Enterprise AI Has a Control Plane. It Now Needs a Value Plane.
Why enterprises need to move beyond AI consumption management toward outcome, productivity and value management
AI is rapidly becoming embedded in the enterprise.
Organizations are investing in AI assistants, coding agents, model gateways, enterprise APIs, agents and increasingly sophisticated AI applications.
At the same time, a new management discipline is emerging.
Organizations are getting much better at answering questions such as:
Who is using AI?
Which models are being used?
How many tokens are being consumed?
Which teams are spending the most?
What does each model cost?
Where should access be restricted?
How can we control AI spend?
These are important questions.
But they leave a much bigger question unanswered:
What value are we actually creating with all of this AI consumption?
I believe this is becoming one of the most important gaps in enterprise AI.
Many organizations are building an AI Control Plane.
They now need an AI Value Plane.
The Token-Maxxing Trap
Tokens are an important AI economic metric.
They tell us how much AI capacity is being consumed and help organizations understand model costs, capacity requirements and opportunities for optimization.
But tokens are fundamentally a measure of consumption, not value.
Consider two AI workloads.
One consumes very few tokens but produces little useful output.
Another consumes substantially more because it is solving a difficult engineering, research or reasoning problem — but creates significant productivity or business value.
If we optimize primarily for token consumption, the first workload can look better.
If we optimize for outcomes, the second may be dramatically more valuable.
This creates what I call the token-maxxing trap:
Optimizing AI consumption instead of optimizing the value created by AI.
The answer is not to stop managing tokens.
The answer is to put tokens in the right place within a broader value framework.
Tokens should be:
- A cost metric
- An engineering metric
- A capacity metric
- A FinOps metric
- An optimization variable
They should not be the ultimate KPI for enterprise AI.
From Tokens to Value
The enterprise AI measurement hierarchy should look something like this:
Each level answers a different question.
Tokens
How much AI are we consuming?
Usage
Where and by whom is AI being used?
Task productivity
Is AI helping people complete work more effectively?
Process improvement
Is AI changing how the organization operates?
Business outcome
Is AI changing something that matters to the business?
Enterprise value
Is the value created greater than the cost and risk of deploying AI?
The further we move up this hierarchy, the harder the measurement becomes.
But it also becomes much more meaningful.
Three Different Types of Enterprise AI
One reason AI governance is difficult is that “AI usage” is not one thing.
I see three fundamentally different categories.
These categories have different economics, different governance requirements and — importantly — different definitions of value.
1. AI Exploration & Innovation
This is where business analysts, developers, innovation teams and AI practitioners explore what AI can do.
They may:
- Experiment with new models
- Build prototypes
- Explore agents
- Test RAG
- Generate and test code
- Compare model capabilities
- Automate processes
- Develop proofs of concept
- Investigate new use cases
The defining characteristic is uncertainty.
We don’t yet know whether the idea will create value.
Therefore, the right question isn’t:
How many tokens did we consume?
It is:
What did we learn, and did that learning create a credible path to value?
This is option value.
A failed experiment isn’t necessarily a failure.
If an experiment quickly establishes that an idea should not be pursued, it may have saved the organization from making a much larger investment.
What should we measure?
The funnel should look like:
The relevant metrics include:
- Experiments completed
- Hypotheses tested
- Prototypes created
- Reusable assets
- Validated use cases
- Production candidates
- Conversion to production
- Estimated opportunity value
- Realized value
The important question becomes:
How much useful learning and future opportunity are we generating per unit of experimentation investment?
2. AI in Production
Production AI is fundamentally different.
Here AI is performing a defined business function.
Examples include:
- Customer service
- Software engineering
- Document processing
- Claims handling
- Compliance
- Knowledge management
- Sales support
- Forecasting
- Enterprise agents
- Operational automation
At this point the organization should become much more rigorous.
Every significant production AI capability should have:
- A business owner
- A clearly defined outcome
- Quality requirements
- Risk classification
- Cost targets
- Performance metrics
- Monitoring
- Value measurement
- Criteria for scaling or retiring the capability
The relevant economic question changes from:
How many tokens does this application consume?
to:
What does it cost to achieve the required business outcome?
