AI Weekly Briefing: Accenture Goes All In on Claude

This week I break down the biggest moves shaping the future of AI. OpenAI’s new enterprise report shows explosive adoption and usage growth, IBM drops eleven billion on Confluent to own the data-streaming backbone, Accenture goes deep with Anthropic in a partnership that could reshape corporate AI, and Runway’s Gen 4.5 quietly leapfrogs the giants in video generation. We also look at Instacart’s new agentic shopping experience inside ChatGPT, Anthropic’s push to bring Claude to nonprofits, and why Mistral’s open models are becoming a serious force.

Episode Transcript

(00:00):
Hi everyone.

(00:00):
I'm Andy, and this is the AI Breakdown.
Welcome to your weekly news edition where I'll cover what happened in AI last week, why it matters all in about 10 minutes. 4 00:00:14,577.687 --> 00:00:22,947.687 Let's start with OpenAI because their new state of Enterprise AI 2025 report is packed with clues about where this is all heading. 5 00:00:23,427.687 --> 00:00:31,827.687 Headline number one, 36% of large US enterprises in OpenAI survey are now using Chat GPT Enterprise or team. 6 00:00:32,472.687 --> 00:00:40,632.687 That's more than one in three big companies with a paid structured rollout, not just a few keen staff playing with the free tier for comparison. 7 00:00:40,662.687 --> 00:00:48,372.687 Around 14% of those enterprises are using Anthropics Claude, so OpenAI still has a very real lead in Enterprise Mindshare. 8 00:00:48,822.687 --> 00:00:50,772.687 The usage data is just as punchy. 9 00:00:50,772.687 --> 00:00:55,932.687 Enterprise message volume on chat, GPT is up eight times year on year. 10 00:00:56,442.687 --> 00:01:03,12.687 Custom GBTs, those company specific assistance that capture your internal knowledge are up 19 times. 11 00:01:03,222.687 --> 00:01:15,552.687 And now make up around one fifth of all enterprise messages and at an individual level, employees using these tools report saving around 45 minutes a day on average with many seeing it's closer to an hour. 12 00:01:16,182.687 --> 00:01:18,822.687 Open AI's leadership is clearly feeling the pressure though. 13 00:01:19,302.687 --> 00:01:27,492.687 Sam Altman sent an internal code red memo over Google's Gemini three basically saying that this is a serious threat and we need to move faster. 14 00:01:28,2.687 --> 00:01:30,972.687 That's part of why they rushed out GPT 5.2 15 00:01:31,32.687 --> 00:01:32,652.687 just weeks after 5.1. 16 00:01:33,252.687 --> 00:01:38,172.687 I think the big takeaway for you is this, the AI race isn't just about model quality anymore. 17 00:01:38,412.687 --> 00:01:40,692.687 It's about distribution and integration. 18 00:01:41,202.687 --> 00:01:44,682.687 So competition in AI is really about ecosystem control. 19 00:01:45,282.687 --> 00:01:48,492.687 Google has Gemini wired into search Gmail and Android. 20 00:01:49,2.687 --> 00:01:56,382.687 OpenAI is trying to answer with deeper enterprise features, custom gpt, and now commerce integrations like Instacart. 21 00:01:56,952.687 --> 00:02:13,771.77790909 For you, the question is, are you treating AI as just another app or as a platform you're actively re-platforming around? Next up a huge bit of plumbing news that won't trend on X, but really matters if you care about building AI products. 22 00:02:14,191.77790909 --> 00:02:18,61.77790909 IBM is buying confluent for about $11 billion in cash. 23 00:02:18,661.77790909 --> 00:02:21,301.77790909 Confluent is basically Kafka as a service. 24 00:02:21,541.77790909 --> 00:02:27,811.77790909 The real time event streaming layer that lets you move data around your organization in milliseconds instead of ours. 25 00:02:28,261.77790909 --> 00:02:33,241.77790909 Think click streams, payments sensor data, customer events, all flowing continuously. 26 00:02:33,241.77790909 --> 00:02:40,621.77790909 I, IBM's, CEO of and Krishna called it the Smart Data Platform for Enterprise IT Purpose built for ai. 27 00:02:41,296.77790909 --> 00:02:42,196.77790909 Translation. 28 00:02:42,466.77790909 --> 00:02:49,756.77790909 If IBM owns the pipe that feeds data into AI systems, it becomes much harder for those customers to rip IBM out later. 29 00:02:50,416.77790909 --> 00:02:55,276.77790909 Investors seem to agree IBM is paying a 34% premium over Confluence. 