Friday, August 28, 2026
FRIDAY – AI FOR THE C SUITE®
Read time: 9-10 min · Read online
Hi, it’s Chad. Every Friday, I serve as your AI guide to help you navigate a rapidly evolving landscape, discern signals from noise and transform cutting-edge insights into practical leadership wisdom. Here’s what you need to know:
1. Sound Waves: Podcast Highlights
This past Monday on AI for the C Suite®, I was joined by Dan Fernandez, founding product manager at Adara and former product lead at Chainguard and CrowdStrike. We get into something that should make every CFO sit up: the same task can cost you $100 or 85 cents depending on which model it gets pointed at… and most companies have no idea which one they’re paying for. Tune in wherever you get your pod on.
Apple · Spotify · iHeart · Amazon · YouTube
Subscribe for free today on your listening platform of choice to ensure you never miss a beat.
2. Algorithmic Musings: Salesforce is Dead. Long Live… Salesforce?!?
I read a press release twice on Wednesday night, which is not how I planned to spend those 5 minutes.
Salesforce and Anthropic announced an expanded partnership called Claudeforce. (Press release here. Trade summary here.) Most of the coverage framed it as two large companies getting larger together. What I saw was one of the biggest software companies on the planet conceding, in writing, the very thing it spent twenty-seven years selling you.
Usual caveat when I write forward: I’m reading a trajectory, not reporting a settled fact. Judge the argument on its merits and give the timing a wide margin.
What Salesforce Conceded.
Benioff’s line in the release is “the UI is the AI.” The supporting language is blunter than the quote. For decades, per Salesforce’s own framing, enterprise software required people to navigate static screens to get work done. Now agents reach the data, workflows, and rules directly, and the software becomes the system underneath every interface.
Read that as a mid-market buyer rather than an investor. Salesforce is not announcing its own obsolescence. It is announcing which half of itself it intends to defend… and the half where they believe value will remain.
The screens are being given away. The data model, the business rules, the permissions, the governance, the enforcement layer that decides whether an action is allowed: those are being fortified. Every action Claude takes routes back through Salesforce so the business rules still fire. Benioff frames it as reasoning plus determinism, each useless without the other.
Which is where the ERPs-are-dead genre has been getting it half right for two years. When I wrote about The SaaS Overhang in May, the argument was that the interface layer was contested territory and the system of record was durable. I did not expect a vendor to confirm both halves in a press release inside of four months.
The Part That Should Make You Uncomfortable.
Everyone hated those screens. Nobody thought about what they were doing.
For twenty years, your CRM and your ERP interfaces were the only surviving explanation of how your company works. Not the documentation, which nobody wrote. The screens. Why there are two overhead pools. Why that opportunity stage requires an approval. Why the field somebody added in 2013 still gets filled in every Tuesday. None of it was written down, but all of it was visible, and people learned it by clicking through it.
Which brings me to Dennis Nedry.
Jurassic Park has a scene Gen X-ers everywhere remember, where Lex sits down at a terminal, recognizes the file system, and gets the park’s doors back online. The line nobody quotes comes much earlier, when Nedry is still in the room complaining about his pay. He boasts that nobody else could have networked the machines and debugged two million lines of code for what he bid the job. He is bragging about the exact fact that dooms the park. The sabotage is beside the point. The park was undone by the fact that its operating logic lived in exactly one head and had never been made legible to anyone else.
Your configuration is your two million lines. The interface was the compiled version, the part your people could read. Salesforce just told you the compiled version is going away and the source is being handed to an agent instead.
So What Do You Do About It.
Three moves, in order, and none of them require you to make a platform decision this quarter.
Stop treating data hygiene as an IT chore. I know you are tired of hearing it. Hear it once more anyway, because the terms just changed. When people worked the screens, sloppy naming and thin metadata cost you some frustration and some rework. When agents work the rules directly, that same sloppiness becomes the thing your automated decisions are made of. Naming conventions, file structure, metadata discipline, and a database that can answer a question without three people interpreting it are no longer hygiene. They are the asset.
Reframe buy versus build as a question about layers. There’s a case for building that has very little to do with software, and it’s the case I’d make to you. Most middle market companies cannot describe their own processes out loud with any precision. Buying lets that continue, because the vendor’s template quietly supplies every answer you never worked out for yourself, and you inherit somebody else’s assumptions about how your business runs. Building refuses to let you skip the step. You cannot configure a rule you can’t articulate. Which means the build becomes the process mapping exercise you’ve been putting off since 2019, except this time you can’t abandon it halfway, because nothing works until it’s finished. You are going to do that work either way. Building is the version where quitting isn’t an option.
