Productivity is not value realisation. Unless organisations redesign work and make deliberate decisions about released capacity, AI benefits can disappear inside the operating model.
Enterprise AI has developed an accounting problem.
Not a financial-accounting problem in the conventional sense.
A management-accounting problem.
Organisations are increasingly able to demonstrate that AI makes individual activities faster. Documents take less time to produce. Analysis can be accelerated. Service requests can be resolved more efficiently. Repetitive work can be automated.
Yet a faster task does not automatically produce a better P&L.
New research this week exposes the gap.
BearingPoint surveyed 1,050 C-suite executives and senior leaders across 13 countries. Among organisations that have implemented AI, 74% report measurable top-line or bottom-line impact.
That sounds encouraging.
Yet only 13% have scaled AI completely in line with the original business case.
More revealingly, 62% report that AI has already created workforce overcapacity of at least 10%.
That number deserves more executive attention than another model benchmark.
Because it raises a question many AI business cases avoid:
What happens to the capacity AI releases?
A saved hour has no intrinsic financial value
Suppose an AI-enabled workflow reduces a ten-hour activity to eight hours.
Two hours have been released.
It is tempting to record that as a 20% productivity improvement and attach an employee-cost calculation to it.
The arithmetic may be correct.
The economics may not be.
If the employee still works the same hours, the organisation still carries the same payroll and the same fixed cost.
The two hours have created capacity, not automatically cash.
That capacity can become economically valuable.
The organisation might use it to increase throughput without increasing headcount. It might absorb growth that would otherwise require recruitment. It might improve service. It might redirect people towards higher-value work. It might remove roles through natural attrition or redesign. It might reduce external expenditure.
Each is a legitimate value mechanism.
Each requires a management decision.
Without one, the saving can simply disappear into the working day.
That is the difference between productivity and value realisation.
The problem starts further upstream
This week’s research also provides an explanation for why scaling remains difficult.
The Hackett Group and ARIS examined process context in Global 2000 organisations.
86% of business leaders say AI agents cannot be deployed reliably without process context. Yet only 22% say their organisation has comprehensive, real-time visibility into its end-to-end processes. Another 59% describe process visibility as fragmented across systems and functions.
Organisations with strong process-context capabilities were five times more likely to report very successful AI outcomes, although that association should not be interpreted as proof of causality.
The underlying problem is straightforward.
You cannot intelligently redesign work you do not understand.
A process diagram is not necessarily enough.
Real work contains:
- decisions;
- hand-offs;
- exceptions;
- informal knowledge;
- approval thresholds;
- controls;
- system constraints;
- duplicated activity;
- queues and delays;
- workarounds; and
- dependencies between functions.
This is where technology-first AI programmes often make a category error.
They ask:
“What can this AI automate?”
The better question is:
“Why does this work happen this way, and what should the future workflow actually be?”
Those are fundamentally different questions.
Automating a task can preserve a bad operating model
Consider a process requiring information to move through six stages and three approvals.
An AI agent might make stages two and four substantially faster.
That does not tell us whether all six stages are necessary.
It does not tell us whether the three approvals reflect genuine risk or organisational history.
It does not tell us whether the information being created at stage two is subsequently recreated elsewhere.
It does not tell us whether one function’s productivity improvement creates a queue in another.
The danger is that AI becomes a very efficient mechanism for preserving yesterday’s workflow.
That is optimisation.
Transformation asks whether the workflow itself should survive.
Workforce adoption introduces another complication
PwC’s new Global Workforce Hopes and Fears Survey provides another part of the picture.
Across nearly 50,000 workers, 64% report using AI at work, up ten percentage points in a year. Daily GenAI usage has risen from 14% to 22%.
Yet those benefits are not evenly distributed.
PwC identifies 14% of workers as AI-enabled “front-runners”, while 56% sit within an operational “engine room” cohort with considerably lower access to AI and development opportunities.
Less than 40% of that core group report having the learning and development resources they need, compared with almost 80% of front-runners.
That matters because enterprise transformation cannot depend indefinitely on a small group of sophisticated adopters.
If the people closest to operational work are not involved in redesigning it, organisations risk creating another form of fragmentation: advanced individual AI use sitting on top of largely unchanged enterprise processes.
The objective is not simply to teach everybody how to prompt.
It is to build the capability to operate a redesigned system of work.
Value needs to be designed backwards from the outcome
A stronger AI investment sequence therefore begins somewhere different.
Start with the economic outcome.
Then work backwards.
What needs to improve?
Is it cost-to-serve?
Cycle time?
Revenue capacity?
Customer retention?
Quality?
Risk exposure?
Working capital?
Employee capacity?
Once the outcome is explicit, examine the workflow producing it.
Where is value leaking?
Which activities genuinely require human judgement?
Which can be augmented?
Which can be automated?
Which should disappear completely?
Only then does the technology decision become meaningful.
The same discipline applies after deployment.
If an initiative releases 5,000 hours annually, management should know what happens to those 5,000 hours.
If throughput increases, measure the throughput.
If recruitment is avoided, identify the avoided cost.
If people move into higher-value activity, measure the resulting outcome.
If quality improves, measure the cost of defects or rework before and after.
Otherwise, the organisation risks claiming theoretical savings while continuing to carry the same economic structure.
TFx experience reinforces the distinction
This distinction has been visible in TFx’s previous enterprise transformation experience.
In an anonymised Fortune 500 environment spanning approximately 15,000 users, AI adoption increased from 15% to 68%.
Measured outcomes included 38% greater efficiency, 73% lower resolution time and 63% lower compliance effort, contributing to $10–12 million in annualised ROI.
Those results should not be treated as a generic benchmark for other organisations.
Their relevance lies elsewhere.
The outcomes required more than access to AI.
They involved workflow integration, operating-model design, governance, workforce enablement and measurement.
The technology changed what was possible.
The organisation had to change how work happened in order to capture the value.
Agentic AI makes this discipline more important
Agents intensify the issue because they can execute work rather than merely assist with it.
Last week’s TFx analysis examined the need for managed autonomy: explicit ownership, authority, controls, measurement and lifecycle management for AI agents.
This week’s evidence adds another requirement.
Context.
An agent cannot be given sensible decision rights if the organisation has not identified the decisions within the workflow.
Runtime controls cannot enforce appropriate boundaries until someone defines those boundaries.
An agent cannot optimise towards a meaningful business outcome if the organisation has not established what that outcome is.
NVIDIA’s newly announced Open Agent Safety Platform illustrates how rapidly technical enforcement is advancing. Its OpenShell runtime is designed to enforce agent boundaries, while Sentry provides an independent hardware-level monitoring and intervention layer.
That is an important development.
It also demonstrates why governance cannot be delegated to technology.
A control system can enforce a boundary.
The operating model must decide where the boundary belongs.
The management question is changing
AI has spent several years forcing organisations to ask whether the technology works.
Increasingly, that is not the hardest question.
The harder questions concern the enterprise around it.
Can we see how work actually happens?
Can we identify where economic value leaks?
Can we redesign the workflow rather than simply accelerate its existing tasks?
Can we distribute capability beyond a small population of advanced users?
Can we govern increasing autonomy inside the workflow?
Can we measure what changed?
Most importantly, can management convert the capacity created into an observable economic outcome?
That is where AI transformation becomes enterprise transformation.
The organisations that solve this will not simply become faster.
They will become better at turning technological capability into organisational performance.
That is a much harder advantage to copy.
The executive question
If AI released 10% of your organisation’s capacity tomorrow, could you explain precisely how that capacity would become measurable enterprise value?
Leave a Reply