If AI is transforming your business, it should show up on the P&L
Here is the question I often ask leadership teams I work with: “If AI is genuinely transforming your business, where does that show up on the P&L?” Very often the honest answer is that it doesn't show. AI spend lands on the cost line as more activity, not lower cost, and never reaches the bottom line.
Revenue is rarely touched, because few organisations are using AI to change what they sell or how they grow. Gross margin sees pilots rather than systemic change, and operating expenditure, where nearly all the activity actually sits, moves in volume, not in cost. Effort is up, output is up, yet profit and revenue is unchanged.
That gap between activity and impact says less about how we measure AI, and more about how we've rolled it out in the first place.
The problem we haven't framed properly
Most organisations didn't really “roll out” AI. In most scenarios, teams were told to adopt tools, yet very few were given the training, context or process redesign needed to use those tools well. The result is a familiar pattern of more output, produced faster, with no corresponding improvement in quality or impact. Researchers have started calling this "workslop" - work that looks finished but creates more effort downstream than it saves. It's also, in P&L terms, activity that shows up as cost and volume, never as margin or revenue.
I'd argue this comes down to three things. There's been an unspoken incentive to label everything AI, whether or not it changes the outcome, because being seen to adopt matters more than proving it works. Teams have used AI to speed up tasks that shouldn't have existed in the first place, using automation to paper over work that needed to be eliminated, not accelerated. And most importantly, AI has largely been built by engineers, handed to non-technical teams, and never accompanied by real education on when and how to use it well.
Automating mediocrity
I see this mostly in workflow and platform-based automation projects. When a project management or approvals platform is already embedded across an organisation, the instinct is to add AI onto the existing process, automating the brief intake, the review stage, and the handoffs. But if that process was inefficient to begin with, AI doesn't fix it and just runs the same broken sequence faster and at a larger scale.The issue isn’t the platforms themselves, but the operating model around them. Before any organisation applies AI to a workflow, it needs to ask whether this process still makes sense, and whether it is structured around how people actually work. Skip that step, and automation hard-wires inefficiency into the system. In simple words, don't automate the mess. Understand the process, remove what shouldn't be there, and only then apply automation and AI on top.
Built for engineers, handed to everyone else
Part of the reason this keeps happening is that AI tools have been designed, tested and refined by engineers, for engineers. That's the discipline that built them. But it means the tools have been optimised for capability rather than adoption, for what they can do rather than why a marketer, an HR lead or a project manager should use them, and how.
We've effectively taken a piece of specialist engineering, made it available to everyone, and assumed the instructions would take care of themselves.
Also Read: Beyond the Click: Why 2027's Winners Will Train Their Ads on Real Sales
Closing that gap means asking two separate questions before any rollout:
- What can this technology do, and
- Why should this particular team be using it?
The first question is easy and gets most of the attention. The second is more challenging to answer, and more important - and it's often the one that organisations skip.
The trust problem, on both sides
This has consequences for trust in two directions. Internally, staff who were told to adopt AI quickly, without guidance, pause and rethink how they're using it, while the pressure to show efficiencies, prove ROI and reduce costs carries on regardless. I think leadership needs to talk about this honestly. If it comes across as a sudden reversal, people will stop trusting the direction they're being given, and rightly so. What we're actually doing is checking that the push we asked for is paying off, and admitting that the first round of guidance didn't go far enough. Framed this way, it builds trust rather than costing it.
Externally, the same dynamic plays out with audiences. Marketing and communications teams have been producing at a volume that AI makes possible, without always pausing to check whether that output is actually landing. Engineering teams are already a few steps ahead here, auditing what generative tools have produced and finding that a meaningful share of it doesn't hold up. Marketing and brand functions are only just beginning to do the same. When they do, I expect many will find that more content hasn't meant more trust with their audiences - it has often meant less.
What responsible adoption looks like
Rather than an argument against AI, this is an argument for sequencing. Before automating anything, organisations should identify which tasks and processes should exist at all and remove the ones that shouldn't. Only once that's done should they look at automation, and only once a process is genuinely optimised should AI be layered on top, with the people using it properly trained and clear on the purpose it serves.
Sequencing is what determines whether the four lines on the P&L actually move: whether cost reflects less spend rather than more activity, whether revenue is touched at all, whether margin moves beyond pilot stage, and whether OpEx reflects efficiency rather than volume. Getting it right means that AI stops being a cost line that grows in step with effort, and becomes something a finance leader can actually point to.