Two conditions on that. Your general ledger is a poor candidate, because auditors, tax authorities, and your lender all hold opinions about it that a custom build satisfies slowly and expensively. And whatever you do build, understand that you have just appointed someone the custodian of your own two million lines. Decide who that is on the way in rather than finding out on the way out. Fair disclosure: mapping processes and defining rules is work I sell, and I run a platform built on infrastructure that appears in this announcement. Weigh both when you weigh the advice.
Ask what you are willing to rent. If more of your cost base shifts from labor toward compute, and I think that shift is underway even if I can’t yet put a defensible number on it, then the rent-or-own question stops being an IT budget line and becomes a strategy question. Renting a frontier model is fast, cheap to start, and leaves you exposed to somebody else’s pricing and roadmap. Owning your own containerized stack costs more up front and buys you control. Neither answer is wrong. Not having asked the question, though, is.
The Thought I Keep Circling.
If the screens are now a commodity, then the most valuable thing Salesforce owns is not the application. It is being the most rock-solid, best-governed, most trustworthy back end in the business. That is a smaller story than the one in the press release and a far more durable one. It also happens to be the same job description your ERP has been quietly holding down all along.
My takeaway? The vendors have started announcing the disappearance of the interface as a feature, and they are not wrong that it is one. Before you celebrate, go pick three configuration choices in your core system and ask who can explain why they were set that way. Not what they do. Why. If the only honest answer is the implementation partner you used in 2019, you’ve uncovered your Nedry problem, and it doesn’t require anyone to sabotage anything.
Hit reply and tell me what you find, or reach me at chad@chadharvey.com. And if this one has you wanting to go back to first principles, Monday’s episode of AI for the C Suite® podcast ep is a back-to-basics session built for exactly that.
3. Research Roundup: What the Data Tells Us
Beyond Benchmark Scores: Why the Task You Pick Matters More Than the Vendor You Select.
On the reg, someone asks me which model they should standardize on or which model is “best.” New research out of Cambridge and the UK’s science ministry suggests that’s the smaller question.
The numbers that matter: Researchers profiled six commercial AI systems across eight cognitive capabilities, then surveyed 400+ employees in six industries about what their actual work demands. The same system ranked first on 17 of 18 work activities tested. And the leading model’s edge came almost entirely from planning, roughly double the next system’s, while its knowledge edge was slight.
What this means for your Monday morning: Switching vendors rarely changes which work comes into scope. Research, administrative work, checking, and data analysis line up well with what these systems can do. Problem solving, decision making, supervising people, and hands-on tool use don’t, and every system fails in the same places.
The catch: Planning is where the leader pulls ahead of the pack, and it is still not good. Planning and causal reasoning are the two places every system tested came up short against what the work actually requires, and they are exactly what agentic products lean on hardest. A stronger model won’t close that gap because the weakness is structural, not a matter of degree.
Action item: Stop scoping AI by job title. Pick one department, break the week into activities, and pilot where the match is strongest. Ask any agentic vendor what scaffolding they’ve built to cover planning, because that’s where the real work is happening.
Read our full analysis of this research at AI for the C Suite®.
4. Radar Hits: What’s Worth Your Attention
Cheaper AI is widening the gap, not closing it. Token prices fell 40% since late June. July AI spend per employee still rose 49% among the top 1% of firms and 9% at the median. The spread between heavy users and everyone else has doubled since 2023. Budget AI as a growing line item, not a shrinking one, and assume your fastest competitors are spending more, not less.
Thomson Reuters built its own legal AI model for $450,000. The full program ran $40 million over two years, but the final training run cost under half a million because they started with an open-weight model and added their own content. If you own a deep proprietary archive, domain-specific models are now a procurement question, not a moonshot. Ask what your data is actually worth.
Claude’s memory now follows you across apps, on by default. Anthropic merged memory between its chat and Cowork products for consumer plans, and it writes to memory mid-conversation now. Team and Enterprise seats stay off until an admin turns them on. That split is the problem. Anyone on a personal Pro account doing company work has retention running outside your controls. Worth an audit this week.
5. Elevate Your Leadership with AI for the C Suite®
If reading this made you realize you can’t name who owns your configuration logic, that’s the problem worth solving before the platform decision arrives. I’m taking Q1 2027 engagements with mid-market leadership teams. chad@chadharvey.com.
Forward AI for the C Suite® to a peer who’s still treating data hygiene as an IT chore.
Stay safe. Stay healthy. Be strong. Lead well.
Chad