Aventary · Insights
The Arguments Worth Having
Eight debates that decide how AI actually lands in the enterprise — and where I come down on each.
Most AI writing tells you what to believe. I'd rather show you the arguments and where I've planted my feet, because in this work the person you can trust isn't the one with the loudest conviction — it's the one who can fairly state the other side and still tell you why they chose.
So here are the eight arguments I've had to settle to do this job well. Each one is a real fight happening right now between serious people. I'll give you both sides straight, then tell you where I land and who moved me there. Nothing here is repeated on faith — every position traces to a real, dated source.
01 Replace the system of record — or embed around it?
The case for replace
Satya Nadella floated it on the BG2 podcast in late 2024 — the business-logic tier of SaaS "collapses" as agents reach into your data directly and orchestrate across it, making the old application layer optional. Bret Taylor, who once co-ran Salesforce and now builds agents at Sierra, argues the interface itself becomes the product and value moves to outcomes delivered on top of whatever's underneath. It's a serious, well-argued position from people who've built the systems in question.
The case for embed
Then reality filed its brief. Klarna tore out roughly 700 support agents, watched quality and retention slide, and rehired humans — and its own AI-maximalist CEO publicly doubted anyone was really going to rip out their system of record. The tell that settled it for me: Salesforce itself adopted the Model Context Protocol in 2025. When the system of record embraces the open standard rather than dying to it, "rip and replace" stops being the smart move.
Where I land
You embed AI around the record, not on its grave. The data, the governance, the audit trail — that gravity is real, and the winners work with it. I'd rather make your system of record smarter than bet your business on replacing it.
02 One model to rule them all — or orchestrate many?
The case for one
Simplicity. Pick the frontier model, standardize, don't fragment your stack chasing marginal gains.
The case for many
The research went the other way. Zaharia and the Berkeley team documented the shift from monolithic models to compound systems — the state of the art comes from components working together, not one giant brain. LMSYS's RouteLLM showed you can route easy work to cheap models and hard work to strong ones without losing quality. And in July 2026 Anthropic put the argument inside the model itself: Opus 5 ships with an "effort dial" and mid-task switching. The routing decision is now so fundamental it's a knob on the tool.
Where I land
No single model or tool wins, and loyalty to a vendor is a liability dressed up as a strategy. My job is to route the work to the right engine — and increasingly, to build systems that route themselves.
03 The interface is everything — or the interface is disappearing?
The case for the interface
Thirty years of software says the screen is where value and stickiness live. The UI is the product.
The case for disappearing
Karpathy's "Software 3.0" talk reframed the LLM as a new kind of operating system with English as the programming layer. Simon Smith wrote the cleanest piece on the interface dissolving into protocol. And it's shipping, not theorizing — ServiceNow went "headless" in 2026, opening the platform so agents do the work under the UI instead of a human clicking through it.
Where I land
The screen was never the point; the outcome was. I build for the world where the interface recedes and the job still gets done — but I don't pretend it happens overnight, because the humans using these systems are still human.
04 Trust the model — or verify it?
The case for trust
The demos are astonishing, the models keep improving, and verification adds cost and friction. Ship.
The case for verify
This is the argument I'll die on. Anthropic's Constitutional AI and the "LLM-as-a-judge" research showed machines can supervise machines at scale. OpenAI's own paper on why models hallucinate is the honest part — they guess confidently because our evaluations reward a confident guess over an honest "I don't know." Hamel Husain turned all of it into something you can actually operate with real evals. And the skeptics keep everyone grounded: Gary Marcus has been blunt that agents remain unreliable in exactly the compounding, multi-step ways that break production.
Where I land
If you deploy agents on someone's business without a verification loop, you're not an operator — you're an incident waiting for a date. AI has to check AI. It's why I source everything, including this article.
05 Turn it on — or change how the team works?
The case for turn it on
Buy the licenses, flip the switch, capture the productivity. The tool is the transformation.
