Where AI earns its place

Most organizations are past the question of whether to use AI. Nearly all of them already use it somewhere. The question that separates the ones getting value from the ones running perpetual pilots is different: where does AI earn a place in the operation, and where does a person stay in charge?

Our position across everything we publish is that the value comes from the system around the tool, not the tool itself. Tool-first adoption buys the technology and then hunts for a problem worthy of it. Workflow-first adoption walks the work and lets the work name the spot. What follows is the workflow-first method in two parts: the questions that find the spot, and the rule that places the human gate.

We did not invent these questions, and we will not pretend we did. They come from Stephen Leitch of Purdue's Daniels School of Business, who teaches them in a graduate AI course there. Our founder worked through them during his graduate study at Purdue and recognized what he had already seen in operations: the diagnosis has to start with the work, not the technology. The wording below is ours. The structure is his.

First, where does work slow down? Every operation has points where throughput dies: a queue in front of a single approver, a report that takes a week to assemble by hand. Slowdowns are the most visible AI opportunity because the baseline is measurable and the pain is already acknowledged.

Second, where is expertise trapped in low-value tasks? Watch what your most expensive people actually do in a day. If the controller is re-keying data and the senior engineer is formatting status updates, judgment is being spent on work that does not need it. AI does not remove the need for that expertise. It removes the friction that keeps the expertise from being used.

Third, where are decisions inconsistent? Same input, different answer, depending on who handles it and what kind of week they are having. Inconsistency usually means a decision rule exists but was never written down. That makes the decision a candidate for AI support, and it is also a warning: technology applied to an inconsistent process locks the inconsistency in.

Fourth, where is information trapped? Knowledge that lives in one veteran's head or in an archive nobody searches. Retrieval is something the technology is genuinely good at, and freeing trapped information is often the cheapest win on the list.

The fifth question is different in kind: which decisions still require human judgment and accountability? Where the first four find opportunities, the fifth draws a boundary. Some calls need context the system cannot see. Some need a person who can answer for the outcome. A model can be consistent at any scale. It cannot be accountable.

When an AI project fails, the post-mortem usually blames what is visible: model quality, data quality, hallucinations, infrastructure. Those problems are real. They also sit above the waterline, where everyone can see them and discuss them.

The projects that die usually die lower, a point we took from the same course material as the five questions and keep seeing confirmed in the field. Below the waterline sit governance, change management, trust, workflow integration, and leadership alignment. These rarely get a line in the project budget, and they sink more deployments than the model ever does. A tool that works in the demo and dies in the operation did not fail on accuracy. It failed on adoption, and adoption was nobody's job.

The pattern shows up in the numbers. Yury Korolev, chief executive of AI Decisions in London, synthesized the major industry surveys in 'Artificial Intelligence in 2026: An Analytical Review of Key Trends,' published this year in the Journal of Human-AI Decision-Making Systems. His headline finding has a name, the Gen AI Paradox: most organizations now use AI, and few extract significant value from it.

McKinsey's State of AI survey found 88 percent of organizations using AI in at least one business function, while only 39 percent report any EBIT impact at all. The BCG figures Korolev reports are harsher still: 5 percent of companies extracting significant value, and 60 percent seeing no return.

Korolev's review also carries the number that prices the iceberg. BCG's 10-20-70 rule puts the barriers to transformation at 10 percent technology, 20 percent data, and 70 percent people, organization, and process. Seventy percent of what stops these projects sits below the waterline. A Deloitte finding Korolev cites states the cause in one line: the problem is not the technology, but the attempt to automate existing processes designed for humans.

Korolev reads the current period as a shift from experiments to transformation through agents, systems that carry out multi-step work on their own instead of answering questions and stopping. We read the same shift as raising the stakes on placement. An assistant that drafts a paragraph is wrong cheaply. An agent that executes a workflow is wrong at scale. The better the technology gets at acting, the more it matters where you put it and who stays in charge of it.

This is not a case against the technology. You just have to know how it behaves, because four tendencies show up reliably in current models, and each one forces a decision about where a person stays in the loop.

The model assumes the prompt is directionally correct. Ask it to help you do the wrong thing and it will help you do the wrong thing well. It reinforces the question instead of interrogating the reason behind it, so someone upstream has to own the why.

It is also a pleaser. It tends to validate the framing it is handed rather than challenge it. Bring it a flawed premise and the flaw comes back polished and confident.

On sensitive calls, it goes vague. This is the dangerous one, because vagueness gets misread as endorsement. A hedged non-answer lands in the meeting as 'the AI did not object,' and the decision proceeds as if it had been checked.

Regulated advice, medical or legal questions, it defers on entirely. That is the right behavior, and it widens the gray zone the third tendency creates: the boundary between what the machine will say and what a person must decide gets blurrier exactly where the stakes are highest.

What the model cannot supply, the operation has to: the reason the question is being asked, and the final call on what the answer is worth.

Every deployment therefore needs a point where a human is in charge. The design question is where to put it. The rule we use comes from the same course material as the five questions, a litmus we have adopted whole: put the human gate at the point where a wrong-but-confident output would cause harm before anyone catches it. Trace the output downstream and find the first place a bad answer becomes an action that is hard to reverse. Money moves, a customer hears it, a record changes, a treatment decision leans on it. The gate goes at or before that point.

The rule cuts both ways, and the second edge is the one most advice skips: a badly placed gate subtracts value. Korolev reviews a medical benchmark in which the model alone scored 92 percent while physician-plus-model teams scored 76. The pairing performed worse than the machine on its own. Human plus AI is not automatically better than either alone. The design of the collaboration decides which way it goes.

Over-gating and under-gating fail differently. A gate on every routine output buries reviewers in volume and kills the efficiency case for the whole deployment. No gate where the harm lands lets the vagueness-as-endorsement failure run straight into production. Gate placement, not gate existence, is the design variable.

Finding the right spot is not enough. A workflow can be the correct target and still be unready: wasteful and undocumented. Readiness is what our other frameworks govern. ESSA prepares any single process in strict order, eliminate, simplify, standardize, automate, because you cannot automate what you have not standardized. The Process-to-AI Pyramid sequences the larger climb, from manual work through automation and data to AI, so the technology lands on a foundation instead of on a mess.

What remains is who stays in charge: a person, standing exactly where a confident wrong answer would otherwise become harm before anyone caught it. If you want the method run against your operation, let's talk.