- Everyday AI
- Posts
- Ep 671: From Automation to Agents: Why Weak Data Makes AI Guess (1)
Ep 671: From Automation to Agents: Why Weak Data Makes AI Guess (1)
Google Releases Deep Research Agent, NotebookLM Ultra Release, TIME teams up with OpenAI and more
Everyday AI
December 12, 2025
Outsmart The Future
Today in Everyday AI
8 minute read
š Daily Podcast Episode: Many companies are investing in AI without investing in their people. We examine the growing AI training crisis and how leaders can close the gap. Find out more in todayās show and give it a watch/listen.
šµļøāāļø Fresh Finds: Target and Walmart upset with OpenAI, Disney AI Controversy, NotebookLM Ultra Released and more Read on for Fresh Finds.
š Byte Sized Daily AI News: Google Released a reimagined Deep Research Agent, Trump signs a new AI executive order, OpenAI reaches year 10 and more Read on for Byte Sized News.
š§ Learn & Leveraging AI: Despite massive spending on AI, most organizations struggle to scale impact. We explore the structural and cultural issues preventing teams from turning AI investment into real results.Keep reading for that!
ā©ļø Donāt miss out: OpenAI releases GPT-5.2, Disney partners with ChatGPT but presses Google, TIMEās Person of the year is AI and more. Check it here!
Ep 672: The AI Training crisis: Why companies are spending money on AI but not educating
Algorithms and automations have been buds for a decade plus. š¤
But the old 'smart' automations were rigid. If one thing was wrong, the automation would bust.
But with LLM-powered agents? Those automations are different. If something's wrong, the agent might just..... guess. š³
Weak data = weaker outcomes.
Here's how to fix it when agents come first and they're gonna finish the job, whether the data is strong or not.
Also on the pod today:
ā¢ AI training crisis exposed šØ
ā¢ Documentation: your secret AI sauce š
ā¢ Grounding models with company data š¢
Itāll be worth your 29 minutes:
Listen on our site:
Click to listen
Subscribe and listen on your favorite podcast platform
Listen on:
Hereās our favorite AI finds from across the web:
New AI Tool Spotlight ā Korgi Manages any project instantly, Kaily is the AI agent that goes beyond the chatbox, Mozart AI Creates custom high-energy trap and genre-blended tracks from your vocals.
AI Holiday Shopping ā AI chatbots like ChatGPT, Gemini and Perplexity are reshaping holiday shopping and stealing attention from Google and Amazon. Retailers are scrambling to adapt, but is the experience actually better for shoppers?
Disney AI Controversy ā Disneyās $1bn AI deal sparks fresh battle over creative workersā rights
NotebookLM Ultra Update ā Google quietly turns NotebookLM into a serious perk for AI Ultra power users
AI Helping Cybercrime Strike ā AI copilot helps Maharashtra police battle Indiaās exploding cybercrime wave
1. Googleās new Gemini research agent sets up an AI showdown with OpenAI š§
Google has rolled out a āreimaginedā Gemini Deep Research agent powered by its latest Gemini 3 Pro model, positioning it as a tool that can digest huge amounts of information and feed those capabilities directly into third-party apps through a new Interactions API.
The company plans to wire this agent into Google Search, Finance, the Gemini app, and NotebookLM, signaling a future where AI agents increasingly handle the looking-things-up part of daily life instead of humans typing queries. Google is also touting Gemini 3 Pro as its āmost factualā model yet and is backing that up with a new DeepSearchQA benchmark and strong results on independent tests like Humanityās Last Exam, even as OpenAIās ChatGPT 5 Pro and the new GPT 5.2 āGarlicā model keep the race tight.
2. Trump moves to block state AI rules as floods, health costs test Americansš§āāļø
In a sweeping and very current move on tech policy, President Trump signed an executive order that aims to stop individual states from heavily regulating artificial intelligence firms, shifting power over AI rules firmly toward Washington and signaling a preference for light federal oversight instead of a state-by-state patchwork.
The order creates a new "AI Litigation Task Force" to challenge state laws that conflict with the administration's vision, and it could tie some federal funding to states backing off rules that force AI systems to change their outputs. The announcement lands as AI surges across the economy, from Disney cutting a deal with OpenAI for its video generator to Wall Street nervously eyeing an AI-fueled market bubble and senators pressing tech leaders to be more transparent about risks.
3. OpenAI at 10: From long-shot bet to near-term āsuperintelligenceā š¤
Marking its 10-year anniversary, OpenAI is signaling that what once sounded like science fiction is now on a visible roadmap, with its leaders saying they expect to build superintelligence within the next decade. In a candid reflection, the company describes its journey from a small, uncertain research group in 2015 to the maker of ChatGPT and GPTā4, tools that it says now outperform most top humans in complex intellectual tasks.
OpenAI credits its rise to a mix of early risky bets on deep learning, large-scale computing, and a controversial but now widely copied strategy of releasing powerful AI systems step by step so the public and regulators can adapt.
