AI tools get bought fast, but the training and workflow redesign to actually use them well lag far behind — which one's the real bottleneck?

This explores the non-hardware, non-license side of AI adoption: the training, process redesign and organizational rewiring that have to happen before AI shows up in the numbers. The corpus doesn't rank these investments with a clean 'number one.' Read together, though, it points to something specific: the scarcest investment is organizational knowledge of where AI fits, more than technical skill.

The frame comes from Why does AI productivity appear weak when investment is high?. Brynjolfsson, Rock and Syverson argue that general-purpose technologies like GPTs need large complementary investments in worker training, new business processes and supporting software. Company accounts don't record these as assets, so early on productivity looks weak even as money pours in. Later, once those investments start paying off, productivity looks unusually strong. That J-curve shape explains a puzzle in the survey data. In Are CEOs actually seeing business gains from AI investments?, 56% of CEOs report no cost or revenue benefit yet. Meanwhile Are companies actually generating real value from AI investments? finds about half of firms capturing value, with leaders earning 2.3–2.8× the returns of laggards. Both can be true if some firms have climbed out of the bottom of the J and others haven't.

What separates the two groups? The less obvious answer is recognition and governance, not tooling. Does easier tool-building actually solve enterprise adoption problems? argues that making AI tools easier to build doesn't fix the real bottleneck. Most workers don't see their own tasks as automatable, and real adoption depends on decisions that cross departments and budget cycles. BCG's leaders fit this picture: AI decision-making authority in those firms has moved out of IT and into the business. The usage data shows the same unevenness inside companies. Who adopts enterprise AI first and how do they use it? finds heavy use among marketing staff and early-career workers but much lighter use among executives. So the people who decide how work gets reorganized are often the ones using the tool least. Does AI adoption follow wealth and mature over time? adds a time dimension: as adoption matures, usage shifts from handing off whole tasks toward collaboration and learning. That suggests a key investment is helping people learn to work alongside AI, beyond automating what they already do.

Two less glamorous investments come up from other directions. The first is maintenance. Are enterprises abandoning custom AI apps for third-party solutions? reports that CIOs are giving up on in-house AI apps because keeping up with constant model changes costs too much. Building a tool turns out to be the cheap part; maintaining it is where the cost lies. The second is verification. When do users stop checking whether AI output is actually backed? describes users accepting fluent AI output without checking it, because checking is costly. An organization that doesn't build review habits and accountability into its workflows may get output without reliability. Relatedly, Why do agent benchmarks not predict real economic value? suggests firms need their own ways to test AI against their actual long, multi-step work, because public benchmarks measure contests rather than jobs.

The surprising takeaway: the investment the corpus emphasizes most isn't skills training in prompting. It's the managerial work of mapping which tasks can change, who is allowed to change them, and how results get checked. That work is the hardest to buy, the hardest to measure, and the main reason the J-curve dips before it rises. A caveat: most of this evidence comes from surveys and one theoretical model. The corpus has no head-to-head study measuring which complementary investment pays off most.