5 Levers That Decide Your AI Investment's Fate

Here's something I keep running into when I cover agentic AI from the security side: the organizations treating these deployments as pure capital allocation problems — ROI calculations, vendor bake-offs, build-versus-buy spreadsheets — are largely skipping the conversation that matters most to me. An agentic workflow touching your accounts receivable isn't just a productivity question. It's a system with access to customer financial data, the authority to act on it, and in many cases, nobody who's clearly accountable for what it does when it hits an edge case.

That's the backdrop I brought to a recent video from Nate B. Jones, who runs AI News & Strategy Daily. Jones isn't a security person — he's an operator and strategist — but his framework for thinking about AI investment decisions is the clearest I've seen on the business side. And it exposes, inadvertently, exactly where the risk surface gets ignored.

Jones's framing starts with a number that's been circulating: Gartner has reportedly predicted that more than 40% of agentic AI projects will be abandoned by the end of 2027. [Editor's note: Buzzrag has requested the specific Gartner report citation; Jones attributes the failure causes to cost overruns, unclear business value, and inadequate risk controls, but readers should treat the specific failure breakdown as Jones's interpretation pending source verification.] Jones's point is that these aren't failures of the technology — they're failures of investment logic. And the enterprise AI failure pattern here is depressingly consistent regardless of which framework you use to describe it.

The workflow is the unit of analysis

Jones's core argument is deceptively simple: you don't have an AI problem. You have a collection of workflows, each with a different shape, and each warranting a different investment decision.

An accounts receivable team, for instance, isn't dealing with one AI use case — it's managing collections prioritization, invoice matching, customer follow-up, exception handling, cash application, dispute resolution, reporting, and escalation. Eight different shapes of work. Possibly eight different investment calls. "If you pile all of them into a single RFP," Jones says, "you're going to get a mediocre tool that does maybe one of them well."

His definition of "workflow" is more rigorous than the word usually implies: "the entire operating loop — what information comes in, what the system is allowed to do, and what good output looks like, who's checking what, what gets escalated, who owns and is accountable for what the result is." The AI model, he argues, is a small component of that loop. The workflow is what you're actually investing in.

From there, Jones lays out five levers: automate, build, buy, hire, or wait. Often, the answer is a combination.

Automate: the easy call that isn't always easy

Automation is the right move when work repeats often, follows a clear pattern, has exceptions you can define in advance, and can be quality-checked cheaply. Jones cites IBM's AskHR system as an example. [Editor's note: IBM has undergone significant HR technology changes; the current state of this system should be verified independently.]

Jones's warning here is the one I'd tattoo on the wall of every enterprise AI demo room: "Don't automate when the exception is where most of the value is." The vendor demo shows you the routine case. The contract gets signed. Production traffic turns out to be 60% exceptions. Someone in the C-suite is now staring at an accuracy number they don't understand, convinced they were lied to. Nobody lied. The buyer just didn't map the workflow before they signed.

From where I sit, there's an additional layer here that Jones doesn't address. Automated workflows, especially ones making consequential decisions, raise serious questions about what data they're operating on and who has access to it. A collections prioritization agent that ingests customer payment history and behavioral signals is, by definition, a high-value data target. Before you decide whether to automate that workflow, you need to know: what data does this agent touch? Who can query its logs? What happens when it makes a wrong call on a protected-class customer and you need to reconstruct the decision chain? These aren't hypotheticals — they're the questions your legal team will ask after the fact if you don't ask them first.

Build: you're the only honest quality reviewer

Build is the right call when a workflow is too specific, exception-heavy, or competitively sensitive to hand off to a vendor. Company-specific data, approval gates, risk thresholds, institutional knowledge — these are the ingredients that make a workflow yours. Jones frames this as an agentic loop potentially involving multiple tools, sub-agents, MCP connectors, and vendor components stitched together into something custom.

