About
Insights
Back to insights

Industry

AI Pilots Without Payoff
The Current State of Enterprise AI

Adoption is nearly universal. Returns are not. The 2025 surveys from RSM and McKinsey expose the same uncomfortable gap—and point to the same way across it.

Joline Zhuang

The current state, in one diagram

AI is everywhere. Business impact is not.

Three years into the generative-AI era, the question is no longer whether companies use AI. It is whether AI changes how the business performs.

0188%

of organizations use AI in at least one business function

0239%

can point to any measurable impact on their bottom line

The payoff gap

Source: McKinsey, The State of AI in 2025. Adoption has become ordinary; measurable enterprise value has not.

01 — The easy part

Adoption is no longer the story

Generative AI has stopped being an experiment. In RSM's 2025 Middle Market AI Survey, 91% of organizations across the U.S. and Canada reported using it—up from 77% a year earlier. McKinsey's global survey tells the same story: 88% now use AI in at least one business function, against 78% the year before.

The most common jobs remain grounded in what generative models do dependably: text generation and summarization, workflow development, forecasting and demand planning, and sales and marketing content.

Leading uses of generative AI — RSM Middle Market AI Survey 2025.
Use caseOrganizations using it
Text generation and summarization49%
Workflow development45%
Forecasting and demand planning40%
Sales and marketing content40%

AI has also moved into the core of operations: 58% use it in data analytics, 57% in IT operations, and 48% in customer service. By every surface measure, the enterprise has gone all in.

02 — The hard part

But adoption is not the same as advantage

Here the story turns. High usage has not translated into high value. In RSM's survey, 92% of companies using generative AI ran into trouble putting it to work. 53% felt only somewhat prepared, and 70% said they needed outside help to get real value.

The enterprise AI funnel — McKinsey, The State of AI in 2025.
StageShare of organizations
Use AI in at least one function88%
Have begun scaling across the enterprise~33%
Report any measurable impact on profit39%
Have AI fully scaled7%
Qualify as genuine high performers6%

Most companies are not failing at AI. They are stuck in the pilot—running something promising that never becomes how the business actually runs.

The scaling gap
03 — The real bottleneck

The model is rarely the problem

When companies say AI is hard, they seldom mean the model is hard. They mean everything around it is. RSM respondents pointed inward—not at the technology, but at the organization's readiness to use it.

Why companies feel unready for AI — RSM Middle Market AI Survey 2025.
ObstacleShare citing it
Data quality issues during rollout41%
Lack of in-house expertise39%
No clear AI strategy34%
Data quality as a reason for unpreparedness32%

None of those are model problems. A world-class model can still sit uselessly on top of a process no one redesigned, fed by data no one cleaned, and owned by no one in particular. AI gets added beside the work instead of built into it.

04 — What the winners do

They redesign the work, not just the tool

McKinsey's small group of AI high performers—roughly 6% of companies—are about 3× more likely than everyone else to have fundamentally redesigned workflows around AI instead of bolting it onto what already existed.

Field-service use cases with measurable payoff — McKinsey, From pilot to profit.
Use caseMeasured outcome
Lead generationMore than $350M in leads across 45,000 customers and 8,000 untapped opportunities
TroubleshootingFirst-contact resolution rose 50%; diagnosis fell from 30 minutes to under 1 minute
SchedulingTechnician capacity rose 40% while overtime fell 6%
Contract analysisA contract agent saved more than €5M a year

Each result is impressive alone. The deeper lesson appears when companies stop treating them as separate tools.

05 — The unlock

Connection is where the value compounds

The biggest gains come from linking use cases—letting one AI step hand off to the next across the service journey, so value adds up instead of sitting in silos.

Connected service workflow

One request, one AI-powered flow

01Issue intake

Request arrives and is logged

02Search documents

AI scans technical manuals

03Root cause

The likely fault is suggested

04Remote check

Can it be fixed without a visit?

05Schedule

The right technician is dispatched

06Predict parts

Spares are prepared in advance

07Service report

The report is drafted automatically

08Invoice

Billing and vendor recovery close the loop

Source: McKinsey, From pilot to profit. The value compounds when every step makes the next one faster.

That is fundamentally different from a chatbot that answers questions. It is an operating layer for intake, diagnosis, scheduling, parts, reporting, and recovery.

Don't deploy one AI chatbot. Build an AI-powered service flow.

The workflow thesis
06 — Our view

Where the value actually gets built

The constraint on enterprise AI is not intelligence—capable models are close to a commodity. The constraint is wiring: clean data, redesigned workflows, governance, real adoption, and people who understand both the business process and what the model can do.

That last ingredient cannot be downloaded. It is why 70% of companies say they need outside help—and why effective help looks less like a software license and more like an engineer inside the operation, mapping the workflow before writing a line of it.

That is the model Kunbyte is built around. We embed with a team, diagnose the workflow first, and build the connected flow—not a demo that dazzles in a meeting and quietly dies in production.

07 — A field guide

Five moves that separate scale from theatre

EY frames the shift from pilot to payoff as five practical disciplines.

Adapted from EY — five actions to escape the AI ROI trap.
MoveWhat it means in practice
Scale through process, not toolsRedesign the whole workflow around AI; do not scatter standalone apps across an old one
Define value up frontTie every build to a measurable KPI: cost, speed, accuracy, revenue, or risk reduction
Govern deliberatelySet enterprise-wide rules, decision rights, compliance, and real ownership
Make data AI-readyCreate clean inputs, traceable lineage, security, and human review where it counts
Run AI as a portfolioBalance safe core bets against a few high-upside experiments

The next winners will not have the most pilots. They will have done the unglamorous work of building AI into how they actually operate.

From pilots to payoff
WorkflowDataGovernanceOwnership