Most organisations are treating AI like someone just handed them a thousand brilliant new employees and said "make use of these." What would they all do? Where would you even start?
That framing surfaced during our first Business Services Challenge Forum this week, and it resonated with everyone in the room. We'd brought together operations leaders from legal services, software, consulting, and compliance to explore a shared question: what's actually working with AI in operations?
The conversation revealed a pattern. Most leaders are seeing results, but not at the pace or scale they'd anticipated. What emerged wasn't a story about AI failing, but about organisations approaching deployment in ways that make scale difficult.
The dual challenge
The discussion surfaced two parallel constraints. Change management (getting people on board, winning hearts and minds) takes six months to a year. Every leader in the room recognised this timeline.
But the conversation kept returning to a second challenge that often gets less attention: infrastructure. Not the AI platforms themselves, which are readily available. The constraint is connecting them to legacy systems, fragmented data, and existing technology estates.
As one participant explained, you can access AI libraries, tailor them, and orchestrate them relatively easily. But integrating all of that with current infrastructure? That's where deployment slows significantly.
"Get the infrastructure right and adoption becomes achievable. Skip that foundation, and you're automating on quicksand."
This creates a compounding problem. Organisations need patience for people to adapt, but their technical foundations aren't ready for what they're trying to build. You can't separate these challenges. Get the infrastructure right and adoption becomes achievable. Skip that foundation, and you're automating on quicksand.
Think like a pizza shop
So what's working? The room converged on an uncomfortable truth: most organisations are trying to do too much.
The pizza shop analogy came from Joe Deeley at Enate, and it captured what several participants had been circling around. Customer asks for pepperoni, you make pepperoni. Not "the most delicious pizza possible" but something specific, repeatable, governable. When you embed a focused purpose within an existing process and let that prove itself, you get trust, compliance, and measurable value.
The generic AI assistant with a thousand purposes? It rarely delivers.
"One success becomes torch paper for the next project. You earn the right to scale by proving you can execute."
This isn't just about starting small. It's about staying small long enough to learn. Boards want rapid outcomes, but the path runs through controlled wins. One success becomes, as someone put it, "torch paper for the next project." You earn the right to scale by proving you can execute.
The measurement trap
The second challenge we explored was harder: how do you measure value when the technology is fundamentally probabilistic?
Past technology cycles offered clearer ROI calculations. ERP implementations, cloud migrations: you could model the returns. AI doesn't work that way. Its outputs vary. Its impact is diffuse. And the traditional horizontal scaling approach (roll it out across the organisation) tends to fail. Participants noted that pilot success rarely translates to enterprise scale, with failure rates commonly cited in the 50-80% range.
What's emerging instead is a vertical approach. Go deep into one function. Use the savings to fund the next project. Build a staircase rather than a ramp. Some private equity firms are applying this across portfolio companies: pick a function, prove it works, replicate.
The implication for services businesses is significant. You can't sell transformation; you have to sell sequences. And you need to be honest with clients about how long real adoption takes. Winning hearts and minds isn't a quarter. It's six months to a year. By which point, the technology has probably changed again.
What we're taking away
Three patterns emerged from the discussion:
First, both change management and infrastructure need attention. Organisations need patience for people and foundations ready for technology. Addressing only one creates obstacles for the other.
Second, specificity outperforms ambition. A narrowly scoped agent that delivers measurable results is more valuable than a general-purpose assistant that disappoints.
Third, success creates permission to scale. Leaders advised starting where you can prove value quickly, then using those wins to secure resources for adjacent implementations.
The Business Services Challenge Forum will continue these conversations in the new year. If you're an operations leader navigating similar challenges, the community is growing. We bring together senior practitioners from different sectors to share real experiences and practical approaches to common problems.
Thanks to everyone who participated in this first session, and to Enate for sponsoring.