Sr Director of Products at Dell and former Agile Alliance Board Chair, scaling businesses from version zero to legacy revival.getty​You approved the AI budget. Six or seven figures. The pilots ran. Some delivered; most stalled—S&P Global found 42% of organizations had abandoned most AI initiatives in 2025, up from 17% the year before. Now you are facing a question no vendor prepared you for: What are you actually betting your operating model on?That is not a technology question. It is a capital allocation decision, an organizational commitment and a talent bet. Getting any one of them wrong is recoverable. Getting all three wrong at once is not.When Klarna ended its Salesforce and Workday contracts, most coverage missed what mattered. Revenue roughly doubled between 2022 and 2025 while headcount fell by nearly half—but Klarna also rebuilt human customer service after satisfaction declined on automated interactions. They now run a disciplined hybrid: AI at volume, humans on complexity. Owning the capability meant owning the failure mode, and having the visibility to correct it. That is the decision you are actually being asked to make.Subscription And Owning The OutcomeMost enterprises will not cancel Salesforce or Workday outright. They will cap seats, shift work to internal agents and renegotiate on different terms. SaaS is not disappearing—the boundary between what you rent and what you own is moving, and faster than most renewal cycles. The vendors adapting fastest know this too; the shift toward outcome-based pricing is them acknowledging the same reality.The SaaS model trained executives to think about capability as a subscription. Someone else owns the complexity; you own the outcome. That model breaks the moment capability becomes autonomous. You cannot subscribe to organizational judgment. You cannot rent the business rules that govern your company’s decisions, because those rules are the company. When an agent executes a refund policy, a credit decision, an escalation rule—that is your operating model running in production. You can outsource where it runs. You cannot outsource what it decides.The Complexity Of Internalization​Here is the question missing from most board decks: When the per-seat SaaS fee goes away, where does the money actually go? This is build versus buy—GenAI shifts the cost curve but does not eliminate it. A $2 million contract typically reappears as 30%–45% in model inference, 20%–30% in data engineering and tooling and two to four FTEs to govern what the agents decide; the first-year outcome is reallocation, not saving.There is a second cost dynamic most business cases are not capturing. AI operating costs inflate in aggregate even as unit prices fall—consumption scales faster than pricing drops. Enterprises that have deployed AI tools broadly are rationing access: employees get enough to experiment, not enough to depend on. The budget to close that gap already exists in the technology portfolio—in SaaS contracts being consolidated, in infrastructure being replaced. Governing your agent estate means having the visibility to direct that spend toward what is actually working.Guideline​sThe line between building and buying has moved—not away from vendors, but up the stack. Today, you buy the model infrastructure, you build the proprietary logic that runs on top of it. DIY no longer means writing code. It means orchestrating agents with the discipline you apply to anything else running your business: versioned, deployed, monitored, governed. A rule I applied running a $380 million P&L: If the workflow touches proprietary data, proprietary judgment or proprietary economics, you own it. Everything else you rent. That rule is crude. More useful than any vendor framework.The real risk is not security—that gets solved at the infrastructure layer. Nor model quality; that curve moves fast enough to embarrass anyone banking on today’s limitations. The real risk is governance—and enterprise RPA is where this movie was previewed. Between 2017 and 2022, enterprises deployed thousands of automation bots with no deployment standard, no versioning, no life cycle management. By 2023, the pattern was consistent: bots in production, nobody certain what they do and engineers afraid to touch them. I have sat in those reviews. Agents are RPA with reasoning and a much larger blast radius. The governance failure mode is identical; the scale of potential sprawl is orders of magnitude larger. If you cannot answer “what agents do we have, who owns them, and how do we roll them back?”—you already have sprawl.Final Thoughts​Organizations that govern their agent estates now will be the informed buyers when the orchestration standard arrives. Those that wait will inherit a shadow estate nobody can inventory, let alone migrate. The gap between those two outcomes is being created right now.Three questions worth putting on your next leadership agenda: Which workflows touch proprietary judgment—and who owns the logic inside them? When you reduce a SaaS contract, have you mapped where that spend goes? And does your team know how to version, monitor and roll back an agent the way they would a production service?If you cannot answer all three, you are not running your operating model—your vendors are.Forbes Technology Council is an invitation-only community for world-class CIOs, CTOs and technology executives. Do I qualify?