More than 70% of companies accelerated AI licence distribution in 2025 and 2026. Employees gained assistants in code editors, email and spreadsheets. But the financial dashboards of those companies tell another story: most recorded stagnant margins and a net shift of work.
Structural mismatch, not failed adoption: companies confused 'typing speed' with 'system efficiency'. The Upwork Research Institute 2024 survey found 77% of employees say AI added to their workload, and 47% do not know how to reach the productivity targets their employers expect. The METR RCT found experienced developers took 19% longer with AI than without.
Leaders measure lines of code, pull requests and drafts. Those are vanity metrics. Generated code does not pay bills; reliable production systems and closed sales do.
LookADev builds system-layer pipelines, not desktop assistants. The work is measured on business outcomes, not on the volume of AI output. The Cainelli scope is an NF-e XML pipeline with Tiny/Olist integration; it is a documented technical scope, not a measured saving claim.
First, the randomised clinical trial (RCT) by METR (Model Evaluation & Threat Research): experienced developers using AI took 19% longer to complete complex tasks in real repositories than the control group without AI. The most alarming finding was the 'Perception Gap': before the test, developers predicted they would be 24% faster; after becoming slower, they still believed they had been 20% faster. The feeling of ease masked the time lost to debugging.
Second, GitClear's 'The Maintainability Gap' report, which analysed more than 200 million lines of code: refactoring plunged 70% while code block duplication surged 81%. Instead of modularising systems, AI tools induced the proliferation of 'workslop' code: copied, disorganised blocks without regression tests.
Third, the Upwork Research Institute survey: 77% of employees said that using AI increased their total workload and raised stress levels, because delivery expectations rose while useful working hours were consumed reviewing inaccurate drafts generated by assistants.
2. The Fallacy of Vanity Metrics
Corporate leadership often measures AI success through metrics that are easy to collect: lines of code generated, Pull Requests opened or emails drafted. These are vanity metrics. Generated lines of code do not pay bills; bug-free systems in production and closed sales do.
Giving individual tools to every operator merely moves the bottleneck: the employee types faster, but the manager spends twice as long reviewing errors and customers still wait hours for real answers.
3. System-Layer Automation vs. Desktop Assistants
LookADev approaches productivity from the opposite perspective. We do not sell assistants to sit at an employee's desk and suggest text. We redesign processes and build closed pipelines at the system layer.
The Sebrae Triagem IA and Cainelli Bebidas projects address different flows: CV screening and processing NF-e XML. At Cainelli, the documented scope includes a parser, product mapping and integration with Tiny/Olist ERP, with review of uncertain matches. This scope does not prove the complete elimination of manual work or measured hours saved.
4. References & Scientific Sources
1. METR (Model Evaluation & Threat Research): 'Measuring the Impact of AI on Experienced Software Developers: A Randomized Controlled Trial' (2025). Controlled trial showing a 19% increase in completion time and a 39-percentage-point perception gap. 2. GitClear Research: 'Coding on Copilot: 2024–2025 Data Shows Downward Pressure on Code Quality and The Maintainability Gap'. Large-scale analysis of 214 million lines of code demonstrating the collapse of refactoring and the escalation of duplicated code ('workslop'). 3. Upwork Research Institute: 'AI and the Productivity Paradox: How Generative AI is Changing Work' (2024–2025). Survey of leaders and employees documenting the review burden imposed by decentralised AI tools. 4. Harvard Business School & MIT Sloan: 'Navigating the Jagged Technological Frontier: Field Experimental Evidence on the Effects of AI on Knowledge Worker Productivity' (Dell'Acqua et al.).