Beyond Incremental Gains: Rethinking Agentic AI's Impact

IBM wants agentic AI to stop tinkering at the margins and start delivering “net‑new impact.” I'll be honest — that’s an ambitious repositioning. Funny thing is, it sounds less like a product roadmap and more like a mandate for rewriting corporate purpose.

Let’s give the pitch its due. The article is right that incremental gains only get you so far; shaving a few basis points off costs doesn’t justify the organizational chaos that AI can unleash. Framing agentic AI as a route to new capabilities — not just faster PowerPoints — is a useful provocation for executives who’ve been hiding behind pilots for a year.

Yeah, no, the trouble starts when “net‑new” becomes a productized promise.

The piece makes a simple, seductive claim: agentic AI can go beyond incremental gains and unlock outcomes firms couldn't get before. That's an appealing marketing line for any vendor with deep pockets and a services arm — IBM fits that profile. But moving from marginal efficiency wins to genuinely new capabilities depends less on the model and more on the messy scaffolding around it: data plumbing, reward design, human workflows, vendor lock‑in, regulatory comfort. You can call an agent "strategic" all you want; if it can't plug into procurement, risk, and frontline ops without exploding, the "net‑new" ends up being new problems.

One reason IBM's framing is clever is that it sells transformation as an engineering problem you can solve with consulting. That aligns incentives: if you promise net‑new impact, customers sign up for big transformation projects and the services revenue flows. But incentives matter; companies will optimize toward demonstrable ROI, which often means rolling back to incremental, safer deployments when something unpredictable emerges. So the claim to ascendancy masks a tug‑of‑war between ambition and the institutional impulse to minimize disruption.

Neuromancer had its Sprawl; the enterprise has governance.

And governance here is not a bolt‑on compliance slide at the end of the deck — it’s the main event. The article treats measurement of "net‑new impact" as a destination. That’s convenient, but measuring novelty is philosophically and technically thorny. What counts as new: faster throughput, a process that previously required human discretion, or a revenue stream that literally didn't exist? Those are different standards. If you demand all three, adoption slows; if you accept fuzzy proxies, you risk downstream harm that isn’t captured in dashboards. IBM can help design the metrics — of course it can — but the more you tailor metrics to prove the case, the more you risk masking negative externalities.

There’s also an operational truth: agentic systems act with some degree of autonomy. They make decisions, propose actions, trigger chains of activity. That's attractive when the cost of human oversight is high; it's dangerous when the cost of failure is systemic. So governance can't be a checkbox module sold alongside consulting hours. It has to be baked into reward functions, audit trails, rollback mechanisms and, critically, cultural expectations about error. That’s expensive, and it dilutes the glossy narrative of quick, net‑new wins.

Look, we’ve seen this movie before. When cloud first arrived, vendors sold “digital transformation,” but most enterprises started with “please just lift‑and‑shift our mess into someone else’s data center.” Only once pricing models, security postures, and org charts adjusted did you see genuinely new things — think of how Netflix used cloud to reinvent release cycles, not just hosting. The lesson: infrastructure changes fast; incentive structures crawl.

That’s why IBM’s strategic repositioning is really a culture test, not a product demo.

If IBM is pivoting messaging from incremental gains to net‑new impact, the company is also asking its customers — and itself — to change how they allocate budgets, hire for skills, and tolerate experimentation. That's harder than it looks. Enterprises have incentives to optimize current systems; compensation plans, audit cycles, procurement rules all favor reliability over novelty. To make agentic AI genuinely strategic, IBM would need to help clients rewire incentives — not just install models. That’s long‑term, bespoke work, and it exposes IBM to blame when experiments fail.

A quiet irony here is that genuine “net‑new” often starts in corners that look trivial on paper. Customer support teams hacking together agent workflows, finance analysts using agents to simulate scenarios, operations leads letting agents negotiate micro‑decisions in logistics — these pockets of experimentation don’t fit neatly into a top‑down “strategic ascent” narrative, but they’re where organizations actually change. By the time the board declares victory, the culture shift started two levels down.

Counter‑argument time: defenders will say agentic AI already automates complex tasks and creates new business models; selling the story accelerates adoption, which is good. True — early adopters will find places where autonomous agents create real, unforeseen value. The caveat is this: those successes will be highly contextual and often require human organizations to adapt in ways marketers rarely admit. Selling a universal "net‑new" promise risks overreach.

So what's the sensible way forward? Treat "net‑new impact" as a research program, not a slogan. Pilot, instrument, and iterate in settings where missteps are contained; focus on composability so agents can be limited or replaced without a full systems rebuild; price outcomes in ways that share risk between vendor and customer. That reduces the sexy headline but increases the odds of durable change.

IBM’s article plants a flag around strategic ascent; the real test will be whether, three budget cycles from now, “net‑new” shows up in how companies redesign their org charts, not just in how they rename their AI roadmap slides.