Every quarter has a magic word, and this quarter it's "agentic" - stapled to the front of software you already own. North Wayne isn't impressed, and she has Gartner's own language to back her up: the analysts call it "agent washing." Her column today isn't a takedown of AI; it's a buyer's framework for telling a bet you'll still be glad you made in two years from one you'll quietly cancel. If you sign off on AI budget, read this before the next demo. The one question she wants in your head: moat, or feature? — Muximus
Every vendor deck this quarter has the same word bolted to the front of a product you already know: agentic CRM, agentic cybersecurity, agentic BI. It feels like the future. It is, mostly, the crowded trade. And if there is money attached to your next AI decision, the single most useful skill you can build this year is telling a durable bet from a reskin - because you are about to be pitched a great deal of the second kind and asked to fund it like the first.
So let me give you the framework, then the trade-off, then a clear call.
The tell
Start before the demo. When the pitch is a two-year-old product with "agentic" appended - the meeting note-taker that now introduces itself as an agent, the dashboard that "agentically" surfaces what last year it simply displayed - ask one question: what actually differentiates this? Nine times out of ten the honest answer is the model underneath. And the model is a rented input, commoditizing fast, that your competitor can rent next quarter at the same price.
That makes it a feature, not a moat. I want to be precise here, because the distinction is the whole decision. A feature is worth buying when it pays for itself this year. It is not worth building a strategy around, because the thing that makes it special isn't yours - it's the vendor's, or the model provider's, and it's for rent to everyone including the people you compete with.
You do not have to take my word for the risk. Gartner projects that over 40% of agentic AI projects will be canceled by the end of 2027, citing escalating costs, unclear business value, and inadequate risk controls. The same analysts named the pattern behind it - "agent washing," the rebranding of ordinary assistants, RPA scripts, and chatbots as agents without the capability to match - and reckoned only about 130 of the thousands of self-described agentic vendors were the real article. Consider the source: Gartner sells forecasts, so read the number as a projection, not a measurement. But I've watched this failure mode in the room more than a dozen times, and its name is always the same - a project that demos beautifully and adds cost, with no business case that survives the second invoice.
The one question: moat or feature
Here is what I want in your head before any vendor opens a laptop. Does this AI create a durable advantage that compounds and is hard to copy - or does it add a capability anyone can rent next quarter?
Durable advantage comes from what you own and accumulate: proprietary data, deep context about how your customers and your operation actually work, workflow lock-in that gets stickier the more the thing is used. A rented feature comes from what everyone can buy: the base model, the obvious wrapper, the label. The first is a bet - it earns budget, patience, and a seat at the strategy table. The second is a purchase - treat it like one.
Where the moats plausibly sit
If the reskins are the crowded trade, the durable bets are one layer down, where AI builds lock-in instead of a demo. Four directions worth positioning around - not four things to buy Monday, but places to watch and build toward.
Memory and personal context: AI that genuinely knows a person or an organization creates switching costs the way a well-kept customer file always has, only deeper - and the moat is the accumulated context, not the model reading it. The agent economy: the move from one agent doing one task to agents finding, hiring, and coordinating each other, where the value is the coordination layer, not another single-purpose assistant. Verification and trust: when anyone can generate anything, the scarce good becomes confidence that the output is right, and whoever makes that cheap and credible sells the thing everyone else now depends on. And anything physical: robotics and the unglamorous plumbing that turns a generated plan into a machine that moves.
None of these is a purchase order this quarter. Several are early enough that the right move for most mid-market companies is to position, not chase - start accumulating the data and context now, so you're ready when the tooling matures, instead of funding a moonshot you can't staff. "Wait" is not the timid answer here. It is frequently the correct one.
A disclosure, because this column doesn't get to skip it: VarOps, which publishes this magazine, sells AI transformation work and builds its own memory-and-context and agent-coordination tooling - which sits squarely next to the first two directions I just praised. That alignment is a reason to trust this framing less, not more. Hold it to the same question I'm handing you: moat, or feature? The test doesn't care who profits from the answer.
The Verdict
Don't fund the same product with an agent glued on as a strategic bet. Buy it as a feature if it pays for itself this year, and go in expecting to cancel it if it doesn't - you'll have plenty of company by 2027. Put the strategy budget where the AI compounds into something you own: your data, your context, your workflows. And when the honest read is that a direction is real but not yet ready, position and wait - that's judgment, not timidity. The leaders who lose money on AI over the next two years mostly won't lose it on a bad model. They'll lose it on a confident purchase any competitor could have made, for the same price, a quarter later.
