Are we bringing old problems into AI?

Originally published in AI in Practice on LinkedIn on 11 October 2026.

It’s Sunday evening, I’m putting the finishing touches on this, and I’m getting flashbacks to the early 2010s again.

Conversations about digital transformation and change management… some of which I’m now having with businesses bringing in AI. You might be heading into another one yourself this week.

Before we accelerate another AI rollout, can we have the awkward conversation about how the work actually gets done?

If a request already bounces between departments because nobody knows who can approve it, that deserves attention. If staff only get a say once the important decisions have been made, that does too.

And now we’re asking everyone to work faster. Apparently those old problems are coming along for the ride 😅

We want people to get more done, so let’s sort out what’s already getting in their way. If we bring the same unresolved decisions and messy handovers into an AI project, we risk ending up in the same damn spot, with higher expectations this time. We can tackle those problems alongside delivery.

Let me walk you through how I’d check whether this is happening in your business. We’ll follow one piece of work, look at what’s holding it up and work out which checks address a real risk. Then we can look at what could change with the people and systems you already have, and where you need more backing.

Follow one request

Take a recent request to use AI for an actual task. Did it reach someone who could make a decision, or keep moving between teams while the person explained it again? If it stalled, find out what was missing and who could resolve it.

There may be a good reason for the wait. Access to sensitive information, for example, deserves proper consideration. But if several people have reviewed the request and nobody can explain what happens next, I’d be getting them together to sort that out.

Then speak to the person who raised it. What happened to the work while they waited? Did they carry on manually, abandon the idea, or find another way to get it done? The approval queue may not show what they did next. If the task involved a customer, check whether the delay reached them too.

I’m seeing people finish tasks using personal AI accounts or tools the business hasn’t assessed. This is part of the shadow AI problem: AI use outside the organisation’s approved systems and processes.

The motivation may be to get the job done. The consequences can still be serious. Company or customer information may end up in a service the business hasn’t assessed, with less visibility over how it is stored or used.

In a September blog post, the UK’s National Cyber Security Centre warned that where security policies cannot meet business needs, employees are likely to keep adopting AI services before the organisation has assessed them and provided approved alternatives. The post recommends finding out what drives that use, so organisations can assess the risks and offer safer options. NCSC: The hidden risks of shadow AI

A difficult process may be one contributor to shadow AI. It doesn’t excuse bypassing safeguards. But if your response is another form and another approval, with the same delivery deadline attached, check whether you’ve addressed what drove the workaround.

The safeguards matter. Australia’s privacy regulator recommends, as best practice, that organisations avoid entering personal information, particularly sensitive information, into publicly available generative AI tools. Making work easier needs to happen within protections like these. OAIC guidance on commercially available AI products

Bring the team into it

Now look at who helped design the approved way. Were employees involved while there was still room to change it, or were they handed a finished process and some training? Ask them to walk through a real task with the people responsible for delivery and risk. You may discover that a requirement that looks reasonable to one team duplicates something another team already checks.

If you have a capable program manager and change support, bring them into this early. But you might be doing this without either, and still be expected to deliver the improvement. That’s a real constraint, and it needs to be visible in the plan.

Someone who’s good with an AI tool can help colleagues learn it. But asking them to manage the change and support everyone alongside their own job is a lot to load onto one person. Give them time and backing, with clear limits. Employees should have a say in how AI changes their work. That doesn’t make them responsible for resourcing it.

If that support isn’t available, let’s get specific about the gap. I’d take one affected task, show where it’s getting stuck and put that evidence in front of whoever can make the call on resources or scope. Then there’s something concrete to discuss about what’s realistic with the people available.

Where you can, work through this in the project meetings and change process you already have. The fix might be removing a duplicate approval or clarifying who can decide. An existing assessment may cover the proposed use, provided the relevant circumstances and risks haven’t materially changed. Where an additional control is needed, explain what it protects and how people are meant to work within it.

Keep moving where it’s safe

We don’t need to wait for the perfect moment. But “embrace uncertainty” shouldn’t become a convenient way to wave through problems we already know about. If leadership wants the team to keep moving, that needs to come with backing to resolve those problems. We can learn as we go. We can also be clear about what we already know needs fixing, who owns it and what can safely continue in the meantime.

That might mean easing off the accelerator where something needs to catch up. If affected work must pause, give someone responsibility for resolving the issue and agree the conditions for restarting. Necessary safeguards need to be ready before that use goes live.

Check how much AI the task actually needs. A system you already pay for might route form enquiries using fixed rules. AI could suggest categories for free-text enquiries, with the person handling them checking against agreed categories before routing. Compare the options, including model usage costs and time spent on review and corrections. We don’t have to give AI the whole job just because it can attempt it.

And be a bit wary of the AI cowboys pushing a wrapper or a Hermes Agent or OpenClaw setup before they’ve understood your business. Ask what problem they’ve actually diagnosed. “Look what it can do” only gets us so far.

Those tools can have a place. Do the discovery and make sense of what it tells you about the business and existing systems. Then show what the proposed approach improves and what it adds in cost and support.

If you’re reading this and thinking, “Yes, but I’m the one trying to hold all of that together,” this is work I can help with.

I work alongside leaders to understand what’s getting in the way and find where AI earns its place and where it doesn’t. We test the approach with your team, put agreed changes into practice and work out what support you need before committing to a permanent role or bigger program.

I usually recommend an initial three-month engagement, with an agreed remit and review points. A focused one-month engagement can suit a narrower question. We’d agree what I’ll own and what needs to come from your team.

If that sounds like where you are, get in touch. We can start with one piece of work that’s getting stuck.