The thing Jones hammers on — and he's right — is that executives commissioning builds frequently can't define what success looks like. "Your team is going to come back to you and they are going to be incentivized to tell you, 'Yep, this is good. Yep, we built the AI thing the executive wanted.'" Without independent criteria for quality, you're just hoping. That's not a governance posture — that's a liability waiting to land.

Buy: the 80% overlap test

Jones separates buy decisions into two categories: primitives — basic components or services that development teams can stack into multiple workflows — and whole-workflow vendors like Harvey, the legal AI platform, which essentially sells you an end-to-end agentic pipeline. [Editor's note: Harvey's current product positioning should be confirmed; this market is shifting rapidly.] Jones also references Stripe as a source of developer tooling that can serve as building blocks, though the "agentic primitives" framing appears to be Jones's characterization rather than Stripe's own product positioning.

The test Jones proposes for whole-workflow vendors is blunt: does the vendor's workflow shape overlap with yours by at least 80 to 90 percent? If not, "you're going to do a lot more work than you think adjusting it. And it's more complicated in the age of AI than it was in the age of deterministic software."

This is where I want to add something Jones leaves out of the buy conversation entirely, and it matters: when you purchase a workflow solution, you're also purchasing its data handling practices, its access model, and its blast radius if something goes wrong. A vendor selling you an agentic legal workflow isn't just selling you task automation — they're getting access to privileged communications, matter strategy, client data. The contract you sign for the software is not the same as a data processing agreement that reflects what the agent actually does. Before any whole-workflow vendor purchase, the questions your security team should be in the room for include: What data does the agent retain? Where does it go? Who at the vendor can see it? What's the incident response plan if the vendor is breached? These aren't bonus questions. They're table stakes.

Intercom's Fin — the customer support agent Jones references as a buy-plus-automate example — is a product specifically designed for high-volume, repeatable support cases. It's a reasonable illustration of the concept, but customer support agents also ingest sensitive customer data at scale. The privacy surface on those deployments is not small.

Hire: specificity over unicorn-hunting

Jones is candid about the talent market: it's broken. Companies are chasing what he calls the "purple unicorn" — a domain expert who's simultaneously an AI builder, systems architect, change leader, and executive communicator. "Sometimes that person exists," Jones says. "More often than not, the market is going to clear out a lot of AI talent from under you while you figure out what you actually want."

His prescription is workflow-anchored: define the specific capability gap your target workflows actually need in six to twelve months, then hire for that gap — not for an impossible composite. If someone on your existing team can close that gap with training, keep them and invest in their development. The agentic AI hiring fog Jones describes — job descriptions no one can define, candidates no one can verify — is a solvable problem if you work backward from the workflow.

Wait: the deliberate choice

Jones's fifth lever is the one that requires the most organizational courage right now. Waiting isn't paralysis — it's the recognition that not every workflow needs AI now, and that sequencing matters. Apply AI first where you get the most leverage. Lower-priority workflows can wait for better tooling, clearer regulation, or a more mature vendor market.

I'll add: for privacy-sensitive workflows specifically, "wait" is sometimes the responsible answer regardless of where a workflow falls on the priority list. If you can't answer basic questions about data residency, access controls, and audit trails for an agentic deployment, that workflow isn't ready — and neither are you.

Jones closes his framework with a line that should be the first item on every pre-deployment checklist: "Do not automate what you cannot describe."

I'd update that for a security-conscious audience: don't deploy what you can't audit, can't explain to a regulator, and can't recover from when it goes sideways. Agentic AI is not software in the traditional sense — it's systems with agency, operating on sensitive data, at the boundary between your organization and the people it serves. The business case and the risk case belong in the same conversation.

Before your next vendor demo, ask one question your CFO probably isn't asking: if this agent makes a consequential mistake with real customer data, who's accountable, and can you prove what happened?

If nobody in the room can answer that, you're not ready to buy. You might not even be ready to build.

---

Rachel "Rach" Kovacs is Buzzrag's cybersecurity and privacy correspondent.