Playbook
Tasteful Skills
“Tasteful Skills” argues that the best agent skills are not documentation or best-practice lists.
Back to Podcast Digest
5 Levers That Separate Winning AI Investments from Disasters
TL;DR
The real unit of AI investment is the workflow, not the model or department — Nate B Jones argues that teams fail when they issue one broad AI mandate for something like accounts receivable instead of separating collections prioritization, invoice matching, dispute resolution, and cash application into distinct decisions.
Gartner’s “40% of agentic AI projects will be killed by 2027” warning is mostly about bad capital allocation, not bad agent tech — he says the usual culprits are predictable: cost, unclear business value, and weak risk controls, often driven by buyers who let vendors define the problem for them.
There are only five real levers: automate, build, buy, hire, or wait — the smart move depends on repeat frequency, error cost, judgment required, company-specific context, market maturity, and whether the next model release is likely to commoditize the task.
His sharpest rule is: ‘Do not automate what you cannot describe’ — if a team can’t explain a workflow’s inputs, outputs, standards, exceptions, and ownership in plain English, they’re not ready to buy, build, or hire around it.
Most flashy demos break on exceptions, not routine cases — he says vendors wow buyers with the happy path, but production traffic is often edge cases, which is why tools like IBM AskHR or Intercom’s Fin only make sense where routine dominates and exceptions are cheap to handle.
The hiring market is broken because companies want a ‘purple unicorn’ instead of a workflow-specific hire — rather than chase a domain-expert-AI-builder-systems-architect-change-leader combo, he recommends hiring for the exact missing capability, like evaluation design or workflow engineering, and leveling up existing staff where possible.
The Breakdown
The Gartner extinction headline, and why he says people are reading it wrong
Nate opens with the Cretaceous-extinction joke and Gartner’s prediction that more than 40% of agentic AI projects will be killed by 2027. His point is blunt: this isn’t proof agentic AI is broken; it’s proof companies are investing badly — overspending, chasing vague value, and skipping risk controls.
Shut the vendors out for a minute and look at the work itself
He tells the story of a finance leader whose CFO wanted AI for order-to-cash, while three vendors pitched three totally different solutions without describing the actual work. That’s the pattern he keeps seeing: businesses know they need help, vendors confidently fill in the blanks, and nobody stops to define the workflow before buying.
“AI investment is not an AI question”
This is the central turn in the video: AI strategy starts with the shape of the work, not the model, dashboard, or vendor. He makes it concrete with accounts receivable and product teams, rattling off sub-workflows like collections prioritization, invoice matching, spec drafting, backlog grooming, and launch coordination to show why one giant RFP usually produces a mediocre, mismatched tool.
What a workflow actually is, and the five levers you can pull
He defines workflow as the full operating loop: what comes in, what the system is allowed to do, what good output looks like, who checks it, what gets escalated, and who owns the result. From there, he says every decision reduces to five levers — automate, build, buy, hire, or wait — and the right choice depends on repetition, judgment, specificity, market options, and whether the category is about to be eaten by the next model release.
Automation works great — until the exceptions are where the value lives
Automation is the easiest lever when work is repetitive, patterned, and cheap to verify, which is why he points to IBM AskHR and Intercom’s Fin as useful examples. But he warns that many enterprise AI disasters start when a demo shows the clean routine case while the real business runs on messy exceptions, leaving executives staring at ugly accuracy numbers and wondering what went wrong.
Building is for the company-specific “secret sauce,” but only if you can define good
Build makes sense when the workflow is full of edge cases, approval gates, internal standards, and company-specific context that an off-the-shelf tool won’t capture. He gets especially animated here: too many execs tell teams to “go build the AI thing” without being able to say what success actually looks like, which practically guarantees confusion, false confidence, and bad output.
Buying and hiring both fail when companies stay vague
On buying, he distinguishes between useful primitives you can stack into many workflows and full packaged systems like Harvey for legal, where the real question is whether your work is truly “Harvey-shaped” with 80–90% overlap. On hiring, he says companies keep chasing a “purple unicorn” while drowning in AI-generated resumes and market noise, when they should instead define the exact missing capability tied to the workflow and often train internal people if they can level up within six months.
Waiting is sometimes the smartest AI move, and he closes with the big rule
He defends waiting as disciplined prioritization, not anti-AI hesitation: if your SQL query layer already works, maybe don’t rebuild it first; put AI where it creates more upstream leverage, like analytics interpretation and storytelling. He lands the video on the line he wants in every investment review — “Do not automate what you cannot describe” — and says the executive job now is better capital allocation around workflows, not a vague top-down order to “have an AI strategy.”
Share
Keep Reading
Two Ways to Follow
The Weekly Echo. The inbox-shaped summary of what mattered.
New editorials announced here.