4. Gunderson Dettmer Goes All-In On AI With Perplexity Enterprise š»
One of the worldās top venture law firms is quietly turning into an AI power user, with Gunderson Dettmer rolling out Perplexity Enterprise firmwide and seeing roughly 80% of its lawyers actively using it only months after launch.
The tool now fields about 35,000 monthly queries inside the firm, effectively becoming a search engine replacement that helps attorneys research faster, stay current on law and tech, and sharpen client advice with better context. This move builds on Gundersonās earlier in-house system, ChatGD, but Perplexityās web-connected research and cleaner interface appear to have unlocked far broader and deeper daily use across legal and business teams.
5. TIME teams up with OpenAI on AI literacy push š„
In a new move that underscores how fast artificial intelligence is changing everyday life, TIME and OpenAI have announced a new partnership focused on teaching people how to use AI effectively and responsibly.
OpenAI is rolling out an AI literacy platform, including its OpenAI Academy and OpenAI Certifications, aimed at helping people understand what AI is good at, how it works, and where it should be used with care. TIME is bringing its editorial muscle, live events, and its TIME100 AI coverage to give this technical training broader context so people are not just using AI tools, but also understanding their social impact.
Literally.
Execs are signing seven-figure checks for enterprise AI licenses, assuming the technology works like an "easy button" for productivity. But hereās the brutal reality nobody discusses in the boardroom.
Most studies show that more than 90% of leaders call AI one of their companyās top priorities. Yet less than a third are actually training their people how to use it.
Wanna know why?
Sames.
So we sat down with Dan Lawyer, Chief Product Officer at Lucid Software, to break down why this gap exists.
The problem isn't the software, he said.
It's that you're treating AI like a tool instead of a new hire that needs an employee handbook.
1. The Last Mile Disconnect š
Everyone thinks grounding AI means connecting it to a data lake.
That is only half the battle.
Dan explained that data only tells the model what happened. It doesn't tell the model how the work gets done. This is the "Last Mile" problem. You can feed a model terabytes of financial records, but if you don't document the specific ten-step reconciliation workflow your accounting lead keeps in her head, the AI will fail every single time.
You cannot automate what you haven't documented.
The companies finding real ROI right now are the ones aggressively documenting their "secret sauce" workflows. They treat process documentation as the bridge between raw data and usable intelligence. If you skip this step, you are just building a faster way to generate confusion.
Try This:
Pick your teamās most expensive recurring workflow and write the ānew hireā version: step-by-step decisions, edge cases, and handoffs (not just the data inputs).
McKinseyās 2025 State of AI survey found 88% of orgs report regular AI use in at least one function, but only 32% say theyāre āscalingā AI and just 7% say itās āfully scaledā ā the gap is usually workflow integration, not model access
2. The Sycophant Trap š¤
We tend to treat AI like a search engine.
Dan argues you need to treat it like a "Sixth Man" on a basketball team. It is a capable substitute player that keeps momentum going, but it comes with a dangerous personality flaw.
It wants to flatter you.
AI models are designed to be helpful, which often manifests as sycophancy. They will validate your bad ideas because they are programmed to agree with the user. In a business context, this creates a confirmation bias loop that amplifies poor decision-making speed.
You have to train the model to be a critic.
Your team needs to understand that getting value from these tools requires pushing back. You have to explicitly instruct the AI to stop being nice and start being critical. If you don't, you aren't getting intelligence. You are getting an echo chamber.
Try This:
Run a "criticism audit" on your next three strategic outputs.
If the model agrees with your premise every time, assume youāre getting flattery, not analysis. A 2025 study across 11 top models found they affirm usersā actions 50% more than humans do, and people trusted the sycophantic responses more anyway. Yikes.
3. Speed Is The New Safety ā”
The technology changes every single week.
Literally.
We have seen major model updates from OpenAI, Google, and Anthropic in back-to-back weeks. If you rely on a traditional corporate procurement cycle that takes six months to approve a vendor, you are effectively choosing obsolescence.
By the time legal approves the tool, the industry has moved on.
Dan suggests implementing a "Fast Path" policy immediately. This isn't about letting employees upload sensitive IP to random websites. It is about creating a pre-approved sandbox where teams can test new models on non-sensitive data the day they launch.
Shadow AI is the alternative.
If you block access, your high performers will just use their personal devices to do the work anyway. That is infinitely more dangerous than a managed experiment. You have to sanction the playground or you lose control of the game entirely.
Try This:
Draft a one-page "Fast Path" agreement today that authorizes specific employees to test new AI tools within 24 hours of release. Define the boundaries clearly by listing exactly which data types are bannedālike customer PII or unreleased IPāand then get out of their way.
The goal is to separate your production environment from your R&D environment. If you wait for permission to play with the new toys, you will be solving last year's problems while your competitors build with tomorrow's capabilities.
Reply