EP. 002: AI for FinOps: April Summit Recap (Agents and Tokens!)

98% of Phenops practices now manage AI spend. Two years ago, that number was 31%. Welcome to the Phenops daily brief. I'm JR Stormin and I am coming to you from San Diego, California. Today, I'm going to be recapping yesterday's April Phenops Foundation virtual summit. The theme for this summit was AI for Phenops. 60 minutes of talks and contents over 2,200 people connected over the 90 minute call 1.5,000 people concurrently online. And there was one big fat dead moment of silence at the end that I will tell you all about at the end of this podcast. So let's get into it. Here's the headline. Token economics has officially arrived. I said this on the call. I'll say it again. All I'm hearing every day in executive conversations right now is about token economics, token value, token costs, attaching value to the scaling costs of AI. This is a real change, C change from where we were six months ago or in reality, even just three months ago. Frederick Poll from SAP put it like this in his Phenops X keynote preview and this line is worth quoting in full. Tokens are where the industry is converging. Our job now as a Phenops organization is to maximize the value of token transformation. Jensen called it the token factory effectiveness. That's what we all need to run toward. Token factory effectiveness. Let's sit with that one for a minute. This phrase kind of reframes the Phenops job. We're not minimizing cloud bills. We're not even maximizing the value of those cloud bills. We're maximizing the output of a token factory whose raw materials is digital and whose product is business value. The whole summit yesterday and away was an attempt to figure out what that actually means for all of you on a Monday morning. First story from the summit. The ITAM forum is now officially part of the Linux Foundation. Chartered as a sister organization to the Phenops Foundation, it's going to be the same structure, a governing board, a directed fund, eventually a technical advisory council, an independent brand, and some fantastic ISO work on the side. A big thank you to Martin Thompson who built that community over many years and is stepping back to focus on other ventures. On the summit yesterday, Ron Brill joined who is the former chair of the Trustee Board, the old Trustee Board, and is now going to move over to the governing board, along with Salame Keat, who is also on that board. And Salame is actually going to be flying all the way out from South Africa for Phenops X to join us there where they're going to have their first governing board meeting of the ITAM forum. So why does this matter? Phenops and ITAM are kind of two sides of a very similar governance coin. ITAM historically has owned risk, compliance, been focused on the contracts and the paperwork. Phenops has owned or focused on value and cost. But the line between these was always somewhat artificial. Now both communities sit under the same roof at the Linux Foundation. Many organizations keep these disciplines separate, but some are integrating them and some are fully merging teams. So expect more cross-pollination between our two sister orgs on best practices on data and on events. We have an ITAM chok tok happening at Phenops X with Ron, Salame, and Omkar Gakal from Salesforce and Savinas, Toykova from TPiCAP. So if you're interested in that area, check out the chok tok at Phenops X. Second story, Jonathan Morley from the Phenops Foundation staff gave us a walkthrough on AI for Phenops as a primer. He did 10 slides in 10 minutes. Nice work, Jonathan pulled it off on time. The framework, he laid out, goes something like this. There is a maturity path here. At the bottom of that path, you have disconnected spreadsheets and manual reporting. Above that, a real data foundation was shared KPIs and standardized data economics. Above that, you start to use generative AI to summarize data and answer questions on demand. And at the top of that framework, Agentate AI enters the conversation to monitor, investigate, and act autonomously. Most companies are not at the top yet, and that's fine. The point is knowing where you are. Two practitioner quotes from Jonathan's deck, "Stuck with me." The first come came from Alex Landis at Snowflake. Alex said, and I quote, "Without having your data foundation built out in a way that actually operationalizes it, AI doesn't really mean anything." Even clear to the point, AI on top of messy data gives you confidence sounding garbage. The second came from a senior practitioner at a North American tech company. That person said, "We see 40% almost 50% action rate on the messages we send out via Slack from our agents. Agents identify the right owners and then send personalized notes. 40% to 50% action on cost alert messages." That's a pretty good number. You've never seen an internal Slack about cost cleanup. You know how effective that is. Jonathan also gave us six use cases that he laid out in the deck, or occurred you to go check out the recording. First one is natural language, dashboards, autonomous waste discovery was high on that list. Pipelines and guardrails that check infrastructure as code before deployment, personalized Slack outreach as we discussed. Automated pull requests for governance findings and contextual labeling that infers missing tags. None of these examples are theoretical. They were taken from our advanced community conversations where people are testing out and implementing these now in production as well. Third story, I met Kenha from do it. I met Kenha from do it. I can say that. Who many of you know from his time previously at Goldman at City on the focus steering committee now on the fed house foundation governing board gave us a crawl walk run framework for AI in thin ups. Crawl is conversational AI as omit put it. You ask a question of your cost data, you get an answer. Walk is beginning to shift left. He said agents review pull requests and infrastructure as code before anything ships run is fully agentic agent to text then investigates recommends and ultimately acts. There's two lines that I'm in mentioned that you should pay attention to. The first one is this. The first thing AI does in thin ups isn't save you money. It shows you how much cleanup you have to do. And if you plug a conversational layer into messy data, the cracks really light up. If you're tagging is only 40% complete. If your allocation has a bunch of tribal knowledge baked into it that's not clear. If