For example, instead of simply measuring tokens for a customer-service AI system, we should understand:
- Cost per interaction
- Cost per successfully resolved case
- Resolution rate
- Quality
- Escalation rate
- Customer satisfaction
- Overall business benefit
Tokens still matter.
But they are now an input into the economics of the solution.
3. AI Democratization: AI as an Enterprise Utility
The third category is AI becoming a normal part of everyday work.
Employees may use AI to:
- Draft communications
- Summarize documents
- Prepare for meetings
- Research topics
- Analyze information
- Brainstorm
- Create presentations
- Explain unfamiliar concepts
- Organize information
- Generate simple code
- Learn new subjects
Some interactions will have obvious productivity benefits.
Others will have diffuse or difficult-to-measure value.
An employee asking an AI assistant to explain a topic may not have a measurable business outcome attached to the interaction.
That doesn’t mean the interaction has no value.
It means the value is distributed rather than concentrated.
This is why enterprise AI should increasingly have a standard utility layer.
The objective is:
Make safe, capable AI broadly available and allow employees to incorporate it naturally into their work.
The organization should not need a business case for every AI interaction.
Three Categories Require Three Management Models
The three categories should not be governed identically.
This is why a single enterprise AI budget policy is unlikely to be sufficient.
Different types of AI work require different economic models.
The Question of Model Selection
As enterprises adopt multiple AI providers and models, another question emerges.
Should every user have to decide which model to use?
I don’t think so.
For many workloads, model selection should become increasingly autonomous.
The user should describe the task.
The platform should determine the appropriate level of intelligence.
The routing layer can consider:
- Task complexity
- Reasoning requirements
- Context length
- Data sensitivity
- Quality requirements
- Latency requirements
- Cost constraints
- Available models
- Historical model performance
A simple task may go to a lightweight model.
A difficult analytical problem may require a stronger reasoning model.
A complex coding or research problem may justify a frontier model.
The objective is not:
Always use the most powerful model.
Nor is it:
Always use the cheapest model.
The objective is:
Use the least expensive level of intelligence capable of reliably achieving the required outcome.
This changes the conversation around token optimization.
The system should not simply minimize tokens.
It should optimize the cost of achieving the required outcome.
Intelligence Should Become an Allocated Resource
This can go further.
The platform doesn’t necessarily need to decide only which model to use.
It can decide:
- How much reasoning to apply
- How much context to provide
- Which tools to invoke
- Whether verification is necessary
- Whether the task should be escalated
- Whether multiple agents should be used
Simple task:
Low reasoning + lightweight model
Complex task:
More reasoning + stronger model
High-risk or difficult task:
Deep reasoning + tools + verification + escalation
This leads to an important idea:
Intelligence should become an allocated resource, just like compute and money.
The platform should dynamically determine how much intelligence to spend based on the complexity, value and risk of the task.
The AI Control Plane
Many enterprises are already building an important part of this architecture.
They have:
- Model gateways
- Authentication
- API keys
- Approval workflows
- Budgets
- Rate limits
- Model access controls
- Token tracking
- Cost attribution
- Provider management
- Observability
This is essential infrastructure.
I would call it the:
AI Control Plane
It answers:
Who is allowed to use AI?
Which models can they access?
How much are they consuming?
What does it cost?
Which provider is being used?
Which team or application is responsible?
Conceptually:
This is the foundation for responsible AI consumption.
But it is not enough.
The Missing Feedback Loop: From Value Selection to Value Realization
There is an important nuance here.
Many organizations are not starting from zero when it comes to AI value.
They already use AI use-case value matrices to decide where to invest. Proposed use cases are often assessed against dimensions such as:
- Business impact
- Expected ROI
- Strategic alignment
- Feasibility
- Data readiness
- Complexity
- Risk
- Time to value
This is good practice. It helps organizations avoid treating every AI idea as equally important and provides a rational way to prioritize investments.
But there is a problem.
The value assessment often stops when the investment decision is made.
Once a use case is approved and deployed, the organization typically moves into a different management mode. The focus shifts to operational performance, adoption, model usage, tokens and cost.
The original value hypothesis can effectively disappear.
