Automate, Build, Buy, Hire, or Wait

Gartner expects more than 40% of agentic AI projects to be killed by 2027 — cost, unclear value, inadequate controls. The reflex is to blame the technology. That's wrong; productive agentic workflows are everywhere, which is why the money keeps flowing. Projects fail because the investment question was framed wrong from the start.

It's not an AI question

AI investment is a question about the shape of your work. The model is downstream of that; the vendor is downstream of that. Most teams skip the step — often because a vendor encourages them to — and pile a dozen different jobs into one RFP. An accounts-receivable team doesn't have one AI problem; it has collections prioritization, invoice matching, dispute resolution, cash application, and more — each a different shape of work that routes to a different decision. The unit of decision isn't the department. It's the workflow: the whole loop of inputs, allowed actions, what good looks like, who checks, what escalates, and who is accountable.

Five levers, one per workflow

Score every workflow on how often it repeats, how costly a mistake is, how much judgment it needs, how specific it is to you, and whether the next model release will simply eat it. Then pull one of five levers:

• Automate — routine dominates, exceptions are definable, output is cheap to check.
• Build — the work is yours: your data, standards, approval gates, secret sauce.
• Buy — an off-the-shelf workflow genuinely overlaps yours (80–90%, not 40%).
• Hire — for the specific gap, not a mythical "AI unicorn."
• Wait — deliberately, where leverage is low, so you spend scarce change-management capacity where it pays.

One line belongs in every AI investment review: do not automate what you cannot describe. The vendor's demo always shows the routine case; your production traffic is exceptions. If you can't state the inputs, outputs, standards, and exceptions in plain English, you're not ready to invest.