30 00:02:55,276.77790909 --> 00:02:59,866.77790909 Last stock price and Confluence shares jumped around 30% on the news. 31 00:03:00,316.77790909 --> 00:03:07,186.77790909 This is their biggest acquisition since Red Hat, which tells you how serious they are about owning the AI infrastructure stack. 32 00:03:07,561.77790909 --> 00:03:10,711.77790909 If you strip away the m and a gloss, the lesson is simple. 33 00:03:11,11.77790909 --> 00:03:14,521.77790909 Most AI projects don't fail because the model isn't good enough. 34 00:03:14,851.77790909 --> 00:03:18,421.77790909 They fail because the data is late, messy, or locked away. 35 00:03:18,961.77790909 --> 00:03:24,781.77790909 IBM is betting that in the AI era, the real money is in solving that data plumbing problem end to end. 36 00:03:25,291.77790909 --> 00:03:34,801.77790909 So if you're building anything AI powered, it's worth asking, are you obsessing over which model to use? While your data pipeline looks like it's held together with duct tape and hope. 37 00:03:36,902.687 --> 00:03:45,662.687 Sticking with big enterprise moves, Accenture and Anthropic have just announced a multi-year partnership that frankly could shift a lot of corporate AI spend. 38 00:03:46,532.687 --> 00:03:57,332.687 Accenture is creating an Accenture Anthropic business group and trading about 30,000 of its people on Claude consultants, engineers, AI reinvention specialists, the lot. 39 00:03:57,872.687 --> 00:04:02,612.687 Claude Code and Tropics developer assistant already has a big chunk of the AI coding market. 40 00:04:02,942.687 --> 00:04:06,932.687 Accenture is now rolling it outta tens of thousands of developers. 41 00:04:07,472.687 --> 00:04:09,782.687 This isn't just another logo slide partnership. 42 00:04:10,172.687 --> 00:04:21,962.687 Accenture is baking Claude into solutions for heavily regulated sectors, finance, healthcare, the public sector where governance order trails and safety matter as much as raw capability. 43 00:04:21,962.687 --> 00:04:25,742.687 Their co-developing tools to help banks automate compliance. 44 00:04:26,102.687 --> 00:04:32,282.687 Pharma firms accelerate r and d, and CIOs actually measure AI value, not just talk about it. 45 00:04:32,852.687 --> 00:04:37,772.687 Why does this matter? Because a lot of enterprise AI is stuck in proof of concept purgatory. 46 00:04:38,102.687 --> 00:04:45,482.687 The tech works, but nobody knows how to integrate it with legacy systems, risk teams are nervous and staff aren't trained. 47 00:04:45,902.687 --> 00:04:48,842.687 Accenture is basically saying, we'll fix that for you. 48 00:04:49,82.687 --> 00:04:51,992.687 And by the way, our weapon of choice is clawed. 49 00:04:52,412.687 --> 00:04:56,942.687 One thing to watch is how quickly anthropic share of the enterprise market grows from here. 50 00:04:57,302.687 --> 00:05:01,292.687 They're already up from roughly a quarter to around 40% by some measures. 51 00:05:01,712.687 --> 00:05:08,852.687 Giving 30,000 Accenture consultants, a strong opinion in Claude's favor is going to tilt a lot of boardroom conversations. 52 00:05:10,249.05063636 --> 00:05:13,39.05063636 Let's bring it down from boardrooms to the weekly shop. 53 00:05:13,669.05063636 --> 00:05:22,369.05063636 OpenAI and Instacart have launched a new Instacart chat GPT app that lets you plan meals, fill your basket and check out without ever leaving chat. 54 00:05:22,369.05063636 --> 00:05:23,59.05063636 GPT. 55 00:05:23,449.05063636 --> 00:05:27,619.05063636 You literally type something like I've got four people coming over, one's vegan. 56 00:05:28,39.05063636 --> 00:05:30,319.05063636 Give me a menu and help me shop and chat. 57 00:05:30,319.05063636 --> 00:05:31,999.05063636 GPT proposes recipes. 58 00:05:32,404.05063636 --> 00:05:39,694.05063636 Pipes, the ingredients into an Instacart cart lets you tweak items in the chat and then runs what they call instant checkout. 59 00:05:40,354.05063636 --> 00:05:42,604.05063636 All of that happens inside chat, GPT. 60 00:05:42,994.05063636 --> 00:05:47,434.05063636 You don't flip over to the Instacart app or a browser tab under the hood. 61 00:05:47,764.05063636 --> 00:05:56,614.05063636 This is built on open AI's, new agentic commerce tooling, which lets chat GPT, not just recommend things, but actually transact on your behalf. 