Generative AI use cases examples

The most revealing number in generative AI right now is not a productivity gain. It’s a 43-percentage-point perception gap — and the story behind it is the kind of thing that should make any engineering leader uncomfortable.

In July 2025, METR researchers published a randomized controlled trial on arXiv: 16 experienced open-source developers, 246 real tasks, with and without AI tools. Not new developers. Not toy projects. These were people who’d averaged five years on their specific repositories — codebases with a median of 1,500 commits, active for an average of 10 years, exceeding one million lines of code.

Before the study started, developers predicted AI would make them 24% faster. After finishing, they believed it had made them 20% faster. Actual wall-clock time showed they were 19% slower. Same people. Same tools — Cursor Pro with Claude 3.5/3.7 Sonnet. Same production tasks.

The tools felt helpful. They were not.

Here’s what makes this finding so uncomfortable to sit with: the failure mode is completely invisible from the outside. The developer feels faster. Their manager sees more commits. The tool’s dashboard shows high usage. Every signal says it’s working — except the one that actually matters, which is how long it took to ship something that holds up in production.

For the record: METR’s February 2026 follow-up note does acknowledge that developers are “likely more sped up from AI tools now” than the original study captured, partly due to improved model capabilities. But because of selection effects in their expanded design, the researchers explicitly state their new data provides “only very weak evidence for the size of this increase.” The productivity paradox has likely narrowed. It has not been overturned for the specific deployment scenario that produced it.

Three years of accelerating adoption have produced a market where McKinsey’s 2025 State of AI survey found 78% of organizations using AI in at least one business function — up from 55% just two years prior. That figure is self-reported and skews toward larger organizations; the survey doesn’t define a minimum deployment threshold, so it captures everything from one employee using ChatGPT to enterprise-wide integration. Worth knowing before you cite it.

The headline masks a structural split that matters a lot more. Only about a third of those organizations have scaled AI across workflows. Only 6% — McKinsey calls them “high performers” — report EBIT impact of 5% or more. Enterprise generative AI spending hit $13.8 billion in 2024, a sixfold increase over 2023, yet meaningful bottom-line impact remains rare for the majority. McKinsey sizes the long-term economic potential of generative AI at $2.6–$4.4 trillion annually across 63 use cases — a directional figure, since it aggregates modeled scenarios rather than measured deployments.

The issue isn’t the technology’s ceiling. It’s deployment targeting. Organizations are buying tools without a framework for deciding which problems those tools should actually address.

2026 is the year pilots are cheap and failures are expensive, because companies have moved from experimentation to scaling. The organizations pulling ahead are not running the most AI experiments. They’re the ones with a decision framework for where gen AI’s asymmetric strengths actually map to business value — and they’re ruthless about not crossing that line prematurely.

Four categories have produced consistent, measurable results in 2025 production deployments. What they share is structural: the task requires relatively little institutional knowledge to complete acceptably, and the volume is high enough that even a moderate efficiency gain aggregates to something real. That combination — low context, high volume — is the closest thing to a reliable gen AI ROI signal that exists right now.

The METR study is the clearest evidence for the mechanism. Experienced developers in the RCT accepted fewer than 44% of AI-generated suggestions. They spent 9% of task time reviewing and cleaning AI outputs, and 4% waiting for generations. That 13% overhead — before any benefit is counted — cancelled the marginal gains from AI-assisted code completion on tasks where the developer already knew the answer. The AI added a step between knowing and doing.

The practical framework: generative AI use cases can be mapped along two axes — institutional context required (low to high) and volume/repetition (low to high). The high-value quadrant is high-volume, low-context. The stall zone is high-context, where the practitioner’s expertise is itself the irreplaceable input.

What’s most important — and most consistently ignored — is that the failure mode in the stall zone is invisible. The cost doesn’t show up in the output. It shows up in the wall clock. Deploying gen AI into the stall zone doesn’t just fail to help. It can actively reverse productivity while convincing the practitioner they’re moving faster. That’s the specific danger the METR study documented.

Pharmaceutical drug discovery sits in a category of its own — medium-to-high context dependency, but with a specific structural advantage: the tasks AI handles (molecular design, virtual compound screening, lead optimization) are computationally bounded in a way that natural language tasks are not.

