Three risks to AI business cases (and how to solve them)

I recently spoke at the Foundry CIO50 conference in Sydney about three key risks that Enterprise AI deployments face. This is an edited version of my speech.

At small scale, AI pilots often work well. It’s when those pilots become mission critical workflows, and gather scale, that cracks can emerge.

Every AI business case considers models, licences, support and data privacy. What is often left unexamined is the assumption is that the underlying infrastructure will cope.

This includes networks. According to Cisco's 2025 AI Readiness Index, 71 per cent of AI "pacesetters" – companies achieving measurable success with AI – say their networks are ready to scale for any AI project, against 15 per cent of companies overall.

So here are three questions I recommend adding into your AI projects’ business cases.

How will the pilot scale to production volumes?

A pilot might have 100 users, one model and one cloud. That’s a low volume of traffic. But the issue isn’t typically agent-to-user traffic, which is often just text.

The problem emerges when agents interact with agents, at machine speed, transferring and accessing significant volumes of company data in the process. They could be referring to company knowledgebases or analysing medical image sets gigabytes in size. 

This data has to move in milliseconds to deliver a great experience and feel natural to the eventual customer on the end of that workflow.

Suddenly the spare network capacity you had overnight for offsite backups may not be there any more – or it might be less predictably available.

AI also multiplies dependencies. It is fantastic at drawing disparate systems and data together and making sense of it.

However, one customer request handled by an agent might rely on five or ten different systems. If one of those systems stalls, the whole workflow may fail – or worse, finish, but be based on incomplete information.

Network infrastructure links those systems together, and it needs a much longer runway to scale than software. Fibre builds can take months or years.

It's why Vocus is building a ducted Sydney to Melbourne fibre route, due in 2029, so new fibre can be pulled through the duct without digging up 1,200 kilometres again – because land access and permits can’t be sped up. 

A ducted route also provides better protection to the fibre, so that terabits of traffic flowing between hyperscaler data centres in the capitals will be much less likely to be disrupted by a farmer hammering in a star picket.

Since my speech, we’ve also announced a new fibre build from Brisbane to Darwin, off the back of significant customer demand. 

So ask: if our best AI pilot grew tenfold tomorrow, which infrastructure assumption could break first?

Will your AI success be delayed by paperwork?

Developers are used to being able to spin up a cloud virtual machine in seconds. However, with traditional telcos, it’s still normal for physical network changes to take an email, an order form and days or weeks of lead time.

Vocus is moving toward a highly automated deployment process for our network. We’re doing this for consistency, not just speed. Uptime Institute found 85 per cent of major outages involving human error came from staff not following procedures, or from flawed procedures.

This matters because the AI agents can work at speed if they can link into APIs, but if they’re not available, they’re not going to call your telco account manager to chase a ticket.

So ask your provider (including us): what can my team see and change directly, and what still waits on a person?

Does your AI deployment have single points of failure?

"Our data stays in Australia" is only the start of the conversation. Boards will want to know where prompts are processed, where models run and what physical path connects everything.

Resilience needs the same scrutiny, particularly for banks, insurers and super funds bound by APRA's CPS 230 risk standard. Two "redundant" links may share a conduit or a building entry, so ask where the paths differ and what they still share.

Vocus designed our national network as a figure of eight – literally a giant ‘8’ lying on its side with a loop around the west of the country and a loop around the east. If one of our core network cables is cut, most data traffic can be rerouted the other way around the loop.

Whatever the provider, the test is simple: can you draw one page showing where my network path runs and what happens if part of it fails?

If you can answer those three questions, your business case is built for production, not just the pilot.