What questions to ask in an AI safety platform demo

Every AI safety platform looks impressive in a demo. The footage is clear, the detections are sharp, the dashboard is clean. You walk out thinking, "That looks amazing."

But demos are designed to look amazing. The vendor controls the environment, the camera angles, the lighting, and the scenarios they show you. The real question is whether the platform will perform like that in your facility, with your cameras, your layout, and the messy reality of a working warehouse or manufacturing floor.

Having sat through my share of safety technology demos, I've learned that the questions you ask during the demo matter far more than what the vendor chooses to show you. The right questions separate the platforms that will deliver measurable results from the ones that will stall in a pilot and quietly disappear.

Here's what to ask, and why it matters.

Detection: what does the platform actually see?

This is the flashiest part of any demo, and the easiest place to be impressed without learning much. Every vendor will show you clean detections in ideal conditions. Push beyond that.

What specific safety events does the platform detect? Get a complete list, not just the headline detections. Computer vision AI platforms vary enormously in what they can identify. Some detect pedestrian-vehicle proximity only. Others cover exclusion zone breaches, vehicle speed, PPE compliance, person-in-area violations, and more. You need to know exactly what the system sees, because any detection gap is a risk that goes unmonitored.

What happens when conditions aren't ideal? Ask to see detections in low light, in cluttered environments, and with partial obstructions. A platform that performs perfectly in a clean, well-lit demo environment but struggles when a forklift partially blocks a camera's view or when a shift runs during twilight conditions isn't going to deliver consistent results in your facility.

How does the system handle false positives? Every detection system generates some false positives. The important question is how many, and what happens with them. High false positive rates erode trust with frontline teams quickly. Ask the vendor to share their typical false positive rate and how the system learns over time to reduce them. If they can't give you a straight answer, that tells you something.

Can you connect to our existing cameras? This is a practical question with significant budget implications. Platforms that work with existing CCTV infrastructure save you the cost and disruption of a full camera replacement. Ask about minimum resolution requirements, supported camera types, and whether you need to connect every camera or can focus on high-risk zones.

Beyond detection: what happens next?

Detection is only the starting point. A platform that flags a near miss but does nothing to help you turn that event into a behaviour change is a very expensive alert system. This is the section of the demo where the best platforms separate themselves from the rest.

How are safety events turned into coaching moments? This is arguably the most important question you can ask. Look for a platform that packages detected events into short, reviewable clips that a supervisor can use in a coaching conversation with their team. The difference between "we detected 47 near misses last week" and "here's a 15-second clip of a specific near miss in your zone that you can discuss with your team on Tuesday" is the difference between data and action.

Who receives alerts, and how? Understand the notification workflow. Can alerts be routed to specific supervisors based on zone, shift, or event type? Can you control the volume so frontline leaders aren't overwhelmed? A system that sends every alert to every manager will be muted within a week.

Does the platform support structured safety walks? Some platforms extend beyond real-time detection to support AI-powered safety walks, where the system helps structure site inspections using data on where events are occurring most frequently. This shifts safety walks from random walkarounds to targeted, data-driven exercises.

Reporting and data: what do you actually get?

This is where many demos gloss over the details. The dashboard might look beautiful, but ask what's behind it.

What leading indicators does the platform track? The whole point of computer vision AI is to give you leading indicators that traditional systems can't capture: near-miss frequency, exclusion zone breach trends, vehicle speed distributions, pedestrian-vehicle interaction rates. Ask the vendor to walk you through the specific metrics available, and how they're calculated. Generic "safety scores" without transparent methodology should raise questions.

Can you show me the reporting a board member would see? Executives and board members need a different view from the one a safety manager uses day to day. Ask to see what executive-level reporting looks like: trend lines over time, site-by-site comparisons, risk reduction percentages. If the vendor can't show you a board-ready view, your leadership team won't get the visibility they need to justify ongoing investment.

Does the system integrate with our existing EHS platform? Most organisations already have an EHS system for incident reporting, compliance tracking, and audit management. Ask whether the AI platform can feed data into that system through integrations, or whether you're creating a parallel data silo. The best outcomes come when computer vision data enriches your existing safety ecosystem, not when it sits alongside it.

Can we export raw data? This matters more than you might think. Ask whether you can export event data, trend data, and coaching records in formats your team can work with. Platforms that lock your data behind their own dashboard limit your ability to do your own analysis or combine safety data with operational metrics.

Privacy and security: the trust foundation

This is the section many buyers skip during the demo. Don't. Your workforce's trust in the platform depends on how it handles their data, and your IT team will block deployment if the security posture doesn't meet their standards.

Where is video data processed? There's a significant difference between on-premise processing (where footage stays on your site) and cloud-based processing (where footage is transmitted to external servers). On-premise processing, like inviol's approach, keeps 99% of data within your physical infrastructure, which simplifies your privacy obligations and reduces your data transfer risk.

