The dominant narrative about AI and work is a story about jobs. A role gets automated, or it does not. A profession survives, or it disappears. That framing is intuitive, and it is the wrong unit of analysis. Google's newly released AI and Economy ATLAS report, the first large-scale study of how people actually use Gemini across 150 countries and 800 occupations, makes this concrete with data instead of speculation. It found that AI use at work is broad but shallow. Adoption reaches 68 percent of occupations and 90 percent of total United States employment, yet within a typical job, AI is used for only about 21 percent of tasks. Fewer than 10 percent of work interactions fully automate anything. The rest is collaboration: ideation, information retrieval, troubleshooting, drafting.
That distinction, task versus job, is not an academic nuance. It is the actual mechanism through which AI creates economic value, and it is the same mechanism Forte Group has been building on the ground for two years. A recent engagement with an infusion therapy provider, built around a platform we call the Virtual Infusion Guide, is a working example of exactly what ATLAS describes at population scale, compressed into a single clinical operation.
ATLAS is built from 15 million aggregated, de-identified interactions across the Gemini App, AI Mode, and the Gemini API, spanning more than 150 countries, 140 languages, 800 occupations, and 4,000 tasks. Dame Diane Coyle of Cambridge and Dr. David Autor of MIT contributed to the analysis, which matters because Autor has spent two decades documenting how technology reshapes labor markets task by task rather than job by job.
Two findings stand out. First, non-routine cognitive tasks, the category ATLAS uses for creative design and hypothesis testing, show up in AI work interactions at nearly double the rate they occur in the economy overall, 65 percent versus 35 percent. People are reaching for AI on the hardest, least standardized parts of their work, not the easiest. Second, adoption is not confined to knowledge work. Auto technicians and industrial mechanics use conversational AI as a live diagnostic collaborator, and they are twice as likely as the average user to work in images and video because a wiring diagram or a worn part is easier to explain by showing it. The pattern is consistent across a hospital billing office and a repair bay: AI shows up where a task is complex enough to be worth assisting, and it shows up as assistance, not replacement.
Local Infusion, a specialized infusion therapy provider, ran into a version of this same bottleneck. Every infusion prescription requires validation of drug interactions, dosing calculations, and insurer-specific requirements before a patient can be treated. The clinical staff who manage that process, called Infusion Guides, were spending most of their day on manual data entry, prescription interpretation, and insurance verification rather than patient care. Prescription validation averaged about 20 minutes per patient and typically required a manual back-and-forth with the referring provider. As referral volume grew, that bottleneck limited how many patients each Guide could support, and it capped how fast the company could open new centers.
Forte built the Virtual Infusion Guide platform to remove that specific bottleneck, not to replace the Guides. OCR pipelines extract structured data from prescription faxes, scanned documents, insurance forms, and lab results. Azure OpenAI's GPT-4 interprets that extracted content, checks Rx completeness against clinical and insurance requirements, and recommends follow-ups. What used to take a Guide 20 minutes now takes five. A retrieval-augmented conversational system, built on OpenSearch and bootstrapped from the actual conversations of the company's most experienced Guides, lets any Guide query patient records, treatment protocols, and insurer-specific rules in natural language from their first day on the job. Low-confidence outputs route to senior Guides for review, and those corrections feed back into the model.
This is the ATLAS finding in miniature. The platform did not attempt to automate the Infusion Guide role. It automated the specific, structured, repetitive tasks inside that role, prescription interpretation, form extraction, insurance rule lookup, while routing judgment calls and patient relationships back to a person. Five engineers built and now maintain the entire platform, which serves as the operational backbone across 47 clinical centers, scaling to 85 by year end.
ATLAS makes a related point that deserves more attention than it gets: over 86 percent of AI interactions in the dataset happen outside of work altogether, much of it on high-friction administrative tasks such as navigating government services, taxes, and licensing. Standard economic metrics are built to measure output and employment. They are not built to measure the hours a household or a clinical staff member no longer spends on friction. That value is real, and it does not show up cleanly in a productivity statistic.
The same undercounting happens inside a single company if leaders only measure headcount. The metric that actually captures what happened at Local Infusion is the ratio of referral growth absorbed to engineering headcount required to absorb it. Existing centers now take on rising referral volume without adding staff proportionally. The data foundation built alongside the clinical product, a Snowflake warehouse with structured pipelines for patient data and lab analysis, has become a second asset in its own right, opening monetization through data-driven partnerships that did not exist before the platform did. And the same automated eligibility and validation capability is now expanding into pharmacy services, a transfer of capability into an adjacent vertical for a fraction of the cost of the original build. None of that shows up if the only question asked is whether AI replaced anyone's job. It did not. It changed what the operating leverage of a five-person engineering team could produce.
ATLAS v1.0 measures usage of Google's own products, so it reflects Gemini and AI Mode behavior specifically, not a comprehensive census of every AI tool in the market, and Google is explicit that this is an early, evolving methodology rather than a settled picture. The Local Infusion engagement is a single client relationship, and the specific ratio of engineering cost to referral growth will not transfer directly to every care delivery operator. Human-in-the-loop review is not a temporary scaffold to be removed once the model improves. Clinical accountability requires it permanently, and any operator building a similar system should plan for that review capacity as a fixed cost, not a phase-out item.
The AI economy is not being assembled job by job. It is being assembled task by task, in accounting departments, repair bays, and infusion centers, by teams willing to isolate the bottleneck before they automate it. The operators who understand that granularity now are the ones who will capture the leverage later.
Let's talk about where you are today and where you want to go - our experts are ready to help you move forward.