A November 2025 ScienceDirect review of five leading AI drug discovery platforms documented that Insilico Medicine’s generative AI-designed drug for idiopathic pulmonary fibrosis (ISM001-055) progressed from target discovery to Phase I in 18 months — compared to the typical five or more years for conventional approaches. Exscientia reports design cycles approximately 70% faster and requiring ten times fewer synthesized compounds than industry norms. These figures come from the company’s own publications and haven’t been independently audited across competitors — treat them as directional. Phase IIa results for ISM001-055, published in Nature Medicine in June 2025, showed dose-dependent improvement in forced vital capacity for IPF patients. That’s a genuine clinical signal, not just a preclinical milestone.

The failure case must sit right next to this. Drug Target Review’s 2025 annual review documented that as of December 2025, no AI-discovered drug had received FDA approval, with most programs still in early-stage trials. The FDA has reviewed more than 500 AI-related submissions since 2016. The agency’s own early experiments with generative AI tools reportedly experienced hallucination issues.

The honest frame for drug discovery in 2026: AI has demonstrably accelerated preclinical timelines and reduced compound synthesis costs. It has not yet demonstrated whether it improves clinical success rates — which is the variable that actually determines pharmaceutical value, since roughly 90% of drug candidates fail in clinical trials regardless of how they were discovered. Faster to Phase I means nothing if the Phase III failure rate stays the same.

The METR perception gap — 43 percentage points between measured slowdown and felt speedup — is not primarily a technology problem. It’s a measurement problem. Organizations that evaluate AI tools on developer sentiment, user adoption rates, or volume of code generated are going to systematically overestimate impact. That’s not a guess. The METR data makes it visible.

Every content strategist who has reviewed AI-assisted copy knows a version of this: the draft arrives faster, and you still spend the same amount of time cutting the hedging, restoring the voice, and fixing the factual slide that happened somewhere in the generation. The generation didn’t cost you time. The false confidence it created did.

Stack Overflow’s December 2025 developer survey registered the first-ever year-over-year decline in AI tool sentiment: 60% positive, down from above 70% in prior years, with only 3% of 49,000 surveyed developers reporting high trust in AI output. Among experienced developers specifically, the trust figure dropped to 2.6%. The gap between adoption and trust is a leading indicator — developers are using tools they don’t fully trust, in contexts where that distrust is partially well-founded.

McKinsey’s high performers — the 6% reporting 5%+ EBIT impact — share three practices the majority don’t. They redesign workflows rather than adding AI on top of existing ones. They scale faster when pilots show results. And they’re more than three times as likely to have senior leaders actively driving adoption and role-modeling usage. The technology is not the differentiator. The organizational operating model is.

The framework is useful. But what do you actually do on Monday morning? Here’s role-specific guidance drawn from the evidence, without the usual hedging.

Three independent data streams — METR’s developer productivity RCT, McKinsey’s high-performer organizational data, and the drug discovery Phase IIa clinical signal — point toward the same dynamic: generative AI’s value in 2027 will be determined by how well organizations have mapped their specific institutional-knowledge distribution against available tool capabilities.

Companies with dense institutional knowledge embedded in mature codebases, established clinical workflows, or proprietary research datasets face the most complex optimization problem. The companies with high-volume, low-context tasks — content, customer service, templated code — have already found the floor of the opportunity and should be scaling.

McKinsey’s data shows AI high performers are “at least three times more likely than their peers to report that they are scaling their use of agents” — a shift from single-model tool use to multi-step, agentic workflows that handle longer task chains without human intervention. The next frontier isn’t better individual code completions. It’s AI systems that carry enough context — of the codebase, the patient record, the molecule library — to work on high-context tasks that are currently out of reach.

Until that capability is independently validated in production, the safest deployment strategy in 2026 is the oldest one in technology: start where you have high volume and low institutional stakes, prove the ROI explicitly, then move up the context ladder one step at a time.

The organizations that look prescient in 2027 won’t be the ones that deployed gen AI everywhere in 2026. They’ll be the ones who deployed it precisely, measured it honestly, and treated the METR finding not as a reason to stop — but as a map of where not to start.