Scaling AI safety from one site to many: lessons from multi-site deployments

Getting AI safety monitoring working well at one site is a genuine achievement. Getting it working consistently across five, ten, or fifty sites is a different problem entirely.

Most organisations that have tried to scale a safety technology deployment know the pattern: the pilot facility runs beautifully, results are strong, leadership approves the rollout — and then everything gets complicated. Different facilities have different camera setups. Site managers have different levels of buy-in. Data ends up in separate systems that no one can easily compare. Coaching processes that worked at Site A don't translate to Site B, because the shift supervisors there do things differently.

None of this is a reason not to scale. The potential value of enterprise-wide safety monitoring is significant: consistent risk visibility, comparable performance data across sites, and the ability to identify what your best-performing locations are doing differently and replicate it. But getting there requires planning that most technology evaluations don't cover in enough depth.

This piece draws on what actually matters when scaling safety technology from a single deployment to a multi-site programme.

Why multi-site safety management is hard even without AI

Before getting into the technology, it is worth acknowledging that multi-site safety management is genuinely difficult — even with mature, well-resourced programmes.

The core challenges are structural. Each site tends to develop its own risk profile based on its physical layout, the mix of tasks performed, and the experience levels of the workforce. A distribution centre with 24/7 operations and a high proportion of agency workers has a different risk environment to a manufacturing plant with stable, long-tenured staff. Even within the same organisation, expecting identical safety outcomes from different facilities is unrealistic.

On top of site-level variation, multi-site operations typically have to manage regulatory complexity. In Australia, work health and safety law is administered at the state and territory level, meaning a company operating in New South Wales (SafeWork NSW), Victoria (WorkSafe Victoria), and Queensland (Workplace Health and Safety Queensland) technically operates under three separate regulatory regimes, even though the model WHS laws have substantially harmonised obligations across most jurisdictions. In New Zealand, the Health and Safety at Work Act 2015 (HSWA) applies nationally, but operational requirements may still differ between site types.

The result is that enterprise safety managers face a constant tension between the need for consistency (to benchmark performance across sites, identify systemic risks, and demonstrate company-wide compliance) and the reality of local variation (different layouts, risks, workforces, and sometimes regulations).

Technology alone does not resolve this tension. But it can make it dramatically easier to manage.

The visibility problem at scale

The most fundamental problem in multi-site safety management is information. When safety data lives in site-level spreadsheets, incident reports, and the heads of individual supervisors, you cannot get a reliable picture of what is happening across the organisation.

This matters enormously for decision-making. The National Safety Council's 2023 data put the total cost of work injuries in the US at $176.5 billion — around $1,080 per worker. The business case for reducing injury rates is clear. But you cannot manage what you cannot measure, and in multi-site operations, measurement is frequently the bottleneck.

Traditional safety monitoring approaches (manual walkthroughs, self-reported near misses, periodic audits) produce data that is inconsistent, delayed, and dependent on whoever happened to be observing at the time. A safety walk at Site A this week captures a snapshot. Whether that snapshot is representative of normal conditions, and how it compares to a snapshot taken at Site B three weeks later, is almost impossible to know.

Computer vision AI changes this by making monitoring continuous and objective. The same detection logic operates at every camera, at every hour, regardless of which supervisor is on shift. This creates data that is genuinely comparable across sites — not because everything is identical, but because the measurement methodology is consistent.

Principle 1: Start with your highest-risk sites, not your most willing ones

There is a natural temptation to begin a multi-site rollout with the locations that are easiest or where site leadership is most enthusiastic. This is understandable but can create problems later.

Deploying first to your lowest-complexity, most cooperative sites generates a set of early results that may not be representative of what you will encounter across the portfolio. When you then move to higher-risk or less cooperative sites, the gap between expectation and reality can undermine confidence in the programme.

A better approach is to sequence your rollout based on risk. Start with the sites where the potential for harm is highest — the highest injury rates, the most complex vehicle-pedestrian interactions, the most acute compliance exposure. These are the deployments where the return is greatest and where objective data is most urgently needed.

This also makes the business case easier to build at board level. Results from your highest-risk facilities are the most compelling evidence for ongoing investment, and they demonstrate that the technology performs under the most demanding conditions.

Principle 2: Camera infrastructure matters more than you think

In theory, computer vision AI can be deployed across any CCTV network. In practice, the quality and placement of existing cameras varies significantly across sites, and this affects what the system can detect and with what accuracy.

