How executives should be thinking about AI (Aug 2026 version)

Every vendor I talk to is convinced they need AI in their tool. A bunch of them are also shipping MCP servers as something more than a data access point, which is the far more interesting move and gets about a tenth of the airtime. Meanwhile my clients are getting hammered by AI sales telling them this agentic thing or that agentic thing WILL 10X OUTPUT. It's Agentic so BUYBUYBUY. Their C-suite is scared shitless they're behind the curve.

So they go buy something. Usually an agent pointed at a problem a script already solved.

Discrete problems want discrete tools. Invoice matching, tax calculation, inventory reconciliation, routing a ticket to the right queue. One right answer, a rule that spits it out every time, and a system of record that has to prove it later. Hand that to a model and you get the same answer slower, at higher cost, and occasionally wrong in a way nobody catches for a quarter.

Where AI earns its keep is the fuzzy end. What's going on with this category. What am I missing before this call. Read six months of this godawful email thread and tell me what actually happened. Help me think through the thing I can't articulate yet. It's a thinking partner and a synthesis engine, and that covers a hell of a lot of the work in most companies. Work that has never had a tool at all, so nobody ever tried to buy one for it.

Your buying decisions feel chaotic because the sales motion doesn't distinguish between those two, and most of your teams can't tell the difference yet either.

Here's my current thinking. I hope it helps.

AI is the interaction layer

I used to say think of Claude the way you think about Excel. It's the base tool work gets done in.

People fight me on that one, so let me make the actual argument. Excel isn't in your stack because somebody evaluated it for a use case. It's the default workbench for white collar work. Nobody gets trained on it, everybody opens it, and it's where the thinking happens before it turns into a deliverable in some other system.

This feels like that. AI tools are becoming where work gets done whether you signed off on it or not. The only real question is whether your workbench got picked on purpose or just showed up one Tuesday.

It comes in WAVES

Every function gets there eventually. They don't get there at the same time, and the order has nothing to do with which team is most fired up about it. It's about what each wave costs to enter and how much risk is sitting on the other side.

Wave one: BI & Data

Execs will still want their static scorecards. But tier two reporting eventually lives in an AI interface. Data teams burn absurd hours building dashboards the business won't open, for use cases three people asked about once. AI eats that.

Entry fee is the one nobody wants to pay: a real data model and a semantic layer. Miss this and you will get different or wrong answers randomly, NOT A GOOD TIME. Pay it and the BI developer's job moves down the stack into defining, training, and validating, and that job gets a lot more valuable. Wave one goes first because it's the prerequisite for most of what comes after.

Wave two: IT and PM.

There are plenty of tools out there now to manage help desk functions. Smart IT peeps are already thinking "don't buy that, an agent with access to our knowledge base and some n8n flows fits better, and we can control it damnit." If an IT person says that to you, give them a raise.

For PM functions it's even more real. My CoS agent pokes and prods me, manages my Linear backlog, and runs my rhythm of business check ins. It completely fails to send me "update this status slide pls" emails. Not sad about it.

Entry fee: a knowledge base worth pointing at, and permissions you actually trust.

Wave three: The revenue-connected functions. Buying, planning, marketing, sales.

Once I can structure and pipe your data somewhere consumable, my focus shifts to process guardrails for these teams. I do sales today, or I try to. My agents chew through my email, meeting notes, and public sources and prep me for calls. I cannot find the pen on my own desk, but Hank makes sure I know the last time I talked to Tony was three weeks ago, what we covered, and the one simple thing he couldn't get working. Does that sound vaguely like CRM tooling to you?

Entry fee: piped, structured data plus guardrails on the process itself, because this is the first wave where the output touches a customer.

Final boss: Accounting, HR, Compliance

These teams are buying bespoke tools from vendors and honestly that's the right call for now.

Which blows a hole in most of what I just told you, so let me own it. Everywhere else I'm saying the tool becomes a backend and the interaction moves up a layer. Not here. When the cost of being wrong is a restatement, a lawsuit, or a regulator with a subpoena, you want a vendor whose name goes on the mistake. Risk allocation is a real feature and you're allowed to pay for it. These teams will use the enterprise tool to summarize email and draft memos, but the actual work stays in their systems until governance catches up.

What this means when you're buying

This is where I watch the money get set on fire, so it gets its own section.

• Be less impressed by the AI features in the demo. The demo is the part your team stops using in month two. Ask what the integration surface looks like instead. Does it have an API worth the name. Does it have an MCP server. Can data get out as easily as it goes in.
• Stop asking "does this have AI in it." Start asking "can my AI reach it." A boring tool with a clean API is worth more to you than a shiny one with a chat box bolted on and a locked box behind it.
• Shorten your contract terms. If a vendor is on its way to becoming a backend, why the hell are you signing three years. I have clients negotiating six to eight months on tools they fully expect to rip out once the platform underneath is ready, and vendors are taking that deal more often than you'd think.
• Think about what you're actually onboarding people to. A lot of the workflow-change pain disappears if you never train the team on the vendor's interface in the first place.

TLDR

• Figure out an enterprise AI platform. Anthropic, OpenAI, Google, something you build. PICK ONE.
• Build the infrastructure and the talent to run it. You need engineers. You need data people. Someone who actually owns governance would be cool too.
• Put someone security and risk minded in the room. They're the one who says "dude, somebody is going to upload a CSV of social security numbers to whatever model got posted on X this morning, let's not."
• Write an AI policy people get behind. Not "thou shalt." More "here's how, responsibly."
• And pay the entry fee for wave one before you go shopping for wave three. That's most of it.