WEBVTT 00:00:00.000 --> 00:00:02.600 Welcome back to AI Unraveled, your strategic 00:00:02.600 --> 00:00:05.059 briefing on the business impact of artificial 00:00:05.059 --> 00:00:08.220 intelligence. This is your weekly rundown for 00:00:08.220 --> 00:00:12.359 December 8th to December 14th, 2025. This week, 00:00:12.480 --> 00:00:14.919 the industry didn't just move, it picked sides. 00:00:15.220 --> 00:00:18.039 In a stunning display of corporate maneuvering, 00:00:18.079 --> 00:00:21.199 Disney has inked a massive $1 billion partnership 00:00:21.199 --> 00:00:24.859 with OpenAI, just days after hitting Google with 00:00:24.859 --> 00:00:27.219 a cease and desist order for copyright infringement. 00:00:27.260 --> 00:00:30.210 The message is clear. If you want the Magic Kingdom's 00:00:30.210 --> 00:00:32.829 data, you have to pay the toll. Meanwhile, OpenAI 00:00:32.829 --> 00:00:35.429 continues its offensive with the release of GPT 00:00:35.429 --> 00:00:38.670 -5 .2 and the hiring of the Slack CEO to drive 00:00:38.670 --> 00:00:41.289 revenue, signaling they are ready to monetize 00:00:41.289 --> 00:00:43.890 the enterprise at scale. And speaking of money, 00:00:44.509 --> 00:00:46.890 Meta is officially pivoting away from open source, 00:00:47.009 --> 00:00:49.869 focusing entirely on profit -generating AI models. 00:00:50.149 --> 00:00:53.109 The free lunch might be over, but the race isn't 00:00:53.109 --> 00:00:55.759 just happening on Earth. We have reports of Jeff 00:00:55.759 --> 00:00:58.439 Bezos and Elon Musk competing to launch orbital 00:00:58.439 --> 00:01:00.579 data centers to escape our energy constraints. 00:01:00.920 --> 00:01:03.439 But before we unravel the rest of the week's 00:01:03.439 --> 00:01:06.840 news, stop marketing to the masses. Start briefing 00:01:06.840 --> 00:01:10.280 the C -suite. You have seen the power of AI unraveled, 00:01:10.280 --> 00:01:13.060 zero noise, high signal intelligence for the 00:01:13.060 --> 00:01:15.959 world's most critical AI builders. We create 00:01:15.959 --> 00:01:18.500 tailored proprietary podcasts designed exclusively 00:01:18.500 --> 00:01:20.980 to brief your executives and your most valuable 00:01:20.980 --> 00:01:46.180 clients. Welcome to the Deep Dive. You know, 00:01:46.180 --> 00:01:48.719 for years, the business world has chased this 00:01:48.719 --> 00:01:51.500 promise of simple automation. That old... If 00:01:51.500 --> 00:01:55.420 X, then Y model. Very rigid. Exactly. The old 00:01:55.420 --> 00:01:59.280 RPA era. But that era is over. We are now in 00:01:59.280 --> 00:02:01.900 a fundamentally different probabilistic age of 00:02:01.900 --> 00:02:04.859 agentic AI. And this is not just a tech upgrade. 00:02:04.980 --> 00:02:07.819 It's a total structural transition. The economic 00:02:07.819 --> 00:02:11.039 rules are being rewritten as we speak. Our mission 00:02:11.039 --> 00:02:13.819 today for you is to strip away the hype around 00:02:13.819 --> 00:02:18.229 digital workers. really apply some rigorous financial 00:02:18.229 --> 00:02:20.650 discipline because you have to validate these 00:02:20.650 --> 00:02:24.550 massive, sometimes volatile investments. So we're 00:02:24.550 --> 00:02:26.729 doing a deep dive into the three layer agentic 00:02:26.729 --> 00:02:29.729 workflow ROI model to bridge that gap between 00:02:29.729 --> 00:02:32.330 promise and actual measurable profit. And that 00:02:32.330 --> 00:02:34.610 financial rigor is the key