A few specific issues come up consistently in multi-site rollouts:

Coverage gaps. Every facility has areas where cameras either do not exist or do not cover the angles needed to detect the highest-risk behaviours. In warehouses, this often means loading dock interactions, blind corners at intersections, and the areas immediately adjacent to battery charging stations. Site-by-site camera audits before deployment allow you to identify gaps and prioritise additional coverage without having to instrument every camera in every facility.

Camera quality. Older cameras operating at lower resolutions or frame rates may produce footage that reduces detection accuracy. Knowing this before rollout allows you to set appropriate expectations for each site or plan targeted infrastructure upgrades.

Network architecture. For organisations with strict data governance requirements, understanding how footage is processed and retained at each site is important to get right before deployment — not after. A platform that processes footage on-premise (rather than sending it to the cloud) addresses the most common enterprise data security concerns and simplifies compliance across jurisdictions.

inviol works with existing CCTV infrastructure and typically focuses on a selection of cameras covering the highest-risk areas, rather than requiring full facility coverage from day one. This makes phased rollouts more practical and means you can prioritise deployment resources where the impact is greatest.

Principle 3: The data consistency problem is real — and solvable

One of the most frequent frustrations in multi-site safety technology programmes is that you end up with data that looks comparable but is not. If Site A has configured its system to detect a certain type of event and Site B has configured it differently, comparing their event rates tells you almost nothing about relative risk.

This problem is not unique to safety technology. It is a general feature of enterprise data management. But it is particularly acute in safety, where the temptation to let individual sites customise their configuration to match local preferences can quietly destroy comparability.

The solution is to establish a common baseline configuration — a defined set of detection types, thresholds, and reporting categories that every site operates against — before extending local customisation. This gives you a consistent enterprise view while still allowing site-level flexibility for genuinely different risk environments.

Data governance matters here too. Defining who owns safety data at the site level versus the enterprise level, how it is aggregated, and who can see what, is a decision that is much easier to make before rollout than after. In practice, the most successful multi-site programmes have a clear reporting cadence: site supervisors review their own event data daily, regional or operations managers receive a weekly cross-site comparison, and safety leadership gets a monthly portfolio view.

Principle 4: The coaching consistency problem is harder than the data problem

If the data problem in multi-site safety is about making information comparable, the coaching problem is about making the response to that information consistent.

In a coaching-first safety approach, the value of detecting a safety event lies not in the detection itself but in what happens next: who reviews the footage, who has the coaching conversation with the worker, what is documented, and what changes as a result. Getting this right at one site, with a motivated supervisor who has been trained in the methodology, is relatively achievable. Getting it right at twenty sites, with supervisors who have different levels of skill, time, and buy-in, is much harder.

What consistently separates high-performing multi-site programmes from average ones is not the technology: it is the coaching infrastructure around the technology. This means standardised coaching workflows that supervisors follow when an event is flagged, accessible documentation that records what was discussed and agreed, and visibility at the enterprise level of whether coaching is actually happening.

Organisations that treat the technology as the solution tend to plateau. Organisations that treat the technology as the detection and evidence layer, and invest equally in building coaching capability, achieve sustained improvement. The leading indicator data generated by continuous monitoring is only as useful as the quality of the coaching conversations it enables.

Principle 5: Benchmarking across sites is valuable — but requires honesty about context

One of the most powerful capabilities of a multi-site safety monitoring programme is the ability to benchmark performance: which sites have the lowest event rates, which have improved most in the past quarter, which are struggling in specific risk categories.

Used well, this kind of benchmarking drives genuine improvement. Site managers respond to peer comparison in ways they do not respond to abstract safety goals. Knowing that Site C has reduced pedestrian-vehicle near-misses by 60% while Site D's rate has barely moved creates a specific, actionable conversation.

The risk is that benchmarking becomes a ranking exercise that drives the wrong behaviours: suppressing event reporting, focusing on the metric rather than the underlying risk, or dismissing good safety practice at a site because its absolute numbers look worse than another site with less complex operations.

The safest approach is to benchmark on rate of change and improvement, not on absolute event counts. A site with a high-risk layout and an aging facility will produce more events than a purpose-built facility with newer infrastructure. That does not make it less safe — it may simply reflect a more honest measurement environment. What matters is the trend: is the rate going down, are coaching conversations happening, are the most serious event types being addressed?

inviol's reporting and analytics capabilities are designed for exactly this kind of multi-site comparison: cross-site event frequency, trending by category, and comparison of coaching activity alongside event data. Average risk reduction across inviol customers is 67%, with a 42% reduction in incidents over three years and a 61% reduction in machine-on-plant incidents. Understanding how your sites sit relative to those benchmarks, and relative to each other, is the foundation of an evidence-based enterprise safety strategy.