spreadsheets are papering over bad data, the ad doesn't know any of that and doesn't have the context it needs. Because some teams may use this as a motivation to finally fix the data foundation. The teams are going to blame the AI is not being effective and stay stuck. So this is where we get to make a choice. This is which group we want to be in. The second line he said is you want to automate the wrong things first. This was a warning he shared from he's seen a lot of people do this, which is don't start automating the things that are hard expensive or hard to undo the one way doors. Don't make purchases, don't do production right sizing. One bad call, CFO pulls the plug on the whole program. So start with the boring stuff when it comes to automation. Had the agents do tagging enforcement, had them do idle resource cleanup in staging environments, focus on non-prod shutdown schedules. These are all things we've talked about in Phenops for years with basic machine learning and other types of automation. Inagentic where there's potentially more trust and more authority, you need to even more than ever build trust on the small things before you start to go big. And I'm ended with this line, which I think is worth repeating as well. He said the AI part is easy getting engineering to listen is hard. So just like in traditional Phenops in a preagentic world, if you're bought just like the problems that engineers need to go out and figure it out and fix it, you're going to be stuck in the same paradigm as before. So ideally your bot is suggesting a specific alternative with a concrete action and a concrete impact or savings or redirection efficiency that engineers can easily act on instead of having to do a root cause analysis and look for a proposal. So be a service to your engineering partners. Don't be a toll booth for them. Fourth story, Dan Berg from Hype to Helpful. Dan is now leading Phenops at Squarespace and he gave us a preview of his upcoming Phenops X Talk, which is titled Hype to Helpful. Deathframing was pretty simple. The gap is between knowing that it's table stakes and knowing what to actually do on Monday morning. He showed two ways to ask the same question that now you've asked was this, hey, AI, can you write me a sequel to find expensive GKE namespaces? You might get a plausible looking query, but it could be the wrong table, the wrong labels, numbers that you might have almost trusted. Then he shared a pattern there. You want to make sure you've got the context in place, schema, label taxonomy, billing export structures. Make it a more constrained ask. Have it focus on the top 10 namespaces by cost this month, potentially grouped by team label. Do a skeptical review of the initial outputs? What assumptions did you make in that process? What assumptions did the AI make? What could be wrong? Some question very different results if you take that approach. Dan's bigger idea is what he calls an AI home base. It's a directory you control. It's a set of skills you built. It's MCP servers connected to your tools. It's memory that compounds over time. The idea is that context needs to live somewhere, so make that somewhere yours. Lower the activation energy on Monday by having a place that already knows who you are. Dan also said something that I want to further underline. He said that AI is the first technology abstraction layer in the history of computing, where we don't have visibility into how the AI layer works. And the old days seek would compile to assembly and you can trace it all the way down the stack Python runs on C and you can trace that. LLMs do something that we cannot inspect. Even a lot of the researchers building them don't fully understand all the deep mechanics happening in there. So that changes the trust contract. You have to be skeptic by default. Fifth story. Focus, the Phenops Open Cost and Usage Specification. An update on Focus came from that CalSert who shared a steady drumbeat of progress in the working group and in billing generator adoption. Yesterday, Redis publicly announced support for Focus 1.2 with documentation and how to download your cost reports. Grafana moved into a general availability at GA with hourly granularity, which is a big ask for 1.2 support this last week. Snowflake, MongoDB, Oracle, and Tencent also have planned 1.3 support coming this year. We've also just opened the conformance program for 1.3, Matt shared that certified data generators will get announced at Phenops X in June in San Diego. The 1.4 release is coming in June with four big additions. One, standalone invoice and billing period data sets. So you can now reconcile invoices against usage line items even across different billing and payment currencies. Two, we are roughly doubling the number of contract commitment columns, which are going to better describe how a commitment might be structured, its lifecycle, and its payment model. Three, a new correction handling attribute will tell you how data generators are surfacing corrections in their data. And four, there's a new structure for commitment program eligibility so you can see how you could be saving even if you haven't bought the commitment yet. Matt told us that he's actually been having fun building this open-source specification through these last couple of releases. And that is sometimes a sentence that you don't hear often when you're deep in the weeds of getting something like this out the door. So I appreciate Matt for his excitement, his enthusiasm, and the whole team, Sean Elpe and Joaquin Prado and Eurena there working hard on getting the focus spec out to all of you. If you're a focus user, I will please do go. I request for you to go request to your providers to ask them to support Focus 1.3. If you're a service or billing generator, go get certified in the upcoming performance program. And we are getting ready to start scoping Focus 1.5 as 1.4 development wraps. So get to Matt or Sean or anybody on the focus team with your ideas for future releases. Six stories. Google Cloud, Shruthy Nambi from Google Cloud walked us through their April releases at Google Next from last week. There's a real shift happening in how the clouds are thinking about Fennops tooling in this AI era and three things stood out from her talk. First thing was that they launched a Fennops AI explainability agent, which is now on the front page of the Google Cloud billing console. It's proactive, not reactive, those terms that we've been promised for a long time from many folks. And this one