conversation. But 00:02:34.610 --> 00:02:36.759 a lot of companies are. frankly, using the wrong 00:02:36.759 --> 00:02:39.379 calculus. How so? They treat AI adoption like 00:02:39.379 --> 00:02:41.259 installing a single piece of software. You just 00:02:41.259 --> 00:02:43.300 can't do that. You have to think of it like managing 00:02:43.300 --> 00:02:45.199 a diversified investment portfolio. Okay, an 00:02:45.199 --> 00:02:47.300 investment portfolio? I like that. Yeah, you're 00:02:47.300 --> 00:02:49.800 dealing with distinct asset classes. Each one 00:02:49.800 --> 00:02:52.400 has wildly different risks, different liquidity 00:02:52.400 --> 00:02:56.039 needs, and different timelines for single return. 00:02:56.319 --> 00:02:58.879 So we break it down into three tiers based on 00:02:58.879 --> 00:03:02.520 autonomy and, most importantly, verifiable ROI. 00:03:03.319 --> 00:03:05.539 And I guess that need for categorization really 00:03:05.539 --> 00:03:08.400 comes from the failures of traditional automation, 00:03:08.719 --> 00:03:11.699 right? We all remember how brittle RPA was. Oh, 00:03:11.699 --> 00:03:13.900 absolutely. You know, a vendor changes an invoice 00:03:13.900 --> 00:03:16.379 format just slightly, and the whole system just 00:03:16.379 --> 00:03:19.199 seizes up. Exactly. And then you need an expensive 00:03:19.199 --> 00:03:21.300 human intervention to fix the chain, because 00:03:21.300 --> 00:03:23.479 for the last 20 years, those systems just followed 00:03:23.479 --> 00:03:26.219 explicit rules. Right. The moment you bring in 00:03:26.219 --> 00:03:28.879 agentic AI, you're shifting to systems that can 00:03:28.879 --> 00:03:31.419 infer intent. They can reason, they can plan, 00:03:31.520 --> 00:03:33.900 they can execute these real - complex multi -step 00:03:33.900 --> 00:03:36.680 workflows that's agency but that agency comes 00:03:36.680 --> 00:03:39.580 with a catch a huge one it introduces what we 00:03:39.580 --> 00:03:42.840 call probabilistic risk an error in the old system 00:03:42.840 --> 00:03:45.439 was a bug you could find it you could fix it 00:03:45.439 --> 00:03:48.199 it was predictable right an error in an agentic 00:03:48.199 --> 00:03:50.879 system is a hallucination a reasoning failure 00:03:50.879 --> 00:03:53.979 something you can't just patch out you have to 00:03:53.979 --> 00:03:56.159 govern its behavior okay so let's start with 00:03:56.159 --> 00:03:58.060 the safest place to begin them the foundation 00:03:58.060 --> 00:04:01.469 layer one Simple automation. Your sovereign bond. 00:04:01.729 --> 00:04:03.849 The sovereign bond of your AI portfolio. Yeah. 00:04:03.949 --> 00:04:06.830 The goal here is simple. It's low risk. cost 00:04:06.830 --> 00:04:09.289 reduction, the human is really just the operator 00:04:09.289 --> 00:04:11.849 overseeing the output. And layer one is where 00:04:11.849 --> 00:04:14.750 we see old RPA really growing up into intelligent 00:04:14.750 --> 00:04:18.089 automation. It handles huge volumes of unstructured 00:04:18.089 --> 00:04:22.050 data emails, PDFs, claims, forms, and turns it 00:04:22.050 --> 00:04:24.310 into structured output. Like intelligent document 00:04:24.310 --> 00:04:26.810 processing IDP. Exactly. That's the most common 00:04:26.810 --> 00:04:29.170 use case. But critically, the workflow itself 00:04:29.170 --> 00:04:31.290 is still pretty rigid, which keeps the risk low 00:04:31.290 --> 00:04:33.709 and efficiency high. And the financial case here 00:04:33.709 --> 00:04:36.319 must be the easiest sell to the CFO. It is. You're 00:04:36.319 --> 00:04:38.699 just comparing a known quantity manual labor 00:04:38.699 --> 00:04:41.879 cost against a predictable compute cost. The 00:04:41.879 --> 00:04:44.180 studies are very consistent here. Intelligent 00:04:44.180 --> 00:04:46.600 document processing can cut your costs by 60 00:04:46.600 --> 00:04:50.560 to 80 percent. 