Principle 6: Compliance exposure is not uniform across sites

For organisations operating across multiple jurisdictions, compliance risk is not evenly distributed. Australia's state-based WHS regulatory framework means that a company operating across multiple states may face different regulator priorities, different enforcement activities, and different incident notification requirements depending on which jurisdiction each site falls under.

In Australia, the model WHS laws have been adopted by all jurisdictions except Victoria (which operates under the Occupational Health and Safety Act 2004). This means operational requirements are broadly similar across most sites, but enforcement styles and regulator focus areas can vary at the state level. New South Wales, for example, introduced an industrial manslaughter offence in June 2024, with maximum penalties of $20 million for a body corporate and up to 25 years' imprisonment for individuals. Other jurisdictions have had similar laws in place for longer.

For multi-site operations spanning New Zealand and Australia, this means that enterprise safety management needs to account for HSWA obligations in NZ alongside state-level WHS frameworks across the Tasman. Building consistent monitoring and documentation practices across all sites makes it significantly easier to demonstrate compliance in multiple jurisdictions when it matters: during a regulator visit, following a notifiable event, or in the context of a board-level due diligence review.

What good looks like at scale

The organisations that manage multi-site safety most effectively tend to share a few characteristics that have nothing to do with the technology they use.

They have clear accountability at the enterprise level for safety outcomes, not just compliance activity. There is someone (or a small team) who owns the cross-site view and has the authority to act on it. They treat coaching as a core operating rhythm, not an occasional response to serious events. And they use data to have honest conversations — including about sites that are struggling — rather than presenting performance in the most favourable possible light.

The technology layer adds genuine value: continuous visibility, objective data, comparable metrics, and the evidence trail that supports both coaching and compliance. But it amplifies what is already there. A strong safety culture at one site, scaled with the right monitoring infrastructure, becomes a strong safety culture at every site. A weak culture with patchy accountability, equipped with the same technology, produces patchy results.

For organisations at the earlier stages of a multi-site deployment, book a demo to understand how inviol's enterprise reporting and coaching infrastructure supports a scaled programme — including what the rollout sequencing looks like in practice.

Frequently Asked Questions

What are the biggest challenges when scaling AI safety monitoring across multiple sites?

The most common challenges are data consistency (ensuring detection configurations are comparable across sites so performance data can be meaningfully benchmarked), coaching consistency (replicating good coaching practices across different site managers), and camera infrastructure variability (sites often have different camera coverage and quality). Sequencing rollout by risk level rather than convenience also significantly improves outcomes.

Do all sites need to be identical in their AI safety monitoring configuration?

No, and trying to force identical configurations often creates friction without improving results. The goal is a common baseline — a consistent set of detections and reporting categories — that enables cross-site comparison, while allowing additional customisation for genuinely site-specific risk environments. Local flexibility is fine; losing comparability is not.

How does Australian WHS law affect multi-site safety monitoring programmes?

Australia's WHS framework is primarily state and territory-based, meaning organisations operating across multiple jurisdictions may face different regulatory priorities and enforcement activities depending on location. Most jurisdictions have adopted the model WHS laws, providing broadly harmonised obligations, but enforcement practices and specific legislation (such as industrial manslaughter laws in NSW, Queensland, and Victoria) vary. Consistent monitoring and documentation across all sites simplifies compliance across multiple jurisdictions.

How do you benchmark safety performance across sites fairly?

The most effective approach is to benchmark on rate of change and improvement rather than absolute event counts. Sites with more complex layouts or higher-risk operations will naturally produce more detected events than lower-risk facilities. What matters is the trend: are event rates falling, is coaching activity occurring, and are serious event categories being addressed? Absolute comparisons without accounting for operational context can drive the wrong behaviours.

What role does coaching play in a multi-site AI safety programme?

Coaching is the mechanism through which detected events translate into sustained behaviour change. Technology provides the detection and evidence; coaching provides the human conversation that changes how people work. The organisations that see the strongest long-term results at scale treat coaching as a core operating rhythm, with standardised workflows and enterprise-level visibility into whether coaching conversations are actually happening across all sites.