You’ve decided your plant could benefit from AI. Maybe it’s predictive maintenance, maybe it’s quality inspection, maybe it’s production scheduling. Now comes the hard part: picking the right vendor without getting burned.

After sitting through dozens of vendor pitches and watching several implementations go sideways, here’s the checklist I wish someone had given me five years ago.

Questions to Ask Before the First Demo

Most vendor conversations start with a flashy demo. Resist the urge to get excited and ask these questions first:

“What manufacturing environments have you deployed in?” Generic AI companies will talk about their technology. Manufacturing-focused vendors will name specific industries, line speeds, and equipment types. You want the second kind.

“Can I talk to a reference site running your system for more than 12 months?” Anyone can find a happy customer at month three. The real test is year two, when the implementation team has moved on and the plant is running the system themselves.

“What does your system need from our existing infrastructure?” This question exposes whether the vendor has actually thought about integration. If they can’t tell you exactly which PLCs, protocols, and data formats they support, that’s a warning sign.

“Who on my team needs to be involved, and for how long?” Legitimate vendors will give you honest staffing estimates. If they say “minimal disruption to operations,” push harder. Every implementation requires floor time from your best people.

Red Flags That Should Stop the Conversation

Years of watching AI projects succeed and fail in manufacturing have made some patterns obvious.

They guarantee specific ROI numbers before seeing your plant. No responsible vendor promises 40% efficiency gains without understanding your current state. Your factory isn’t their previous customer’s factory.

Their team has no manufacturing experience. Ask how many of their engineers have spent time on a factory floor. Software developers who’ve never dealt with production pressure, shift changes, or equipment that’s been running since 1998 will struggle with the realities of your environment.

They want a multi-year contract before a pilot. Any vendor confident in their product should be willing to prove it works at small scale first. If they need a three-year commitment upfront, ask yourself why.

They can’t explain how their system handles bad data. Factory data is messy. Sensors drift, connections drop, operators enter things incorrectly. If the vendor’s answer to “what happens when data quality is poor” is a blank stare, move on.

They dismiss your existing systems. A vendor who wants to rip out your SCADA setup and start from scratch is selling you their platform, not solving your problem. Good AI systems work with what you have.

Integration Requirements Worth Documenting

Before you sign anything, get written answers on these integration points:

• PLC compatibility. Which specific PLC models and firmware versions are supported? Siemens S7-1500 and Allen-Bradley ControlLogix behave differently.
• Communication protocols. Do they support OPC UA, Modbus TCP, EtherNet/IP, or MQTT? Which ones have they actually deployed in production, not just listed on a spec sheet?
• Network requirements. What bandwidth do they need? Do they require cloud connectivity, or can everything run on your local network? Many plants have restricted or air-gapped networks for good reason.
• Data storage and ownership. Where does your production data go? Who owns it? Can you export it if you switch vendors?
• IT/OT security. How does their system handle authentication, encryption, and network segmentation? Your IT security team will ask, so get the answers early.

According to Gartner’s research on manufacturing technology adoption, integration complexity remains the top reason AI projects stall in industrial settings. Getting these details in writing before you commit saves months of frustration later.

Realistic Timelines You Should Expect

Here’s what a responsible implementation timeline looks like for most manufacturing AI projects:

Weeks 1-2: Site assessment and scoping. The vendor visits your plant, maps your systems, identifies data sources, and defines what success looks like. Companies like AI Development Company Melbourne specialists can usually do a site assessment in a week. If a vendor wants to skip this step, they’re guessing.

Weeks 3-6: Pilot setup. Installing sensors, connecting to your systems, getting data flowing. This almost always takes longer than quoted. Budget for delays caused by legacy equipment, network issues, or scheduling around production runs.

Weeks 7-12: Model training and tuning. The AI needs to learn your specific environment. This is iterative. Expect regular check-ins and adjustments as the model encounters real-world conditions it wasn’t trained on.

Months 4-6: Supervised operation. The system runs alongside your existing processes. Your team monitors outputs and provides feedback. This is not optional—it’s where you build confidence and catch problems before they matter.

Months 6-12: Full deployment and optimisation. Gradual expansion to additional lines or use cases, based on pilot results.

Any vendor promising production-ready AI in four weeks is either working on a trivially simple problem or setting unrealistic expectations.

The Bottom Line

Picking an AI vendor for your plant isn’t like buying office software. The stakes are higher, integration is harder, and a bad choice shows up on your production line.

Take your time. Ask hard questions. Talk to real references. And don’t let anyone rush you into a decision based on a polished demo and a promise of transformation. The right vendor will welcome your scrutiny because they’ve earned it.