60 to 80. Wow. So like a $15 invoice 00:04:50.560 --> 00:04:53.160 becomes what, three or four bucks? Pretty much. 00:04:53.220 --> 00:04:56.040 You're seeing average costs per invoice drop 00:04:56.040 --> 00:04:59.660 from that $15 to $40 range all the way down to 00:04:59.660 --> 00:05:02.639 $3 to $8. It's a massive drop. But it's not just 00:05:02.639 --> 00:05:05.089 the cost per transaction. Yeah. It's the speed 00:05:05.089 --> 00:05:07.769 that has its own benefits. Precisely. The cycle 00:05:07.769 --> 00:05:10.610 time reduction is enormous. You go from a manual 00:05:10.610 --> 00:05:13.110 process taking five to seven days down to maybe 00:05:13.110 --> 00:05:16.529 six to 12 hours. That's a 90 % reduction. A 90 00:05:16.529 --> 00:05:19.370 % reduction. And that speed has an immediate 00:05:19.370 --> 00:05:23.420 verifiable financial benefit. You can now capture 00:05:23.420 --> 00:05:25.639 early payment discounts from your suppliers. 00:05:25.879 --> 00:05:28.579 Which are usually, what, 1 % to 3 % of the invoice 00:05:28.579 --> 00:05:31.180 value? Exactly. And that single benefit, that 00:05:31.180 --> 00:05:34.160 1 % to 3%, can often pay for the entire system 00:05:34.160 --> 00:05:36.339 by itself. And something we often overlook here 00:05:36.339 --> 00:05:39.279 is accuracy. Traditional OCR, it always capped 00:05:39.279 --> 00:05:42.399 out around, what, 85%, 95 %? At best. And that 00:05:42.399 --> 00:05:45.199 last 5 % or 15 % required expensive humans to 00:05:45.199 --> 00:05:47.680 go in and validate everything, which, you know, 00:05:47.699 --> 00:05:49.540 eats away at your speed gains. So what's different 00:05:49.540 --> 00:05:53.170 now? Well, LLM driven IDP systems use semantic 00:05:53.170 --> 00:05:55.790 understanding. They're not just recognizing characters, 00:05:55.970 --> 00:05:58.649 they're reading and interpreting the data. This 00:05:58.649 --> 00:06:02.569 lets them get to over 99 % accuracy. Over 99%. 00:06:02.829 --> 00:06:04.810 So the people who used to do reconciliation can 00:06:04.810 --> 00:06:07.550 now be moved to higher value work. That's the 00:06:07.550 --> 00:06:09.589 idea. There's a great case study in health care, 00:06:09.670 --> 00:06:12.269 actually, for oncology research. Oh, yeah. Manual 00:06:12.269 --> 00:06:15.230 data extraction by five physicians took seven 00:06:15.230 --> 00:06:18.550 months. The LLM took 12 days. Seven months versus 00:06:18.550 --> 00:06:21.670 12 days. A 91 % reduction in physician hours. 00:06:21.910 --> 00:06:25.670 And here's the kicker. The LLM actually exceeded 00:06:25.670 --> 00:06:28.269 human performance in capturing specific survival 00:06:28.269 --> 00:06:30.810 events from unstructured notes. So it's not just 00:06:30.810 --> 00:06:32.629 cheaper and faster. It's actually better. In 00:06:32.629 --> 00:06:35.009 these data -heavy, well -defined tasks, yes, 00:06:35.129 --> 00:06:37.509 it's demonstrably superior and it offers the 00:06:37.509 --> 00:06:40.029 fastest time to value. So layer one is the cash 00:06:40.029 --> 00:06:42.569 cow. It's giving you immediate, reliable ROI 00:06:42.569 --> 00:06:45.569 in, say, six to 12 months. And that generates 00:06:45.569 --> 00:06:47.550 the capital to invest in the riskier layers. 