The case for changing the work
The evidence is lopsided. Ethan Mollick has said it every way there is — individual gains don't automatically become organizational ones. The Harvard/BCG "jagged frontier" study measured it: big gains inside the frontier, and people getting confidently wrong outside it. McKinsey's framing is the one I quote most — a roughly 1:3:5 ratio of technology to process to capability-building. Most companies invert it, spend everything on the tool, and wonder why nothing moved. Klarna is the cautionary tale in the flesh.
Where I land
Turning AI on is the cheap, easy part. Changing how a team actually operates is the whole job — and it's the part I get hired for.
06 Build — or buy — in an era where building is cheap?
The case for build
It's never been cheaper. Retool's 2026 report found a third of enterprises have already replaced some SaaS with custom software. If you can build it in a weekend, why rent it forever?
The case for buy
Cheap to build is not cheap to own. Sutton's "Bitter Lesson" — the oldest idea on my list and still the most load-bearing — warns against betting on your own cleverness over general methods that ride compute. The clever custom thing you build today is the maintenance burden and the single point of failure tomorrow.
Where I land
The build-vs-buy line moved, but the judgment got more valuable, not less. Knowing what to build, what to buy, and what to leave completely alone is most of what I'm actually paid for.
07 Automate the people — or augment them?
The case for automate
Marc Benioff reframed headcount around "digital labor," and Salesforce's own support org shrank as agents took over conversations. The economic pull toward replacement is real and I won't pretend it isn't.
The case for augment
Erik Brynjolfsson's "Turing Trap" warned that chasing human-replacing AI concentrates wealth and power, while augmentation that complements people is the better road — for the business and the society. Mollick's "Choosing to Stay Human" sharpened it: AI as a shortcut erodes skill, AI as a tutor builds it. And Anthropic's own economic research found the heaviest users are often the most optimistic about their work, not the least.
Where I land
I'd rather build the kind of system that makes people better than the kind that makes them redundant — and I think, over a long enough horizon, that's also the better business. Being a good human as the roles change isn't a soft value bolted on the side. It's the strategy.
08 Is any of this even real — or is it a bubble?
The case for bubble
Ed Zitron's "Case Against Generative AI" argues it's circular money with no profitable models and revenue dwarfed by spend. A widely-cited MIT report claimed 95% of enterprise GenAI pilots deliver no measurable P&L impact. These are not cranks; they're describing something true about the froth.
The case for real
Anthropic's Economic Index shows genuine learning curves — people who've used these tools for months succeed more and route their highest-value work to the strongest models. The skill compounds. The adoption is real even where the hype is not.
Where I land
Both camps are partly right, and that's exactly the point. There's a bubble in the narratives and a real, compounding shift in the work. My job is to tell you which is which for your business — and I'll change my mind the moment the receipts change. The Klarna reversal changed mine. I'll tell you the next time it happens, too.
Who taught me this
I didn't reason my way to any of this alone, and pretending otherwise would be its own kind of lie — the exact kind this whole piece is arguing against.
Vernon Keenan's independent analysis of the Salesforce ecosystem taught me to read the platform's real direction underneath the marketing. Ian and Nate have challenged my thinking and kept me honest in the places I most wanted to cut a corner. The VCs I've talked to sharpened a different muscle than the builders did — harder questions, faster kills, the discipline of pattern over enthusiasm. And the wider community — the Salesforce ecosystem and the people building across every AI system I touch, the ones who answer a hard question late at night for nothing in return — is the reason any of this is possible. Community isn't a nice-to-have in this work. It's the substrate.
Before any of them, my Dad, who taught me what "doing it right" means in a way that has nothing to do with software and shows up in everything I build. And under all of it, my faith — the ground I actually stand on.
Why this matters if you're deciding whether to work with me
I won't tell you something I can't source. I'll give away credit freely, because I owe it. And I'll change my mind in front of you when the evidence turns, instead of defending a position past its expiration date.
That's not a pitch. It's how I show up. If that's the person you want in the room when you're betting real money on this, you know where to find me.
— Mendy, Aventary
Every idea referenced here is drawn from a real, dated, primary source. The full ledger — 70+ influences with links and dates — is right here.