actually tells you why your costs change before you even ask. It can break down spend by model, by modality, by token input versus output. Interesting. By API key, by service account, and the skill set is growing throughout the rest of the year. Second thing that stood out was spending caps on Google Cloud budgets now in private preview. The shift here is from alerting to enforcement, setting an actual cap. If you hit that cap, the usage on the specific service and the specific project gets blocked. Doesn't delete the resources, it doesn't do collateral damage to other projects, but it does stop further usage. This is currently just available for GIMNI API agent platform Cloud Run, Cloud Run functions, and maps. And is an interesting play by Google in the sense of providing a way to handcuff usage versus alerting on thresholds when they are exceeded. Third thing that stood out from Shruthy's talk is that anomaly detection at Google Cloud went into GA. And they added some AI-generated thresholds of default alerts on every project and protection from day one, even on accounts with no spend history. You can add billing account groups, consolidate cost reports, commitment tracking reports, and see the playbook easy to see control and predict. So I think this is going to be a big theme at PhenomSex. We're expecting from all of the clouds, Google will be their AWS, Microsoft, and Oracle. We're expecting a lot of new feature announcements from the clouds at the conference as we saw last year. In fact, what's really interesting is that last year some of them chose to announce some of those features at the PhenomSex conference and not at their own. And it's a really big signal about where the PhenomS practitioners are gathering in the PhenomS ecosystem right now. So the last moment I want to share with you from the summit was something that surprised me a little bit. I would say we'll call it the crickets moment. Toward the end of the call, one of the very last questions in the Q&A, I asked the panel of esteemed experts a very simple question. I said, can one of you share an example where in your own company or with your clients, you had a demonstrable value outcome from applying AI to your PhenomS practice rather than say a six hour cloud session without an outcome. And the result was silence on the call. No one raised their hand. I want to sit with that for just a second. We spent an hour talking about agents and token economics and explainability and crawl walk run. We have practitioners who are clearly building, we have action rates of 40 to 50% on Slack outreach. And when I ask for demonstrable outcomes and stories there, the room would quiet. So here's what I'm thinking. We're still really early. Nobody has really figured this out. It really reminds me of 2017, 2018 PhenomS land where nobody really had the answers on cloud yet either. So people are showing up at the conference. They're sharing what they're trying. They're sharing what didn't work. And we're going to build this AI management practice together out in the open, both how we're applying AI to PhenomS and how we're using PhenomS to manage the value of AI. And that's exactly where things are right now. No one has figured it out. Not the practitioners, not the clouds, not the vendors, not the PhenomS foundation. We're all in this room together. And the people who are going to lean in and help set those patterns will be those who are setting the stage and allowing us to build frameworks and best practices for the next few thousand teams that need to follow and do this in the coming years. So if you shift, some shift, something that's worked well, even if it's something small, we want to hear about it at PhenomS X or in our Slack or post a link to it. And if you've shipped something that didn't work, we especially want to hear that. The failures are how the practice gets built. So quick wrap, PhenomS X is 38, no, 37 days out now. It is June 8th through 11th in San Diego. The day one key notes are going to orient roughly around PhenomS for AI token economics, AI value. The day two key notes are going to orient roughly around AI for PhenomS applying the magic of AI to the PhenomS practice. You can see the full agenda at x.phenomS.org. We have two new staff joining the PhenomS foundation announced on the call was Bo Nelford, who is joining us from AnglePoint. And he's going to be a focus technical community architect bridging the gap between the technical practitioners in the community and the focus specification and joining up the billing generators to help enable more focus adoption. We also have Enhiki, compost Amaran, who is joining us from crayon. He is currently running the Danish PhenomS community and is joining as our European community leader. Welcome to you both, and we're going to see them both in the conference. So if you can make it out there to the conference in San Diego here, stop by, see those folks, reach out to them now. Maybe set up time to say hi while you're there. I'm going to leave you with five takeaways from the summit that I want to summarize. The first one is that token economics has arrived as a core theme. Your job right now is really changing from a cloud cost or other SaaS management headspace to one of value maximization across the token factory, whether it be in cloud on prem or coming from a platform. Two, you really need to get the data foundation first. AI can do a lot of great stuff, but on top of messy data, it lights up every gap that you've been papering over with manual processes. Three, the AI part is actually the easy part that data we mentioned and getting engineering engaged with the right harness and the right actions and the right trust and automation on the back end is the more difficult part. So lead with concrete alternatives, not generic flags. Four, start with the boring automation, start small, build trust before you go big. I'm very fond of saying progress, not perfection. This is the best place to apply that. And five, nobody has this figured out yet. So show up in San Diego, ready to share, ready to build the practice out of the open. If you can't make it there, We'll have virtual keynotes dreamed. We're going to be sharing a lot online. If you're struggling to make the case or yourself funding, reach out to Joe Daley about options for discounted tickets. And that's the brief. I'm really looking forward to seeing many of you in 37 days. Let's go figure out this new world of token economics together. I'm JR Stormant. We'll see you soon.