The real question is not agentic versus deterministic. It is whether your operating model can spot the few problems that need both, and fund them.
A bank recently sorted through more than 3000 ideas for using artificial intelligence (AI) and concluded that most of them did not need an agent at all. That is the finding making the rounds. The more useful number is the smaller one hiding underneath it: the share of ideas that were worth real money, and what those had in common.
Enterprise leaders are under pressure to show they are doing AI. That pressure quietly funds agents for problems that never needed one. Every misplaced agent is budget, risk, and attention spent to reach an outcome a simpler system already delivers. But there is a quieter, more expensive mistake on the other side: passing over the few problems where judgement and deterministic execution together would have paid, because the organisation had no way to find them.
The screen that matters is not "does this need AI." It is "does this need both reasoning and deterministic work, and can we actually run that." That is an operating-model question, not a technology one.
The binary is the trap
Most of the current advice stops at a warning: do not over-agent, most things are deterministic, avoid theatre. That is true, and it is also where the thinking usually ends. It is a cost-avoidance posture. It tells you what not to build.
The harder and more valuable insight is about composition. The problems that justify serious investment are rarely purely deterministic and rarely purely agentic. They need intelligent reasoning or judgement in one part and reliable, rule-bound execution in another, working together to produce the outcome. Screening those in, and screening out the process fixes wearing an AI label, is the actual work.
This is not only Massivue's view. Commenting on Google Cloud's 2026 agentic push, GlobalData analyst Beatriz Valle observed that governability and observability, not raw model capability, are what decide whether live deployments succeed or fail, as reported by Verdict. The capability gap is in how organisations govern and run these systems, not in the models themselves.
What a real screen looks like
A concrete, public example. In mid-2026, at Google Cloud Summit Sydney, Bendigo Bank's Chief Technology Officer (CTO) Kieran O'Meara described how the bank runs ideas through a series of gates before any funding, as reported by iTnews.
By his estimate, roughly 80 percent of what got pitched was not truly an agentic problem. It was deterministic work the bank could solve in other ways. He called the pattern "hammers looking for nails." The ideas the bank chose to progress usually needed both: reasoning or judgement and deterministic work to reach the outcome.
The gates are practical. One is economics. In one call-centre example O'Meara cited, an agentic interaction was priced near nine dollars against roughly three and a half dollars for the human-supported path. A perfect model that costs more than the process it replaces is not a win. The others test whether an agentic version would be auditable, what happens if it fails, and whether the data, the integrations, and the team are actually ready.
None of this required a better model. It required a decision discipline applied before the money.
This is an operating-model problem, not a tooling one
Here is the part most organisations get wrong. Running a screen like this once, in a workshop, is easy and nearly worthless. Running it across a portfolio of hundreds or thousands of ideas, repeatedly, with someone accountable for the calls, is a capability. That is the difference between a good slide and a durable result.
Massivue's consistent position is that AI initiatives fail for lack of structured adoption, not for lack of technology. The organisations getting a return are not the ones deploying the most agents. They are the ones with a standing way to tell a deterministic problem from a hybrid one, in order, every week, and to put money only behind the second kind.
How Protum runs the screen
Protum is Massivue's AI operating model. It builds six business capabilities for AI and human integration: Data Culture, Adaptive Structures, Augmented Craft, Responsible Intelligence, Flow Based Interactions, and Impact Prioritisation.
Impact Prioritisation is the screen described above, made permanent. Deciding which problems deserve investment, and in what order, becomes a standing capability rather than a periodic guess. Protum gives it an owner and a cadence through a small team model, and pairs it with a decision framework for resolving conflicts between agents quickly, so the screen has someone accountable for it and a rhythm to run on.
The distinction Massivue's founder, Sandeep Joshi, draws is between fitting an AI to a workflow and looking at the problem to be solved through an operating-model lens. The first produces demos. The second decides what should exist before anything is built.
A screen you can run this quarter
An entry-level version of the discipline. Four questions to put to any AI idea before it is funded:
- Does the outcome genuinely need judgement or reasoning, or is it deterministic work a rule, a script, or an existing system could already do? If it needs both, that is the signal it may be worth building.
- If it goes wrong, is the process auditable, and how large is the damage?
- At the volume you actually run, does the agentic path cost less than the path you use today?
- Are the data, the integrations, and the team ready to deliver it?
An idea that needs both judgement and deterministic work, and clears the other three, is worth real money. An idea that fails the first question was never an AI project. It was a process fix in disguise.
The four questions handle one idea. The harder problem, and the one worth an operating model, is running that screen across every idea in the pipeline without it collapsing back into a workshop.
The posture, not the summary
Going beyond AI theatre is not about doing less AI. It is about the discipline to find and fund the few problems where judgement and deterministic execution together earn their cost, and to solve the rest the cheaper, more auditable way. As Sandeep Joshi puts it, AI is "not a technical problem to be solved, it is a leadership challenge to be embraced."
If your AI backlog has outgrown a workshop, the screen becomes an operating-model question. That is the conversation Massivue's AI Transformation practice is built for. See massivue.com/services/ai-transformation.
Sources: Kieran O'Meara's remarks are drawn from iTnews' report of Google Cloud Summit Sydney (8 July 2026). Analyst commentary is from Verdict's coverage of Google Cloud's 2026 summit programme.