62 00:05:56,914.05063636 --> 00:06:09,394.05063636 Once you've given it permission, why is this interesting beyond dinner? Because it's one of the cleanest demos yet of conversational commerce, going from intent to plan to paid order in a single dialogue. 63 00:06:10,24.05063636 --> 00:06:12,604.05063636 For Instacart, it's a new acquisition channel. 64 00:06:12,904.05063636 --> 00:06:18,964.05063636 They suddenly show up wherever chat GPT users are brainstorming meals for OpenAI. 65 00:06:19,54.05063636 --> 00:06:20,944.05063636 It's a potential new revenue stream. 66 00:06:21,214.05063636 --> 00:06:23,854.05063636 They'll take a small fee on sales that go through chat. 67 00:06:23,854.05063636 --> 00:06:25,804.05063636 GPT, zooming out. 68 00:06:26,134.05063636 --> 00:06:30,724.05063636 This is what AI agents are going to look like for consumers, not sci-fi robots. 69 00:06:30,934.05063636 --> 00:06:34,594.05063636 Just very competent flows that quietly remove whole chunks of friction. 70 00:06:35,164.05063636 --> 00:06:41,524.05063636 Today, it's groceries tomorrow, it's booking travel, renewing SaaS tools or reordering office supplies. 71 00:06:41,884.05063636 --> 00:06:48,814.05063636 If you run an e-commerce or SaaS business, it's worth thinking about how your product might live inside someone else's AI interface. 72 00:06:50,984.50518182 --> 00:06:55,304.50518182 On the creative side of AI runway has pulled off a proper David versus Goliath. 73 00:06:55,304.50518182 --> 00:06:59,234.50518182 Moment in video generation, they've launched Gen 4.5, 74 00:06:59,594.50518182 --> 00:07:03,284.50518182 a new text, a video model that shot straight to the top of the video arena. 75 00:07:03,284.50518182 --> 00:07:17,204.50518182 Leaderboard run by an independent evaluator called Artificial Analysis and blind Tests where humans pick their favorite clip without knowing which model made it runway's output beat Google's latest VO three and open the eyes. 76 00:07:17,234.50518182 --> 00:07:18,374.50518182 So a two pro. 77 00:07:18,929.50518182 --> 00:07:20,759.50518182 The highest ELO score of any model. 78 00:07:20,759.50518182 --> 00:07:24,389.50518182 So far, the big improvement is in physics and motion. 79 00:07:24,809.50518182 --> 00:07:25,169.50518182 Earlier. 80 00:07:25,169.50518182 --> 00:07:33,509.50518182 Video models often produce that weird floaty look, objects sliding instead of walking fabric that doesn't quite behave water, that feels like jelly. 81 00:07:34,229.50518182 --> 00:07:35,309.50518182 Gen 4.5 82 00:07:35,309.50518182 --> 00:07:47,279.50518182 has been trained on lots of high frame rate footage and tuned for motion consistency, so things like weight, momentum, smoke, hair, and camera moves all hang together much more convincingly over time. 83 00:07:47,909.50518182 --> 00:07:52,139.50518182 Crucially, they claim they've kept the speed and efficiency of the previous Gen four model. 84 00:07:52,439.50518182 --> 00:07:54,209.50518182 So you're not waiting forever for renders. 85 00:07:54,539.50518182 --> 00:07:58,649.50518182 For filmmakers, marketers, and game studios, this is a serious new tool. 86 00:07:59,99.50518182 --> 00:08:09,179.50518182 You can pre-visualize scenes, spit out concept shots for a pitch, or generate social content with something much closer to cinematic quality without needing a full VFX team. 87 00:08:09,719.50518182 --> 00:08:14,819.50518182 And the angle I like here is that Runway did this with a team of about 100 people. 88 00:08:15,329.50518182 --> 00:08:22,799.50518182 As their CEO Cristobal Valenzuela put it, they managed to outcompete trillion dollar companies by staying extremely focused. 89 00:08:23,99.50518182 --> 00:08:31,979.50518182 How cool is that? So if you are a smaller player, building on a niche, this is a nice reminder that you absolutely can beat the Giants on specific problems. 90 00:08:33,468.59609091 --> 00:08:38,388.59609091 Moving on to a quieter story that I think is genuinely encouraging clawed for nonprofits. 91 00:08:38,988.59609091 --> 00:08:47,88.59609091 Anthropic has partnered with Giving Tuesday to offer eligible nonprofits a 75% discount on Claude's team and enterprise plans. 92 00:08:47,478.59609091 --> 00:08:52,878.59609091 On top of that, they've built connectors into tools that social good organizations already live in. 93 00:08:53,298.59609091 --> 00:08:55,998.59609091 Platforms like Blackboard, candid and Benty. 