can ask is deceptively simple: where does this actually run?
Some AI functionality lives natively inside the platform. Some of it depends on external models, cloud services, vector databases, or partner technologies stitched in behind the scenes. Neither approach is wrong on its own; a distributed architecture can let an organization pick the right model for the right job. But customers deserve a straight answer about which is which, because the architecture decision quietly shapes security exposure, IP protection, performance, who’s on the hook for support, and ultimately, cost.
An assistant can appear to be a native part of your PLM environment while actually routing through several external services, each with its own licensing, consumption charges, and security posture. That’s a reasonable design if you know about it going in. It’s an unpleasant surprise if you find out at renewal time. So, push your vendors to answer plainly: what stays inside the core platform, what leaves it, where do the models actually execute, what data crosses a security boundary to get there, and who owns support for each piece? These aren’t technical footnotes. They define the real risk and cost profile of the investment.
Time saved isn’t the same thing as money saved
Nearly every AI business case I see starts with a time-savings story: an engineer finds information in minutes instead of hours, a quality reviewer covers more records without opening every file, a service tech gets a summarized product history before the call even starts. Those are real benefits. They’re also only half the math.
To get to actual economic value, you have to net out what it costs to enable that productivity in the first place: new entitlements, cloud consumption, model usage, storage, integration work, data prep, monitoring, security, and the professional services and training to stand it all up. A task running 30 percent faster doesn’t give you a 30 percent business benefit. Some of that saved time gets reinvested in verifying what the AI produced. A pilot that looked cheap can look very different once it’s running at enterprise scale and consumption scales with it.
This is exactly the conversation CFOs are starting to bring to the table: Is this AI actually lowering the cost of doing the work, or is it just adding another operating layer whose costs get spread quietly across software, infrastructure, integration, and governance line items?
Watch for a hidden tax on your own data
Enterprise AI runs on enterprise data, and in engineering and manufacturing, that data represents decades of investment: product structures, requirements, configurations, drawings, simulation results, supplier records, quality findings, compliance evidence, service history. That information belongs to the customer, full stop, even when it lives inside a vendor’s application or cloud.
As AI gets woven deeper into enterprise software, it’s fair for customers to ask a pointed question: Are we being charged simply to access our own information through an AI layer? There’s real, chargeable value in securely exposing that data, preserving permissions, orchestrating services, and maintaining traceability; vendors should be paid for that work. But the commercial model needs to clearly separate value-added AI capability from basic access to data customers already own and have already paid to manage. Ask specifically whether new charges are tied to governed APIs, model consumption, infrastructure, prebuilt agents, orchestration, data transformation, or implementation, not just a vague “AI” line item.
Plenty of the executives I talk with have lived through this pattern before, when a broad platform promise becomes slowly fragmented into a stack of add-on modules, connectors, tiers, and fees. It’s reasonable to expect the same scrutiny applied to AI pricing now.
One AI vendor is not an AI strategy
It’s unlikely that a single model will be the right answer for every use case an enterprise has. A model tuned for summarizing documents may be a poor fit for technical analysis. A public model may not clear the bar for a regulated or defense program. A highly specialized model might produce excellent output but be overkill, and overpriced for routine tasks.
Organizations will keep choosing AI services based on the nature of the work: data sensitivity, required accuracy, deployment environment, latency, explainability, and cost. That means your AI strategy shouldn’t be anchored to a single vendor or a single model. Models will keep evolving fast, pricing will shift, and today’s differentiators will become commoditized. The durable investment isn’t any one model – it’s the foundation around it: governed data, permissions, business rules, APIs, traceability, orchestration. If you can’t swap in a new model without rebuilding the digital environment that gives it context, you haven’t reduced vendor lock-in. You’ve just relocated it.
Moving from an AI feature to an AI operating model
A great demo proves an AI feature is technically possible. It doesn’t prove that the capability can be operationalized safely across a real enterprise.
Making AI operational means answering harder questions: who’s accountable when a recommendation is wrong, which actions an agent can take autonomously versus which require a human sign-off, how prompts and models get updated, and how those changes get tested before they touch production work. In engineering environments specifically, these aren’t abstract governance questions – product decisions made or influenced by AI can touch quality, safety, compliance, cost, and long-term service obligations. So, the AI has to do more than produce a plausible-sounding answer; it has to operate inside the same controls that already govern the underlying work, preserving permissions, configuration context, lifecycle state, and a clear record of what informed the outcome.
How I’d tell a customer to evaluate AI quality
Start with the data foundation. Is the AI working against governed, connected enterprise information, or is it searching disconnected documents for similar-sounding text? Does it understand how records relate to each other, or is it just retrieving matches?
Then look at the architecture. A vendor should be able to tell you plainly where your data goes, how it’s processed, which models are involved, and what happens if one of those outside services changes or goes down.
Governance matters just as much. AI should inherit your existing permissions, produce traceable outputs, and let people review, approve, correct, or reverse consequential decisions – not create a shadow process that routes around the controls you already trust.
Finally, ask for evidence, not confidence. How has the capability actually been tested for accuracy, hallucination, latency, security, and scale? Is it running in real production scenarios today, and has it measurably reduced cycle time, improved quality, or lowered risk? A good AI solution proves itself through architecture, governance, and repeatable results, not through the polish of the pitch.
Where Aras fits into this conversation
At Aras, our starting point is that effective engineering AI has to begin with trusted product context, not a chatbot bolted onto the top of PLM. The value in your product information was never just the records themselves; it’s the relationships between requirements, structures, configurations, designs, changes, suppliers, quality events, manufacturing decisions, compliance obligations, and service history. That connected context is the digital thread, and our goal is to let AI work inside it without loosening the permissions, lifecycle controls, and traceability that engineering organizations depend on.
This is the thinking behind Aras InnovatorEdge. Edge API gives partners and customers a controlled way to expose digital thread information and actions. Edge Builder supports building focused applications for specific roles and tasks. Edge AI is meant to be the foundation for building, orchestrating, and governing AI agents that work directly with Aras Innovator®. Not every AI capability has to live inside the core platform. Organizations should be free to bring in different models and services based on their own technical, security, and economic requirements. What has to stay constant is the governed foundation underneath: one authoritative product record, one set of permissions, one traceable history, instead of a patchwork of disconnected AI experiments.
I’ll also say plainly: none of us, vendors or customers, have fully mapped where AI creates the most value yet, or finalized how adoption should look. That’s exactly why Aras is working so closely with our partners and customers to figure out where AI produces real, practical value, what it takes to deploy and govern it responsibly, and how it fits alongside the engineering work that’s already underway.
Demos will keep getting more impressive. The organizations that win with AI will be the ones who made sure the foundation underneath got stronger too – and who asked the hard questions before they signed, not after.
If you’re in the middle of evaluating enterprise AI for your engineering organization and want to talk through what to look for, I’d welcome the conversation. That’s a big part of what my team and I are here for. Connect with me on LinkedIn or reach out to me in the comments below.