Does the platform identify individual workers? Ask specifically about face blurring, anonymisation, and whether the system tracks or profiles individuals. A coaching-first platform should detect safety interactions (person near vehicle, person in exclusion zone) without identifying who the person is. This distinction is critical for worker buy-in and regulatory compliance.

What security certifications does the vendor hold? Look for SOC 2 Type II and ISO 27001 at a minimum. These represent independently audited commitments to data security, not just self-declarations. For organisations subject to GDPR, confirm the vendor can support your Data Protection Impact Assessment with documented data flows and processing purposes.

Implementation: how does this actually work in practice?

The demo shows you the product. These questions show you the project.

How long does deployment take from contract to live? Get a realistic timeline, including any site assessments, camera compatibility checks, network requirements, and calibration. Ask for examples from similar-sized deployments, not best-case scenarios.

Do we need new infrastructure? Some platforms require dedicated servers, specific camera models, or network upgrades. Others work as an overlay on existing infrastructure. The infrastructure requirements directly affect your total cost of ownership and deployment speed.

What does onboarding look like for our team? The best technology fails if your safety managers and supervisors don't know how to use it. Ask about training, ongoing support, and whether there's a dedicated customer success contact. Understanding the onboarding process tells you a lot about how the vendor treats customers after the contract is signed.

What does a pilot look like, and what defines success? If the vendor suggests a pilot (and they should), ask how success is measured. A good pilot has clear metrics defined upfront: event detection accuracy, false positive rates, coaching session completion, and ideally early indicators of risk reduction. If the vendor can't define what success looks like, the pilot will drift.

The questions most people forget to ask

Beyond the product itself, there are a few questions that can reveal more about a vendor than any feature walkthrough.

Can I speak with a current customer in my industry? Reference customers are worth more than any sales presentation. Ask specifically for a customer in your industry and at a similar scale. The conversation you have with them, about implementation challenges, adoption hurdles, and actual results, will give you information the vendor never will.

What results can I realistically expect in the first 90 days? Push for specifics, not aspirational numbers. A vendor that tells you "67% average risk reduction" should also be able to tell you what the trajectory looks like: what you'll see in week one versus month three, and what factors influence the speed of improvement.

What happens to my data if we cancel? This is the question that tells you how confident the vendor is in their product. Ask about data portability, retention policies, and what happens to your historical safety data if the contract ends. Vendors who make it easy to leave tend to be the ones whose customers don't.

Making the demo work for you

A demo is a sales exercise. That's not a criticism; it's a reality. The vendor's job is to show you their platform at its best. Your job is to understand how that platform will perform at your worst: during the night shift, in the loading dock with poor lighting, when your team is stretched thin and doesn't have time for another tool.

The questions in this guide are designed to get you past the polish and into the substance. Ask them, take notes on the answers, and compare what you hear across vendors. The platform that gives you clear, specific, honest answers is almost always the one that will deliver clear, specific, honest results.

When you're ready to see how inviol answers these questions, book a demo and put us to the test.

Frequently Asked Questions

What should I look for in an AI safety platform demo?

Focus on four areas: detection capability (what events the system identifies and how it performs in non-ideal conditions), what happens after detection (how events become coaching moments, not just alerts), reporting and data (what leading indicators are tracked and whether the platform integrates with your existing EHS system), and privacy and security (where data is processed, how workers are anonymised, and what certifications the vendor holds). Push beyond the standard demo scenarios and ask to see how the platform handles the messier realities of an operating facility.

How do I compare AI safety platforms during evaluation?

Create a standardised evaluation framework that you apply consistently across every vendor demo. Include detection types, false positive rates, infrastructure requirements, integration capabilities, privacy architecture, deployment timelines, and customer references. Ask each vendor the same set of questions and compare their answers side by side. The platforms that give specific, evidence-backed answers rather than vague assurances will stand out quickly.

What is the most important question to ask a safety technology vendor?

Ask how the platform turns detected safety events into coaching moments. Detection alone doesn't reduce risk. The platforms that consistently deliver measurable incident reduction are the ones that package AI-detected events into practical, shareable coaching tools that supervisors can use with their teams. If the vendor's answer focuses only on alerting and dashboards without a clear coaching workflow, the platform is unlikely to drive lasting behavioural change.

Should I run a pilot before committing to a full deployment?

Yes, and the pilot should have clearly defined success metrics agreed upon before it starts. A well-structured pilot typically covers a defined area of your facility for a set period (often 60 to 90 days), with measurable targets for detection accuracy, false positive rates, coaching session completion, and early risk reduction indicators. The vendor should be able to tell you exactly what success looks like, not just suggest that you "try it and see."

How long does it take to deploy an AI safety monitoring platform?

Deployment timelines vary depending on the platform's architecture. Platforms that overlay onto existing CCTV infrastructure and process data on-premise can typically go live within days to weeks, depending on site complexity and camera compatibility. Platforms requiring new cameras, dedicated server infrastructure, or extensive network upgrades may take significantly longer. Ask the vendor for deployment timelines from comparable sites, not just their best-case example.