00:06:47.790 --> 00:06:50.089 That's the strategy. You're tracking cost per 00:06:50.089 --> 00:06:52.329 transaction and straight through processing rate. 00:06:52.470 --> 00:06:54.269 Simple as that. All right, let's move up that 00:06:54.269 --> 00:06:57.470 risk curve. Layer two, collaborative augmentation. 00:06:57.910 --> 00:07:01.379 If layer one was the bond, this is the... growth 00:07:01.379 --> 00:07:04.279 equity of the portfolio. The realm of co -pilots. 00:07:04.319 --> 00:07:06.720 And the focus shifts. Right. It's not about cost 00:07:06.720 --> 00:07:08.660 reduction anymore. It's about capacity expansion. 00:07:08.939 --> 00:07:11.540 The human is now the collaborator. They keep 00:07:11.540 --> 00:07:13.500 the final say. And this is where we're seeing 00:07:13.500 --> 00:07:16.720 the most widespread adoption right now. The data 00:07:16.720 --> 00:07:18.980 on speed is pretty compelling. I mean, studies 00:07:18.980 --> 00:07:22.000 show developers using GitHub Copilot complete 00:07:22.000 --> 00:07:26.639 tasks 55 % faster. 55%. That's huge. It is. Time 00:07:26.639 --> 00:07:29.040 dropped from nearly three hours down to just 00:07:29.040 --> 00:07:31.439 over one. And we see similar things with, say, 00:07:31.519 --> 00:07:34.360 consultants using GPT -4. They finish tasks 25 00:07:34.360 --> 00:07:37.500 % more quickly. The premise is solid. Augmentation 00:07:37.500 --> 00:07:40.379 amplifies throughput. Okay. That speed is undeniable. 00:07:40.420 --> 00:07:42.759 But does this productivity boost come with a 00:07:42.759 --> 00:07:44.699 hidden price? This is where we hit the quality 00:07:44.699 --> 00:07:47.360 paradox. We do. Research analyzing code from 00:07:47.360 --> 00:07:49.620 these tools found a pretty big spike in what's 00:07:49.620 --> 00:07:51.779 called churn code. Code that's written and then 00:07:51.779 --> 00:07:54.220 almost immediately rewritten or deleted. Exactly. 00:07:54.459 --> 00:07:56.800 And they also found a decrease in code reuse. 00:07:57.120 --> 00:07:59.579 It strongly suggests that the savings you get 00:07:59.579 --> 00:08:01.579 in drafting time are essentially being mortgaged 00:08:01.579 --> 00:08:03.560 against future maintenance time. You're creating 00:08:03.560 --> 00:08:07.139 technical debt. The AI is fast, but the output 00:08:07.139 --> 00:08:09.680 might be messy. Right. And that debt will have 00:08:09.680 --> 00:08:11.759 to be paid down later by your very expensive 00:08:11.759 --> 00:08:14.920 senior engineers. And this leads right into what 00:08:14.920 --> 00:08:17.959 we call the human verification tax, because the 00:08:17.959 --> 00:08:21.000 AI operates on what researchers call the jagged 00:08:21.000 --> 00:08:23.459 technological frontier. That's such a great term 00:08:23.459 --> 00:08:26.500 for it. It is. It means the AI might nail a really 00:08:26.500 --> 00:08:29.740 complex task, but then fail at a super simple 00:08:29.740 --> 00:08:33.440 logic puzzle. The risk is unpredictable. Precisely. 00:08:33.440 --> 00:08:36.120 So the core cost of Layer 2 is the time the human 00:08:36.120 --> 00:08:38.899 spends reviewing and verifying, not just doing 00:08:38.899 --> 00:08:40.980 the work. In fact, there's evidence that for 00:08:40.980 --> 00:08:43.360 complex tasks, experienced developers reviewing 00:08:43.360 --> 00:08:46.240 AI code actually took 19 % longer to finish. 