The Enterprise AI Gap Is Three Gaps: Documents, Deployment, Direction

First problem: documents still break the agents

Most white-collar work in healthcare, legal, finance, and insurance lives in PDFs and spreadsheets. That is Reducto CEO Adit Abraham's starting point on the October 21 a16z episode: document processing remains the critical bottleneck for AI agents in professional sectors, even as frontier models keep improving. In regulated industries, a misread document is not a cosmetic glitch, Abraham argued — in his framing, a wrong checkbox is not an oops but a patient's outcome. Agents need deterministic, high-recall extraction before they can be trusted with real decisions. m:…2374

The stakes are measurable. Abraham cited one striking figure: roughly 80% of Medicare prior-authorization appeals result in the original decision being overturned, which implies a large share of denials rest on readings of documents that do not survive scrutiny. His proposed answer is specialized infrastructure rather than bigger general models: a hybrid pipeline that uses traditional computer vision to preserve document structure, with visual language models as a correction loop. Reducto, founded in 2023, has unified this into its R1 model for end-to-end processing. Abraham was explicit that Reducto does not try to compete with the labs on reasoning quality, describing its work as tangential to general intelligence. m:…2374

Second problem: almost nobody has an agent in production

Tenex's Alex Lieberman and CJ Hess, who work with Fortune 2000 companies moving toward AI-native software development, put a number on the implementation gap: only 5% of enterprises with more than 1,000 employees have an AI agent in production for non-engineering work. The bottleneck, they say, is not model capability but the final mile of diffusion — and above all, data permissioning. Enterprise CIOs and CISOs are stuck on role-based access control when deploying agents, and that governance layer is the actual blocker. m:…3291

Tenex's formulation of the problem is that automating a process that does not exist cannot work. Companies without standardized processes get nothing from automation layered on top of them. Tenex's own operating model illustrates the discipline: engineers paid on story points to reward output rather than time, AI strategists incentivized on customer retention and growth, and a so-called meta harness of structured documentation plus a command-line interface to standardize AI-driven engineering workflows. One data point they offered: a mid-tier SaaS replacement cost $1,000 to build, replacing a legacy system that cost $40,000 per year. m:…3291

Tenex is also bearish on vertical-specific AI solutions over the long term, unless the company holds a strategic moat such as compliance advantages or proprietary data. Their simple answer for what constitutes that moat is access to data that products like ChatGPT cannot replicate. Their forecast is a market in which a small set of companies owns the user interface and other agents plug into those central platforms. m:…3291

Third problem: strategy, not tool access, creates impact

BCG managing director Vinciane Beauchene, speaking on the October 20 So What podcast, closed the loop with the strategy layer. Employees with strong strategic guidance on where to deploy AI are five times more likely to achieve measurable impact than employees with only strong tool access. Individual productivity gains, she argued, do not necessarily translate into collective organizational value creation. m:…6030

Her prescription is focus: avoid spreading AI thinly across the organization — what she called peanut-butter spreading — and instead concentrate on a few areas where it can make a measurable difference, led by cross-functional transformation teams that connect technology, business, strategy, HR, finance, and risk. And put human input at the front end of AI-driven processes rather than relying on human revision at the back end — in BCG's experiments, front-loading human input on complex elements inaccessible to AI produced the best forecasting results. m:…6030

The leadership warning, as Beauchene put it, is that the worst outcome for a leader is moving too fast and hard-coding abiding operating decisions while the technology keeps evolving. With three in four employees already using AI regularly, adoption is no longer the scarce resource — judgment is. BCG also found that labeling output as coming from an AI worker rather than an AI tool reduced personal accountability and employees' ability to spot mistakes. m:…6030

The through-line for builders

Three problems, three owners. Infrastructure teams own document determinism: until extraction is reliable, agents cannot touch regulated workflows. Engineering leadership owns process standardization and permissioning: without a defined process and RBAC-safe deployment, agents never reach production — and Tenex's 5% figure shows how rarely they do. Executives own the strategic focus: tools without direction produce productivity theater, not measurable impact. m:…2374 m:…3291 m:…6030

For developers building enterprise AI, the checklist writes itself: standardize the process before automating it, gate agent data access through existing permissioning systems, treat document pipelines as the actual product, and measure collective value rather than individual productivity. The model is increasingly the commodity. The gap is everything around it. m:…2374 m:…3291 m:…6030