So in other words: we are carrying more extinction risk from AI alone than most analysts assign to all other known threats combined across the whole 21st century? Ohhhkaaay.

The ppl selling you tokens told you that you should measure your productivity based on your token consumption…. and you believed them? NGMI

i.e., if your workflow has n sequential steps, and each step independently succeeds with probability p, then the probability that the whole workflow succeeds is p^n that means: a step that is right 95% of the time might sound great, but with 5 steps the workflow has a 77% chance of success, with 10 steps 60%...and so on. identifying where to make things deterministic vs LLM calls is going to be a big skillset of FDE, Applied AI Engineering, you don't want to just count on "smarter models", you need to push to make the per-step reliability as close to 1.0 as possible

Use coding agents to make your agentic workflows as deterministic as possible (in the right places, of course…) That’s it, that’s the tweet.

Extended thoughts on the buy vs build debate and the impact on SW in the AI era... I think the software companies most immediately vulnerable are infrastructure software companies and thin SaaS products that sit close to technical execution. Infra software is bought by technical teams to help build, deploy, secure, monitor, move, or operate technology AND the purchase maps closely to a technical capability AND the buyer is usually a technical person who can decide whether to keep buying, use OSS, build internally, consolidate into the hyperscaler, or have an MSP operate something more tailored. We have already seen this movie in cloud. Hyperscalers take open source projects, operationalize them, and sell them back as managed services. In AWS land: RDS, ElastiCache, EMR, OpenSearch, etc. That pattern will likely continue with MSPs starting to do this themselves, and AI makes it easier for more infra capabilities to be rebuilt, wrapped, automated, or operated locally. But business application SaaS is a different animal. SAP, Salesforce, Workday, ServiceNow, Epic, etc. are not primarily bought because an enterprise cannot technically build it, they are bought because the business wants a trusted system to operate an important process: finance, HR, revenue, support, procurement, clinical operations, supply chain and they want to transfer that risk (operational, compliance, etc) completely to the SW vendor. They are generally business operating model decisions not pure SW decisions. The budget owner is often the business, the risk owner is often the business, and when something goes wrong the buyer wants institutional trust and vendor accountability. They want to be able to tell the board, “We picked Salesforce like every other top company,” not “we vibe-coded our own CRM workflow and Bob from IT is maintaining it.” I saw this at AWS: Customers would often move SAP last, not because SAP was technically impossible to move, but because the business risk was so high that the board was often involved. The cost of writing lines of code may trend toward zero, but the cost of operating software is still very real. Uptime, security, compliance, permissions, integrations, data quality, support, change management, upgrades, audits, and accountability do not go to zero. And the typical enterprise still does not have the structure, talent model, or operating discipline to run software at scale like a software company. Maybe that changes over time, but I do not see it changing quickly. Cloud was already a massive opportunity for enterprises to modernize how they thought about software, and most did not become cloud-native software companies. They moved what made sense to cloud, kept some things on-prem, and went SaaS for the rest. I think AI follows a similar pattern: it compresses the cost of building around the edges, but that is different from saying the trusted system of record goes away. cc: @glbeaty

Looking forward to trying this out. I have the perfect test for it in a Healthcare project.

This is called outcome based account management (OBAM) and IMO it’s the only way to do strategic selling. AWS introduced it to its salesforce in ~2015 and it was a key unlock for us to get a seat at the table in those conversations. I’m super passionate about this style of selling.

Yes the “consulting” industry is going to have to change, but one thing I’m 100% confident on is there is no AI w/o professional services. It’s just that the services will look, feel and be priced like software.

“…. while you get to focus on the big picture and critical thinking, which are the more fun parts of the role anyway” Yep. I often frame it like when selling to folks who are reticent on AI, i.e. most clinicians: “what if you were able to just do the parts of the job you love” and it clicks for them immediately.

I’ve been spending a lot of time thinking about & discussing the correct implementation of “AI “in Enterprises. The decision to build an app or build a workflow with AI is going to be something every enterprise faces in perpetuity. Long form thoughts below….. cc: @matt_slotnick

Great insights from Matt as always. The enterprise diffusion of AI will be interesting to watch. My POV is that we will see the same thing happen that did in cloud: 1- Enterprises will “do AI” like they “did cloud” early on.. ie board will ask for a cloud strategy so the customer consumes cloud even if the workload isn’t a fit 2- cloud sold as a panacea 3- Then questions came on why the bill is so high, what value are we getting 4- customers then level up on understanding cloud first principles, architecting for the cloud, why static workloads need care, org re-design, new skills needed 5- new mental models established to understand how to extract real value There is a parallel to the above with inference spend that is taking place right now in the early stages (1-2) of the above. By the end of this year the inference spend vs value delivered questions will be in full swing (3). Meanwhile enterprises that leaned in early will be iterating on 4-5 in a loop and be far ahead.