00:08:46.379 --> 00:08:49.159 Wait, wait. Say that again? 19 % longer? Longer. 00:08:49.159 --> 00:08:51.659 They spent more time debugging the AI's clever 00:08:51.659 --> 00:08:53.940 but flawed solutions than if they had just written 00:08:53.940 --> 00:08:57.100 the code themselves from scratch. Wow. That completely 00:08:57.100 --> 00:08:59.600 changes the ROI equation. The whole point is 00:08:59.600 --> 00:09:02.529 to save time. It does. And it means simply tracking 00:09:02.529 --> 00:09:05.269 hours saved is really misleading. You have to 00:09:05.269 --> 00:09:08.330 focus on quality adjusted output and the acceptance 00:09:08.330 --> 00:09:11.450 rate of the AI suggestions. For GitHub Copilot, 00:09:11.549 --> 00:09:15.019 that hovers around 30%. Only 30%. But to be fair, 00:09:15.179 --> 00:09:17.080 you also have to look at employee satisfaction. 00:09:17.580 --> 00:09:20.039 Layer 2 boosts that immensely because the AI 00:09:20.039 --> 00:09:22.620 handles the drudgery. And for general knowledge 00:09:22.620 --> 00:09:26.259 work, Microsoft claims an ROI of up to 353 % 00:09:26.259 --> 00:09:28.679 for small and mid -sized businesses. Hold on. 00:09:29.340 --> 00:09:32.799 353%. That number seems... Aggressive, especially 00:09:32.799 --> 00:09:34.399 with the technical debt and the verification 00:09:34.399 --> 00:09:36.340 tax we just talked about. It has to be qualified. 00:09:36.440 --> 00:09:38.799 Absolutely. So who's actually getting that return? 00:09:38.980 --> 00:09:41.539 You get that return when the tasks are not mission 00:09:41.539 --> 00:09:43.659 critical, where the cost of a mistake is low 00:09:43.659 --> 00:09:45.799 or where the human is just brilliant at prompting 00:09:45.799 --> 00:09:48.620 and correcting. If you apply that 353 percent 00:09:48.620 --> 00:09:50.759 number to high stakes work where that 19 percent 00:09:50.759 --> 00:09:53.820 verification tax kicks in, the ROI just plummets. 00:09:53.820 --> 00:09:55.720 OK, so that critical eye brings us to the highest 00:09:55.720 --> 00:09:58.470 risk tier, layer three. Autonomous decision making. 00:09:58.590 --> 00:10:00.970 The venture capital bet. This is the strategic 00:10:00.970 --> 00:10:03.649 frontier. It's all about scalability and labor 00:10:03.649 --> 00:10:06.730 substitution. The human role shifts again from 00:10:06.730 --> 00:10:10.409 collaborator to just a supervisor. Human on the 00:10:10.409 --> 00:10:13.129 loop. And the potential upside here is just. 00:10:13.789 --> 00:10:16.929 It's exponential. Agents can monitor and react 00:10:16.929 --> 00:10:20.330 24 -7. Your output is completely decoupled from 00:10:20.330 --> 00:10:23.149 human hours. And they're being priced very aggressively, 00:10:23.309 --> 00:10:26.450 maybe 20 to 60 percent of a full -time employee's 00:10:26.450 --> 00:10:28.730 costs. But this is where the financial risk becomes 00:10:28.730 --> 00:10:31.759 truly acute. If Layer 2 had a verification tax, 00:10:31.980 --> 00:10:34.580 Layer 3 has the massive risk of inference costs. 