Last week I posted the DIY AI Arc, the 11-step rollout we're seeing most Kiwi mid-market businesses running.
A customer yesterday confirmed what we're seeing:
"It's exactly the journey we're on. We're rolling out business unit by business unit, everyone's having those aha moments and sharing use cases within finance, those little wins. Mainly individual or within team. But what's beyond that? And is the opportunity beyond that for all businesses, or do some not have as much AI opportunity?"
That came from a customer in a strategy role, and the answer for me splits into two paths.
The first asks what exists beyond individual and team-level wins?
The second asks whether the ceiling is a capability problem, which is fixable, or a property of the business itself?
For anyone who missed the original post, the arc runs something like this.
Everyone gets a Copilot licence and some training, and the rollout is declared a success.
People have private "ohh sh!#" moments, usually alone at a desk with a meeting transcript.
Work looks cleaner and tidier, because AI amplifies each person's existing strengths.
Rivalry kicks in, and people show off outputs without sharing how they made them.
Everyone feels more productive, while the work inside the business looks exactly the same.
One enthusiast, trained by hours of YouTube, becomes the de facto AI lead.
Their builds work for one person, but not for a business's governance model or multi-player mode (can more than one person partake in using the AI collaboratively with another?).
People paste client data into personal AI tools, because the sanctioned ones frustrate them.
Execs ask why all the spend on AI has not moved any numbers on the P&L.
IT is asked to productionise the builds, finds them incomplete, and pumps the brakes on further AI work.
An exec, sold by LinkedIn or X on what is possible, goes looking for outside help.
Most Kiwi businesses are in the experimental phase of that arc right now, somewhere between steps one and five = wins are staying individual.
The businesses that started a year earlier are now arriving on calls at steps nine to eleven, at the ceiling, asking for help.
The ceiling they're hitting is not an unfortunate ending, it is the built-in result of a single-player rollout. McKinsey & Company found 9 of 10 companies now have AI in at least one function while 94% report no significant value. That is the ceiling I'm referring to.
The signs the ceiling is close are consistent across the calls I sit on:
Wins get shared as screenshots of outputs, hardly ever as the method that produced them.
One person is the answer to every "how does this work" question, and the builds stall when that person is on leave.
Crucially, each build works for one person but hasn't been built with the capability in the platform for more people to partake (multiplayer mode), so the bottlenecks between people to getting the work done still exist. The time to get from bottleneck to bottleneck has just sped up.
Client data is finding its way into personal AI accounts, because the sanctioned tools frustrate people.
Nobody can point to how the work has been redesigned in the business, or the ROI from it. AI looks like a cost centre, nothing more.
Before answering what's beyond the plateau, let's outline why the plateau is a problem, because the individual wins do exist and it'd be silly to dismiss them as wasteful.
The problem is competitive rather than operational. Every competitor can buy the same licences tomorrow and run the same training the week after, so individual AI proficiency is on its way to becoming table stakes, the way building a website, using email, and rolling out Microsoft Office did. Within a couple of years it will be assumed in every hire that you proficiently use AI in every activity, rather than an advantage. The trap is that the spending feels like progress right up until step nine when the CFO and CEO ask for the return on the investment.
Competition then does to these gains what it always does to efficiency gains, it passes them through to customers.
Everyone quotes a little faster, everyone answers email a little sooner, and the floor of industry performance resets while the ceiling of the business and its potential stays the same.
The following question should be at the tip of every leader's tongue as they join their next Lead Team meeting about their AI investment: what is the point of additional spend on something that only gets the business to the starting blocks, if it cannot outperform the market from there?
A rollout that stops at individual proficiency gets them to the game, but doesn't add anything beyond that once the market catches up.
There is an additional layer to this worth calling out. And it's because for the first time the capability and productivity gap between a small business and a large business has closed.
A 200-person manufacturer in Hamilton has access to the same model output as the best-resourced technology company in the world, and what separates them is not the technology available, it is whether the work has been redesigned to use it.
Previous technology waves rewarded businesses with a healthy balance sheet to make investments, whereas the AI wave rewards leadership ambition, decisiveness to act, and decision making, which is the one thing a smaller business can make faster.
So, to the first question, the answer to what sits beyond the plateau is the shift from single-player to multiplayer AI. Not a bigger rollout, but a different unit of work design.
The challenge is that an aha moment lives inside one person's chat window. The value of a business is created between people. It's when a quote crosses sales, operations and finance, or an insurance claim that crosses approvals in three functions.
While every win stays personal, the work coming out of the business looks the same. This is the concept I mentioned above in action, the bottleneck is that work between people is still dependent on the human moving the work through the pipeline. They might just do it faster, but inevitably that slowdown still exists and is inflamed with more people in the loop.
The map I use with clients resembles some version of the following, and each level redesigns a bigger unit of work.
Task augmentation is AI helping one person do one task faster, so research that took 45 minutes takes 5 minutes. Training lands people here, and 50% of businesses stop here.
Workflow automationputs AI inside steps of the existing process, while the sequence, the handoffs and the waiting all stay, so everything runs faster and nothing changes shape. The other 50% of business stop here.
Work redesign changes the question to how the work would happen if the outcome were designed today around AI (not AI bolted onto the workflow). Account management stops being periodic check-ins and becomes continuous monitoring, with a person stepping in where judgement or the relationship is required. Half the old steps stop existing, because they only existed because humans drove them, all based on outdated thinking.