94 00:08:56,538.59609091 --> 00:09:01,458.59609091 In practice, that means a charity can ask Claude questions directly against its donor database. 95 00:09:01,998.59609091 --> 00:09:10,818.59609091 Pull grant information from Candid or use Benevity data to find local organizations working on say, homelessness, all from within the chat interface. 96 00:09:11,298.59609091 --> 00:09:19,698.59609091 Andro has also launched a free AI fluency for non-profits course to help teams that don't have in-house data scientists actually use this stuff. 97 00:09:19,698.59609091 --> 00:09:28,308.59609091 Well, and this isn't just theory, they highlight groups like the International Rescue Committee, which uses Claude to analyze field data faster in crisis zones. 98 00:09:28,683.59609091 --> 00:09:35,583.59609091 And research orgs like ID insight, which report working up to 16 times faster on Serbian analysis work. 99 00:09:36,183.59609091 --> 00:09:37,563.59609091 I like this for two reasons. 100 00:09:37,563.59609091 --> 00:09:42,393.59609091 One, nonprofits are usually lost in line for new tech because budgets are tight. 101 00:09:42,843.59609091 --> 00:09:48,663.59609091 A 75% discount plus training and integrations is a serious attempt to change that. 102 00:09:49,113.59609091 --> 00:09:54,213.59609091 Two, it's a concrete example of AI amplifying human work rather than replacing it. 103 00:09:54,603.59609091 --> 00:09:58,53.59609091 Freeing obs scarce staff to focus on relationships and impact. 104 00:09:58,428.59609091 --> 00:10:08,745.86881818 While the model chews through the paperwork, let's finish with Misra because they're taking a very different route to the big US labs and it's starting to pay off. 105 00:10:09,105.86881818 --> 00:10:25,95.86881818 The Paris based team has released Misra three, a family of 10 open weight models, ranging from a big frontier class multimodal model, down to tiny 3 billion and 8 billion parameter models that can run on a laptop, a server at the edge, even a drone. 106 00:10:25,470.86881818 --> 00:10:32,70.86881818 Crucially, they've published the weights under permissive licenses so you can download fine tune and host them yourself. 107 00:10:32,520.86881818 --> 00:10:33,720.86881818 No API lock-in. 108 00:10:33,840.86881818 --> 00:10:37,680.86881818 No sending sensitive data to someone else's cloud if you don't want to. 109 00:10:38,40.86881818 --> 00:10:46,860.86881818 RA's argument backed up by customer work is that the huge majority of enterprise use cases don't actually need a monster 70 billion parameter model. 110 00:10:47,610.86881818 --> 00:10:50,850.86881818 Businesses often prototype on a giant closed model. 111 00:10:50,935.86881818 --> 00:10:51,505.86881818 Then discover. 112 00:10:52,170.86881818 --> 00:10:53,910.86881818 It's slow and expensive at scale. 113 00:10:54,480.86881818 --> 00:11:03,0.86881818 At that point, they come to Mara, fine tune a smaller model on their own data and get 80 to 90% of the quality at a fraction of the cost. 114 00:11:03,660.86881818 --> 00:11:11,430.86881818 We've already seen their models turning up on platforms like IBM's Watson X with Mistral three that're doubling down on being the open source champion. 115 00:11:12,510.86881818 --> 00:11:17,580.86881818 If you're running in regulated environments or you just hate vendor locking, this really matters. 116 00:11:17,820.86881818 --> 00:11:26,550.86881818 It means you can design an AI stack where some workloads run on closed models from open AI or anthropic and others sit on open models you control fully. 117 00:11:27,240.86881818 --> 00:11:31,410.86881818 That mix and match world is probably where a lot of serious enterprises will end up. 118 00:11:31,860.86881818 --> 00:11:39,660.86881818 And from a European perspective, it's significant that one of the most credible open model families is coming outta Paris, not Silicon Valley. 119 00:11:40,155.86881818 --> 00:11:43,785.86881818 It slightly rebalances who gets to shape the AI ecosystem. 120 00:11:45,738.59609091 --> 00:11:47,628.59609091 That's all for this week's AI roundup. 121 00:11:47,958.59609091 --> 00:11:52,38.59609091 If you found value in this breakdown, please leave a rating and hit subscribe. 122 00:11:52,248.59609091 --> 00:11:53,268.59609091 See you next week.