00:10:34.860 --> 00:10:37.120 This is the infinite loop trap, and it's crucial 00:10:37.120 --> 00:10:39.440 for financial planning because an agent has to 00:10:39.440 --> 00:10:41.320 think and generate a whole chain of reasoning 00:10:41.320 --> 00:10:43.299 before it gives you an answer. And all that thinking 00:10:43.299 --> 00:10:46.139 costs money. It consumes tokens. Exactly. So 00:10:46.139 --> 00:10:48.679 the cost volatility can just explode. There was 00:10:48.679 --> 00:10:51.159 a scenario analysis done on a fleet of just 500 00:10:51.159 --> 00:10:54.139 agents. They got stuck in a reasoning loop. They 00:10:54.139 --> 00:10:57.159 burned through $144 ,000 in compute costs in 00:10:57.159 --> 00:11:00.080 a single night. $144 ,000. thousand dollars in 00:11:00.080 --> 00:11:04.480 one night from one bad loop. Even a simple answer 00:11:04.480 --> 00:11:08.100 can require like six times the tokens and back 00:11:08.100 --> 00:11:10.460 end processing just to verify its own logic. 00:11:10.639 --> 00:11:13.220 That cost volatility is exactly why Gartner is 00:11:13.220 --> 00:11:15.299 warning that over 40 percent of these agentic 00:11:15.299 --> 00:11:17.779 AI projects are going to be canceled by 2027. 00:11:18.299 --> 00:11:20.539 But it's not just the financial cost, is it? 00:11:20.559 --> 00:11:23.320 It's the legal liability. When an agent is fully 00:11:23.320 --> 00:11:25.820 autonomous, it's making decisions for your company. 00:11:26.330 --> 00:11:29.289 And you own those decisions. The Air Canada precedent 00:11:29.289 --> 00:11:31.950 is the landmark case here. Right, the chatbot. 00:11:32.029 --> 00:11:34.350 The chatbot, acting as an agent, hallucinated 00:11:34.350 --> 00:11:37.129 a refund policy. A customer demanded it, and 00:11:37.129 --> 00:11:39.429 the court ruled that Air Canada was fully liable 00:11:39.429 --> 00:11:42.049 for its chatbot's negligent misrepresentation. 00:11:42.230 --> 00:11:44.549 You can't blame the AI. You carry the liability. 00:11:44.970 --> 00:11:47.070 So that's the professional risk. And then there's 00:11:47.070 --> 00:11:49.169 the brand damage risk, which can destroy value 00:11:49.169 --> 00:11:51.870 in an instant. Like the DPD delivery chatbot 00:11:51.870 --> 00:11:54.019 that was tricked into... cursing and criticizing 00:11:54.019 --> 00:11:56.480 the company. Or the Chevy dealer bought that 00:11:56.480 --> 00:11:59.620 agreed to sell a Tahoe for $1. Right. These examples 00:11:59.620 --> 00:12:02.220 show the lack of guardrails and the cost of losing 00:12:02.220 --> 00:12:04.879 one high value customer who might have a $5 ,000 00:12:04.879 --> 00:12:07.139 lifetime value because of a bad bot interaction 00:12:07.139 --> 00:12:10.299 is so much greater than any savings on that one 00:12:10.299 --> 00:12:12.700 call. And we're seeing companies realize this 00:12:12.700 --> 00:12:15.399 now. There's this reversion trend. Klarna, for 00:12:15.399 --> 00:12:18.080 example, they aggressively replaced staff with 00:12:18.080 --> 00:12:20.559 AI. And now they're reemphasizing human agents 00:12:20.559 --> 00:12:23.779 for that top 20 % of complex high -stakes queries. 00:12:24.179 --> 00:12:26.700 They learned that autonomy is great for volume, 00:12:26.799 --> 00:12:29.840 but it can be catastrophic for complexity. You 00:12:29.840 --> 00:12:32.179 still need that high -quality human safety net. 00:12:32.360 --> 00:12:34.720 An extremely high -quality one. Okay, so what 00:12:34.720 --> 00:12:37.740 does this all mean for your strategy? Let's synthesize 00:12:37.740 --> 00:12:40.080 this into a portfolio. The strategy comes down 00:12:40.080 --> 00:12:42.399 to your risk tolerance and time horizon. Layer 00:12:42.399 --> 00:12:45.519 one, automation is your cash