AI is everywhere, yet productivity is flat: The AIOps paradox

Most companies are “using AI” and still can’t point to higher productivity or fewer jobs. That is not a hot take, it is what large executive surveys are reporting. One widely cited survey of roughly 6,000 executives across the US, UK, Germany, and Australia found about 90% said AI had no noticeable impact on productivity or headcount over the last three years (reported in outlets like Fortune, ITPro, and The Register).

So the puzzle is not “why aren’t companies trying AI,” because estimates put enterprise AI adoption around 70 to 78% by 2025. The puzzle is why the scoreboard is not moving.

The AI productivity paradox

This is the AI productivity paradox: lots of activity, not much measurable output. If you have lived through any prior tech wave, this should feel uncomfortably familiar. Economists have a name for it, the Solow paradox, which basically means “you can see the new tech everywhere except in the productivity stats.”

The important part is the why. AI is showing real gains, but mostly in narrow, routine tasks where you can measure the before and after cleanly. Research on generative AI in customer support is the cleanest example, where response suggestions can increase throughput and quality, especially for less experienced agents. Those wins are real, but they do not automatically translate into firm wide productivity, because most of the enterprise is not a call center.

How to transform ITOps into AIOps

Now zoom into IT operations, where the sales pitch for AIOps is basically a dream scenario: fewer useless alerts, faster incident response, less 2am chaos.

AIOps (AI for IT operations) means using machine learning or language models to:

• help detect issues,
• group alerts,
• suggest causes,
• trigger fixes... and more!

The paradox shows up here too. AIOps pilots often look promising, then stall in the messy middle where data quality, tool sprawl, and organizational inertia live. The most common blocker is not model quality, it is inputs and plumbing.

If your incident data is fragmented across ticketing, chat, dashboards, and tribal knowledge, the AI cannot learn consistent patterns, and you cannot trust its recommendations. If your alert stream is 80% duplicates and misrouted noise, the AI will confidently improve the wrong thing.

Keep humans in the loop

If nobody agrees on what “good” looks like, you get demo wins and production confusion. This is why governance ends up being the boring hero. Human in the loop is not a compliance box, it is how you keep automation from quietly turning into a new failure mode.

The research takeaway is blunt: teams that treat AIOps like a product feature tend to get disappointed, and teams that treat it like an operating model change tend to get compounding returns. If you want a simple mental model, think of AI as a force multiplier for already healthy processes, not a substitute for them. That means incident management basics matter more than ever, like clear severity definitions, clean postmortems, and consistent tagging of root cause categories.

How to track AIOps success

Successful introduction of AI also includes choosing the right KPIs (key performance indicators) so you can tell the difference between “the AI is busy” and “the business is better.” Three practical KPIs usually beat twenty aspirational ones:

1. Alert volume per on call shift, because reducing noise is the first find for human attention.
2. Time to acknowledge and time to mitigate, because speed matters when customers are waiting.
3. Change failure rate, because automation that increases incident frequency is negative productivity.

Get the C-suite on board the AI train

There is an extra twist that executives should not ignore. The same survey work suggests many senior leaders are light users of AI themselves, with a meaningful slice of the C-suite not using it at all and most using it less than two hours per week. That matters because enterprise transformation does not happen through memos, it happens through habits. If leadership treats AI like “a thing the teams do,” the organization will tend to cap AI at local optimizations instead of redesigning workflows end to end.

This is also why employment effects look muted so far. If AI is mostly improving task level throughput inside existing roles, you will see output quality changes before you see headcount changes. You should also expect redistribution before reduction, meaning the work shifts to different tasks, different teams, and different expectations.

Select AI tools built for your reality

So what should a corporate leader, journalist, or analyst do with all this? Stop asking “did you deploy AI,” and start asking “which workflows changed, and what metric moved.” Ask whether AI is being applied to a process with clean inputs, clear ownership, and a measurable definition of success. Ask whether the organization invested in integration work (data, permissions, connectors), because that is where most AIOps projects quietly die. And ask whether there is a safety model, meaning human approval for high impact actions and audit trails that survive the inevitable incident.

Tools can help here, but only if they are built to respect reality. For example, in Splunk operations, platforms like Deslicer AI focus on context aware automation that inspects live configurations and keeps humans in control, which is exactly the posture research implies is needed for safe, measurable gains. If you had to pick one takeaway, it is this. AI is not failing, but enterprise productivity gains are being bottlenecked by the unglamorous work of process change, data quality, and governance.

Where are you seeing the paradox most clearly right now, in customer facing work, back office workflows, or IT operations? Contact us to discuss your challenges and explore if Deslicer AI could benefit your organization.