Date: 2026_05_17 Source: https://www.youtube.com/watch?v=LIkYVsxMpS8 Duration: 1666 Platform: YouTube Creator: AI News & Strategy Daily | Nate B Jones

When to Automate, Build, Buy, Hire, or Wait on AI¶

Overview¶

Gartner projects >40% of Agentic AI projects will be killed by end of 2027 — not because the tech fails, but because of cost, unclear business value, and inadequate risk controls. This video provides a decision framework for investing in AI workflows. The core argument: AI investment is not an AI question — it's a question about the shape of your work. The model and vendor questions are downstream of that. Before talking to vendors, you need to understand the work itself.

The Core Framework: Work First, AI Second¶

The Fundamental Mistake¶

Most teams skip the work-analysis step — especially when vendors encourage them to. Instead of asking "What does this work need?" they ask "Which AI tool should we buy?" The result: expensive implementations that don't match how the work actually happens.

The real decision hierarchy: 1. Shape of the work → what it needs 2. Model choice → downstream of that 3. Vendor choice → downstream of that 4. Dashboard / reporting → downstream of that

The Anatomy of a Workflow¶

When Nate says "workflow," he means the entire operating loop, not a prompt: - What information comes in - What the system is allowed to do - What good output looks like - Who checks what - What gets escalated - Who owns and is accountable for the result

The AI model is "a tiny tiny part of that loop." It makes the whole thing go, but it's not the only thing. The workflow is what you are actually investing in — and it's what gives you leverage if you do it better.

Why You Can't Pile All Work Into One RFP¶

A finance leader told Nate her CFO wanted to "do AI in orders to cash" and got three vendor quotes for three different shapes of solution — none of which described the actual work she was doing. This is the default state of AI procurement.

The accounts receivable team doesn't have one AI problem — it has 6-8 distinct shapes of work: - Collections prioritization - Invoice matching - Customer followup - Exception handling - Cash application - Dispute resolution - Reporting - Escalation

These are all very different shapes of work that route to very different investments (automate some, build some, buy some). If you pile them all into a single RFP, you get a mediocre tool that does maybe one well and doesn't adequately cover what you actually need.

The same is true in product: User research synthesis, spec drafting, backlog grooming, design review, experiment analysis, roadmap judgment, launch coordination, customer escalation — all different shapes of work requiring different decisions.

The unit of decision for AI is not a department head or a role — it's the specific work you're naming.

The Five Levers¶

Every workflow can be evaluated against these five options (often in combination):

The critical insight: Often the right answer is a mixture of multiple levers. You might buy one piece that sits inside a loop you build. You might automate part of a workflow and build the rest. You have to think about moving multiple levers at once.

Automate: The Easiest Call — With a Caveat¶

Automation is the right call when: - Work repeats often - Follows a clear pattern - Has recognizable exceptions you can define - You can check quality cheaply

The bad enterprise AI demo failure mode: The vendor shows you the routine case — it looks impressive — the buyer signs the contract. But the buyer's production traffic is mostly exceptions, not routine cases. The executive team ends up staring at a very low accuracy number wondering why they were lied to. Nobody was lied to. The buyer just bought the wrong thing.

The other automation failure: Don't automate when the exception is where most of the value is. Many teams automate the routine 80% and discover the 20% of cases that actually drive business value were the exceptions that required human judgment all along.

Example: IBM ASK HR is a place where automation makes sense — routine cases dominate, exceptions are easy to understand. But most enterprise work doesn't look like that.

Build: The Hardest Call — And Why Most Teams Get It Wrong¶

Build is the right call when the workflow is: - Not suited to purchasing because it's unique - Has lots of edge cases and exceptions - Has company-specific context that matters (your data, your standards, your approval gates, your risk thresholds) - Your team's way of doing the job — the "secret sauce"

The hard part: Most teams don't have bounds around the task. They don't understand the edges of the workflow. They can't answer: "What does good look like at the end?"

The team incentive problem: Your team will be incentivized to tell you "Yep, this is good. We built the AI thing. It's amazing." But can you be the honest third-party eyes that say whether it actually works? Can you distinguish between "the AI thing works" and "the team told you what you wanted to hear"?

The build failure pattern: You go to your team and say "Build it." They say "We can't afford to buy it and can't afford to hire people, so you go build it." And then they build something and you can't tell if it's good. That is not a solution.