cow. Reliable, fast 00:12:45.519 --> 00:12:48.500 ROI in three to nine months. Layer two, augmentation 00:12:48.500 --> 00:12:51.139 is your growth engine. High throughput, medium 00:12:51.139 --> 00:12:54.340 risk, ROI in maybe six to 12 months, as long 00:12:54.340 --> 00:12:56.539 as you manage that quality tax. And layer three, 00:12:56.659 --> 00:12:59.419 autonomy is the venture bet. It offers exponential 00:12:59.419 --> 00:13:01.659 scaling, but it requires tight governance and 00:13:01.659 --> 00:13:04.399 a longer 12 to 24 month time horizon to see a 00:13:04.399 --> 00:13:06.700 return. And I think a crucial insight here, one 00:13:06.700 --> 00:13:09.080 that people miss, is that autonomy does not mean 00:13:09.080 --> 00:13:12.690 zero human cost. Not at all. If your agents handle 00:13:12.690 --> 00:13:16.889 the easy 80 % and escalate the hard 20%. The 00:13:16.889 --> 00:13:19.389 humans you have left have to be incredibly expensive 00:13:19.389 --> 00:13:22.509 experts who can untangle the AI's complex mistakes. 00:13:22.870 --> 00:13:25.289 That's the human -in -the -loop cost curve, exactly. 00:13:25.330 --> 00:13:27.610 You bring the volume of human work down, but 00:13:27.610 --> 00:13:29.809 the average hourly cost of your human workforce 00:13:29.809 --> 00:13:32.470 goes way up. They have to be elite specialists. 00:13:32.889 --> 00:13:36.309 So to mitigate those huge Layer 3 risks that 00:13:36.309 --> 00:13:39.169 cost volatility, we have to implement PhenOps 00:13:39.169 --> 00:13:42.899 for AI. This feels essential. It is non -negotiable. 00:13:42.980 --> 00:13:45.700 Traditional IT spending is predictable. AI spending 00:13:45.700 --> 00:13:48.759 is not. FindOps here means putting strict, automated 00:13:48.759 --> 00:13:51.299 financial controls on these probabilistic resources. 00:13:51.820 --> 00:13:54.340 So how do you stop that $144 ,000 night from 00:13:54.340 --> 00:13:56.600 happening? The most practical solution is a circuit 00:13:56.600 --> 00:13:59.100 breaker. It's a hard limit on token spend. For 00:13:59.100 --> 00:14:01.200 example, you kill any agent process immediately 00:14:01.200 --> 00:14:03.899 if its reasoning path costs more than, say, $5 00:14:03.899 --> 00:14:06.279 per task. So it's a technical failsafe tied directly 00:14:06.279 --> 00:14:09.090 to the budget. Exactly. You couple that with 00:14:09.090 --> 00:14:12.009 rate limiting on API calls so one runaway agent 00:14:12.009 --> 00:14:15.090 can't hog all your resources. And you absolutely 00:14:15.090 --> 00:14:18.230 must have observability tools to see the agent's 00:14:18.230 --> 00:14:21.169 chain of thought. If you can't audit the reasoning, 00:14:21.330 --> 00:14:23.409 you can't fix the problem. And what about on 00:14:23.409 --> 00:14:26.110 the security side, protecting it from bad actors? 00:14:26.309 --> 00:14:29.679 You need agentic security. Agents have to operate 00:14:29.679 --> 00:14:32.580 with the principle of least privilege. An agent 00:14:32.580 --> 00:14:34.860 that processes refunds should never have access 00:14:34.860 --> 00:14:37.279 to the full customer database. Right. And you 00:14:37.279 --> 00:14:40.500 must have robust prompt injection defenses to 00:14:40.500 --> 00:14:43.100 stop people from tricking your agent into, well, 00:14:43.159 --> 00:14:45.500 selling a Tahoe for a dollar. So we've laid out 00:14:45.500 --> 00:14:47.960 the clear strategic steps for you. Start with 00:14:47.960 --> 00:14:50.529 layer