noticias radar ia agentes ia gobernanza ia adopcion ia

Yesterday's radar celebrated that agents are already being sold bundled within CRMs; today it's time for the flip side, which is just as important: most of those agents don't survive the jump from proof of concept to production. NVIDIA says it bluntly, Cognizant is setting up an entire unit in Europe to fix it, and in the background, the four giants are spending a record figure on the infrastructure that sustains it. The core takeaway: in 2026, the advantage isn't in which model you use, but in whether you know how to operate it.

NVIDIA: The problem is no longer building the agent, it's keeping it alive

Marcio Aguiar, director of NVIDIA's Enterprise division for Latin America, summarizes the phase shift: building an agent that executes specific tasks has become relatively simple, but keeping it running every day is the real obstacle. He points out three recurring mistakes: underestimating computing resources when scaling from testing to massive use, postponing security and governance instead of designing them from the start, and scaling without clarity on the return. For your company: an agent that runs 24/7 in the background is not an experiment, it's a fixed cost; before launching it, budget for its continuous operation —compute, access to internal data, supervision— and define how you will measure ROI, because the competitive advantage of the next two years will be in operating reliably under your rules, not in having access to the latest model. Source

Only 4 out of 33 agent pilots reach production

Cognizant opened an AI unit dedicated to EMEA on July 28th with a very specific goal: to take agents out of the pilot cemetery and into production. It cites data from IDC that puts numbers to the problem: 88% of agent proofs of concept never reach broad operation —for every 33 pilots a company starts, only 4 go into real operation—. For your company: the fact that the big consulting firms are seting up an entire division around this gap is the signal that it is real and expensive; the lesson is not to launch more tests, but to industrialize one or two —choose a process with clean data and a clear owner, and take it all the way to the end before opening the second front. Source

The giants' AI bill: 725 billion that you will end up noticing in your rates

Amazon, Google, Meta, and Microsoft will allocate about 725 billion dollars this year to AI infrastructure —data centers and chips—, 77% more than in 2025, and to finance it, the group has raised about 159 billion dollars in debt; Microsoft and Meta report results with AI under pressure to justify such spending. For your company: without giving stock market advice, the operational reading is twofold —that infrastructure bill pushes prices up, but fierce competition between providers and cheaper "flash" models push them down—; don't tie long contracts to a single provider and review rates every quarter, because the token you pay dearly for today could cost half as much in a few months. Source

What to watch tomorrow?

On August 2nd —this Sunday— the transparency obligations of the EU AI Act come into force: if you have a chatbot or generate customer-facing content with AI, this weekend is the real deadline to put up the notice and label what is generated. And with Microsoft and Meta's results on the table, the operational question that brings you back to the beginning: Is your next agent designed to survive a pilot, or to work every day?