Generative AI Reality Check: Expectation and Reality

AI

When we first put generative AI on the agenda, the same sentence was circulating at most tables: “Let’s do something—so competitors don’t outrun us.” That instinct isn’t bad; it’s even healthy. The problem is this: once we start running, we realize the track isn’t asphalt—it’s a construction site. Pilots launch, demos get applause, but business outcomes don’t arrive at the expected speed. That’s where a real reality check is needed.

At CBOT, we repeatedly see the same breaking point in large organizations’ AI journeys: those who manage generative AI as a “technology project” quickly get stuck; those who manage it as a “business transformation” generate lasting gains in both productivity and quality. McKinsey’s data makes exactly this distinction visible: there is adoption, there is excitement, but there is also a discipline-demanding distance when it comes to scaling and financial impact.

Expectation 1: “We’re using it” = “We’ve scaled it”

One of the most common sentences we hear in the field is: “We use generative AI.” But what does that mean? If one team uses a tool a few times a week, it’s called “we’re using it”; the same is said when the approach is integrated into ten processes and measured.

McKinsey’s State of AI findings show that usage is rising, but the conversion into enterprise-scale value is not progressing at the same pace. What that translates to on the ground is: there are bright wins at the unit level, but there is no “shared operating system.”

“Scale is not growing one use case; it’s being able to manage ten use cases with the same principles.”

Reality check: Scaling is less about adding technology and more about designing management.
What works?

  • We establish common components from the start: knowledge source, access model, logging, evaluation metrics, approval checkpoints.

  • We build a “repeatable template” instead of a “one-off pilot,” so every team doesn’t start from scratch.

Expectation 2: ROI comes from the model

Another strong assumption is: “If we choose the right model, ROI will follow.” Model choice matters. But what truly produces ROI is how the work flows. One of the critical points in the McKinsey report is that value capture is more visible among those who redesign workflows.

In CBOT’s field experience, it shows up like this: everyone adds an “assistant” layer (text generation, summarization, drafting). Speed increases, but because the end-to-end process hasn’t changed, the gain remains “scattered.”

Reality check test (very simple):

  • How many steps did this work take before generative AI?

  • How many steps now? Which steps disappeared, which moved, which became control gates?

If the number of steps is the same, you’re probably just accelerating typing. Transformation hasn’t started yet.

Expectation 3: A digital worker is an “autonomous worker”

The concept of a digital worker is exciting—because we’re talking about systems that can plan and execute parts of multi-step work. The McKinsey report highlights rising interest in digital workers and notes that many organizations are experimenting in this area.

But the reality in the field is this: taking a digital worker live is not assigning a task to a model. It means designing process access, authority, oversight, and rollback.

“Taking a digital worker live is like opening a new role in the company: it has authority, boundaries, oversight, and performance measurement.”

What do we do? (CBOT approach)

  • We progress with supervision: human approval at critical steps, especially in processes involving financial transactions and customer statements.

  • We grant authority gradually: read → recommend → initiate transaction → complete transaction.

  • We make every step traceable: what data was used, what decision was made, what output was produced.

This approach grows both speed and trust at the same time. Because “control” isn’t a delaying brake; it’s the seatbelt of scale.

Expectation 4: Risk management can be added “later”

In generative AI, risk usually shows up at two extremes: either it’s never discussed, or it’s discussed and then left on the shelf. Yet risks are no longer theoretical in the field. McKinsey notes that a significant share of companies using AI have experienced at least one negative outcome, and one of the most common issues is incorrect output.

Reality check: Adding risk management later is like reinforcing the foundation after the building is finished. Possible—but expensive.

Minimum set (CBOT’s “must-have” list):

  • “When is human approval mandatory?”—clear and written.

  • “Single source of truth” approach for corporate knowledge: current documentation, ownership, versioning.

  • Evaluation metrics: accuracy, consistency, coverage, risky-expression detection.

  • Logging and audit trail: understandable not only for technical teams but also for business and compliance teams.

Expectation 5: Culture will adapt on its own

The most critical yet most invisible reality: generative AI changes how work gets done. When the way of working changes, role definitions and responsibilities change too. If this isn’t managed, two things happen:

  • “Pilot fatigue” begins.

  • Trust breaks at the first friction point.

Reality check: Managing expectations is as important as the technology.
The approach that works for us in the field: tie small but measurable wins to “transformation.” Not just saying “we saved time,” but being able to say: “we removed this step in the process, added this control, and track quality with this metric.”

CBOT Playbook: 4 steps that turn expectation into reality

Especially in operations-heavy organizations such as finance, retail, airlines, e-commerce, and customer service, the sequence below accelerates value:

1) We choose a process, not a use case

We target the processes that repeat most, produce the most errors, and cause the most delays.

2) We redesign the workflow on paper first

Where does generative AI engage, where does it stop, where are the control gates? If this isn’t clear, it won’t become clear in production; it will grow in production.

3) We build the trust layer from the start

Authorization, verification, logging, evaluation, and compliance processes… These are not “later”; they are part of the design.

4) We manage it like a product

SLA, quality metric, usage metric, risk metric. Not a one-off project, but a living product.

The reality check in generative AI tells us this: the real issue isn’t the model—it’s the operating system. As McKinsey points out, usage is rising; but for scale and value, workflow design, governance, and discipline are essential. The clearest distinction we see in the field at CBOT is exactly this: those who talk about models run pilots; those who talk about process and governance scale.

For organizations that design the digital worker correctly, generative AI becomes not a “hype,” but a lasting competitive advantage that reduces cost, increases quality, and speeds up decision-making. And the key to this transformation is simple: align expectations with reality, redesign the process, build trust into the design, then scale.