one for reliable costs. Use that to fund 00:14:50.529 --> 00:14:53.090 the others. Train your experts to be skeptics 00:14:53.090 --> 00:14:55.570 for layer two and impose strict phenops and security 00:14:55.570 --> 00:14:58.870 on every single layer three deployment. The organizations 00:14:58.870 --> 00:15:00.570 that are going to win this transition will be 00:15:00.570 --> 00:15:03.009 the ones that master the financial and the operational 00:15:03.009 --> 00:15:05.889 orchestration of this complex, half human, half 00:15:05.889 --> 00:15:08.809 AI workforce. It's about mastering the tech and 00:15:08.809 --> 00:15:11.070 the ledger. So since layer three is the highest 00:15:11.070 --> 00:15:13.330 risk, but also promises to fully decouple revenue 00:15:13.330 --> 00:15:15.570 from headcount, here's a final provocative thought 00:15:15.570 --> 00:15:18.009 for you to chew on. If your most strategically 00:15:18.009 --> 00:15:20.610 valuable decisions are being made by probabilistic 00:15:20.610 --> 00:15:22.750 systems that you can't fully audit or predict, 00:15:22.950 --> 00:15:25.730 and yet your company still carries 100 % of the 00:15:25.730 --> 00:15:28.850 legal and financial liability, what happens to 00:15:28.850 --> 00:15:31.549 your organizational structure? It forces a complete 00:15:31.549 --> 00:15:33.409 rethinking of corporate risk management that 00:15:33.409 --> 00:15:36.190 goes far beyond compliance and asks a new question. 00:15:36.450 --> 00:15:39.009 How do you ensure against algorithmic behavior? 00:15:39.889 --> 00:15:42.350 That wraps up this week's strategic briefing. 00:15:42.529 --> 00:15:44.850 The signal for this week is commercialization. 00:15:45.500 --> 00:15:48.240 We are moving past the era of wild experimentation 00:15:48.240 --> 00:15:52.360 and into the era of hard business. Meta is closing 00:15:52.360 --> 00:15:54.940 its doors on open source to focus on revenue. 00:15:55.539 --> 00:15:58.480 Content giants like Disney are choosing exclusive 00:15:58.480 --> 00:16:01.120 partners and suing the rest. And governments 00:16:01.120 --> 00:16:03.679 are regulating the flow of chips to China, not 00:16:03.679 --> 00:16:07.500 to stop the trade, but to take a 25 % cut. AI 00:16:07.500 --> 00:16:10.559 is no longer just a technology. It is the world's 00:16:10.559 --> 00:16:13.580 most valuable commodity. Before we go, a reminder 00:16:13.580 --> 00:16:15.830 for the leaders listening. Stop marketing to 00:16:15.830 --> 00:16:19.049 the masses. Start briefing the C -suite. Leverage 00:16:19.049 --> 00:16:21.309 our proven methodology to own the conversation 00:16:21.309 --> 00:16:24.070 in your industry. We create tailored proprietary 00:16:24.070 --> 00:16:27.190 podcasts designed exclusively to brief your executives. 00:16:27.610 --> 00:16:30.190 Stop wasting marketing spend on generic content. 00:16:30.809 --> 00:16:33.110 Start delivering strategic intelligence directly 00:16:33.110 --> 00:16:35.730 to the decision makers. Ready to define your 00:16:35.730 --> 00:16:38.190 domain? Secure your strategic podcast consultation 00:16:38.190 --> 00:16:41.110 now at the link in our show notes. This podcast 00:16:41.110 --> 00:16:43.769 is created and produced by Etienne Noman, senior 00:16:43.769 --> 00:16:46.190 software engineer and passionate soccer dad from 00:16:46.190 --> 00:16:48.929 Canada. Please subscribe and share. Until next 00:16:48.929 --> 00:16:51.289 week, keep unraveling the future. Let's go.