Role redesign follows once the process is rebuilt, because jobs shift from doing the work to owning the system that does the work. The design question becomes what the human is now responsible for.
Operating model redesign reorganises the function around its rebuilt processes rather than its old departments (this is full multiplayer mode).
Organisational redesign reorganises the whole business around what the previous five levels made possible.
The aha moments in that quote earlier in this article are level one, and the plateau he can now feel is the roof of level two. The P&L starts shifting at level three, because that is the first level where the work changes rather than the speed of the work. Not automation of the old sequence, but reinvention of it.
Being super clear, there isn't a business at level six (organisational redesign) that I'm aware of. Maybe some West Coast business (Shopify? Every?), or an AI-native business by design. It's more a vision than a realistic opportunity for most.
To bring this to life, I'll use my favourite analogy from history.
When factories swapped steam engines for electric motors, most owners kept the same floor plan, with machines still crowded along the line where the drive shaft used to run.
Productivity barely shifted for close to thirty years.
The winners rebuilt the factory around the motor, arranging machines in the order the work flowed, because each machine could now carry its own power.
The gain was never in the motor itself, it was in the redesign the motor made possible, and levels three to six of the ladder are that redesign in modern form.
The rebuild also needs a destination, because "redesign the work" without a target organisation in mind produces random acts of automation.
The target is an operating model rather than an org chart with copilots. Intelligence sits at the core and runs the repeatable work end to end, while humans work at the edge, sometimes starting the work, guarding it and steering it. Judgement, accountability and relationship stay human by design, we're all well aware of this now.
Context (or intelligence) accumulates with every cycle, so the organisation gets smarter in the system instead of the knowledge leaving in people's heads at the end of the day.
One economic marker to monitor to watch is whether revenue decouples from headcount.
If revenue per head is rising each quarter, then the operating model changed.
If it is flat, the business added tools and kept its old shape.
Getting to that organisation is less glamorous than the big AI labs would have people think (despite all of them investing in or buying their own enterprise level consulting firms to support implementation of AI), and it rests on three rather boring foundations.
Leadership mapping: leaders who understand what the technology makes possible and can lay that over their own business, function by function, rather than delegating the question to the AI enthusiast who's spent enough time on YouTube to sound smart.
The house in order: data the systems can trust, governance that defines what AI is allowed to touch and who signs off, and operations management that treats every AI system as an operational asset with an owner, a baseline and a target metric.
Sequencing: the discipline of proving one workflow before starting the next, which the rest of this write up comes back to.
None of these arrive with a licence rollout, and all of them are prerequisites for anything above level two, or Workflow automation.
The harder truth sits underneath all of this, and it is the part I see most often from the inside. Most businesses cannot yet tell a relevant opportunity from an attractive one. A relevant opportunity crosses several desks, carries a number the exec team already watches, and sits at the weakest essential step of a workflow, the step that caps the return of everything around it. An attractive opportunity demos well in a lunch-and-learn and changes nothing downstream. Choosing between the two is where the money is won or lost, and it takes genuine expertise, because the choice depends on understanding both the technology's jagged capability and the anatomy of the business's own workflows.
The same expertise gap shows up in who gets asked to do the work. People who learned to build agents in their spare time are now being asked to architect and operate multi-player systems, with governance, permissions and uptime expectations attached.
Building a single-player agent and operating a production system that a whole team depends on are different disciplines, in the same way that owning a ute and running a freight company are.
it's not a failure of the AI enthusiast either. It's more that the leadership direction wasn't right, because the rollout model assumed productivity would grow into engineering an organisational redesign (or maybe the leaders weren't aware of how that works).
To be fair to the enthusiasts, the answer is a pairing rather than a swap. The knowledge that makes any system reliable lives with the people who run the process today, because they know which customer always needs a manual override and which exception turns up give scenarios.
The discipline that turns that knowledge into a governed, multi-player system is what most businesses need to bring in. Asking either side to do the other's job is how step ten of the FI AI Arc happens.
That backdrop makes the second question easier to answer.
The ceiling NZ businesses are hitting right now is a design ceiling. Every business reaches it at roughly the same height because it is produced by the same rollout model of licences, training and hope, rather than by anything about the AI industry.
To ground ourselves in a reality for a sec, the size of the prize does differ between businesses.
A firm heavy in codified, repeatable knowledge work holds more level-three (**Work redesign)**opportunity than one whose value is mostly physical. What does not differ is the location of the current ceiling, because no business's structural limit sits at the aha-moment (level one, **Task augmentation)**stage.
The part most teams skip is the arc does not have to be completed before it can be exited.
The customer's instinct to move business unit by business unit is right.
Rolling out licences unit by unit reproduces all eleven steps of the arc inside every unit.
Rolling out redesigned workflows unit by unit skips the middle of the arc entirely, because governance, shared context and the multiplayer design arrive before the frustration does.
What cannot be skipped are the foundations, because every attempt from level one to level five without the house in order has failed. Plus the business needs to build the muscle of experimentation to determine where AI works or doesn't work in their business, and identify where the gaps are in their org. Here's a low risk way to do so.
In practice that means picking the one workflow in one unit that crosses multiple teams, and already has a metric that the exec team can monitor for change.
The businesses that lead with AI are not better resourced or luckier with their industry. They accepted earlier that individual capability is the entry fee to the game rather than the prize, and they changed the unit of design from the person to the workflow.
Passionate about all things AI, emerging tech and start-ups, Mike is the Founder of The AI Corner.
Subscribe to The AI Corner