What building actually requires: - Repeatable agentic loop with skills, MCP connectors, plugins, data calls, sub-agents - Understanding of what data goes in - Clear definition of what good output looks like - Ability to verify output quality - Bounds around the task and understanding of edges

The conversation most teams aren't having: Do you have bounds around the task? Do you understand all the bits that go into it? Do you know what good looks like at the end? If you can't answer those questions, you cannot successfully build.

Buy: Two Distinct Shapes¶

Shape 1: Buying Primitives (The Easy Version)¶

Buy basic components and services you can stack into many agentic workflows you build. These are fairly easy decisions because your dev team can play with them, build with them, and if they like them they'll use them a lot. No big purchase decision — just starting to build and experiment.

Example: Stripe has put a lot of agentic primitives out there that are easy to get started with. Dev teams can put solutions together without a major commitment.

Shape 2: Buying a Full Workflow Solution (The Hard Version)¶

This is like Harvey for legal — they sell the whole agentic pipeline, and you have to decide if it works with your workflow. The question is: if you were to look at the work you're doing today, is it "Harvey shaped"? Do you know their product well enough to answer that?

The 80-90% overlap test: If you're buying a vendor's workflow solution, you need to know with confidence that there's an 80-90% overlap between the shape of their work and how they envision the workflow versus yours. If there isn't, 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.

The integration question: When buying a workflow solution, you need to understand: - What is the underlying substrate the purchase will sit on? - Is it a data substrate? Where is it going to sit in your system? - How do you know your dev team will be able to integrate it well and actually get you value?

Hire: The Often-Misused Lever¶

Many companies are trying to find "the impossible hire" right now — someone who has deep AI expertise, understands the specific domain, and can translate between the two. The reality is that the talent market for this is extremely tight.

The combination approach: Often the answer is hire + build together. You need someone who can help you make the right calls about what to build, how to build it, and what "good" looks like. But that person doesn't exist in isolation — they need to work with the team that's doing the actual building.

The risk: Hiring someone to "go figure out AI" without a clear definition of what problem they're solving or what success looks like is a recipe for wasted resources and frustration.

Wait: The Underutilized Lever¶

"Do nothing and wait" is a legitimate option when: - The workflow is still too volatile - The AI capability isn't mature enough for this specific type of work - The ROI case isn't clear enough yet

The trap: Teams often use "wait" as a way to avoid making a decision, not as a deliberate strategic choice. The difference matters: a deliberate wait with clear criteria for when you'll revisit is different from inertia.

The Five Inputs for Evaluating Any Workflow¶

Once you've identified a workflow, there are five obvious inputs for evaluation:

1. How often does it repeat? (High frequency = higher automation ROI)
2. How costly is a mistake in that workflow? (High cost = higher stakes, need more human oversight)
3. How much judgment does it need? (High judgment = harder to automate, may need human-in-the-loop)
4. How specific to you is that workflow? (Highly specific = build; generic = buy)
5. Is the market solution mature enough? (Will the next model release eat this workflow?)

The Vendor Reality¶

The reason all of this matters: "There are about 10 million vendors knocking down the door saying, 'Here, we'll sell it to you. This is what you need. I promise you, this is what you need. You don't understand AI, but this is what you need.'"

You need to shut that door temporarily and have the conversation inside the house first — about where you invest, why you invest, and what is likely to yield success.

The procurement conversation you should be having: Before you talk to vendors about AI for a specific workflow, you should be able to answer: What does this workflow do? What does good output look like? How often does it repeat? What are the exceptions? What's the cost of a mistake? How much judgment is required? Is this specific to us or generic? Has the market solved this? If so, how well?

If you can't answer those questions, you're not ready to have the vendor conversation.

The Overarching Message¶

40%+ of Agentic AI projects will fail not because the tech doesn't work — but because teams are making investment decisions without understanding the shape of their own work.

The fix is simple to describe but hard to execute: Before you pick a model, before you pick a vendor, before you pick a dashboard — understand the work.

The workflow is what you're actually investing in. The AI model is just one component of the workflow. Get the workflow right, and the model question becomes much easier to answer. Get the workflow wrong, and no model in the world will save you.

🦐 Summary by Thrawn the Prawn — Strategic Analysis Division