This creates a disconnect:
Before deployment, we ask:
Is this AI use case valuable enough to build?
After deployment, we ask:
How much is it being used, how well is it performing, and what does it cost?
What is often missing is the question that connects the two:
Did the deployed AI capability actually deliver the value we expected when we approved the investment?
That is the missing feedback loop.
The Value Hypothesis Should Survive Deployment
Every significant AI use case should begin with a value hypothesis.
For example:
If we introduce AI into this process, we expect to reduce processing time by 30%, maintain quality above 95%, and increase throughput without additional headcount.
That statement should not become irrelevant once the system goes live.
It should become the baseline against which the production system is evaluated.
The organization should be able to compare:
Expected
- Expected productivity improvement
- Expected quality
- Expected cost
- Expected adoption
- Expected business impact
with:
Actual
- Actual productivity improvement
- Actual quality
- Actual AI cost
- Actual adoption
- Actual process performance
- Actual business outcome
The objective is not to demand perfect attribution.
The objective is to establish an evidence-based feedback loop.
The Missing Layer: The AI Value Plane
There is an obvious question at this point:
How do you actually build a Value Plane?
It is tempting to think of it as another dashboard showing AI usage alongside business KPIs.
That is not enough.
The real challenge is connecting AI telemetry to the work being performed and the outcomes generated by that work.
An AI gateway can tell us that a request consumed 80,000 tokens and cost €0.40.
It cannot, by itself, tell us whether those tokens helped close a €500,000 opportunity, resolve a customer issue, complete a compliance review or simply generate an interesting answer.
That context lives elsewhere in the enterprise.
The Value Plane therefore needs to connect several layers of evidence:
The AI gateway provides the consumption data.
The application provides the workflow context.
Enterprise systems such as CRM, ERP, ITSM or HR systems provide operational outcomes.
The data and BI layer provides the ability to correlate these signals over time.
The mechanism can be as simple as a use-case ID, workflow ID, application ID or correlation ID carried through the AI interaction and the downstream business process.
For example, an AI sales assistant might generate research for a specific opportunity.
Get Sachin’s stories in your inbox
Join Medium for free to get updates from this writer.
The AI layer can capture:
- Model used
- Tokens consumed
- Cost
- Request ID
- Application
- Timestamp
- User or team
The CRM can capture:
- Account
- Opportunity
- Sales stage
- Opportunity value
- Win/loss
- Time to close
The Value Plane can then connect the two.
This does not mean claiming that one AI prompt caused one closed deal.
That would create a false level of attribution precision.
Instead, the objective is to build an increasingly strong evidence chain between AI investment, work performed and business outcomes.
This distinction matters.
The Value Plane is not simply a reporting layer. It is an attribution and evidence architecture connecting AI consumption to enterprise activity and outcomes.
Measuring Category 3: From Time Saved to Capacity Created
One of the most attractive metrics for enterprise AI democratization is time returned.
It is also one of the easiest to misuse.
An employee may report:
“This AI tool saves me two hours every week.”
That is useful information.
But it is not necessarily evidence that the organization gained two hours of productive capacity.
The employee might spend those two hours on something else, take on additional work, improve quality, collaborate more, or simply finish earlier.
Self-reported time savings are therefore a signal, not a financial outcome.
The measurement hierarchy should be:
At the individual level, we can ask employees which tasks AI helped with and how their work changed.
At the task level, we can measure changes in cycle time where system data exists.
At the process level, we can look at throughput, backlog, quality and rework.
At the business level, we can ask whether the additional capacity actually translated into something valuable.
For example:
AI saves a service employee 30 minutes per day.
is interesting.
But:
The service organization now handles 15% more cases with the same staffing level while maintaining quality.
is much stronger evidence.
And:
The additional capacity reduced backlog, improved customer response times and avoided incremental operating cost.
is stronger still.
This also does not require turning the enterprise into a surveillance system.
In many cases, the relevant evidence already exists in business systems:
- Case volumes
- Processing times
- Cycle times
- Backlogs
- Throughput
- Quality
- Rework
- Service levels
- Customer outcomes
The goal should not be to monitor every employee’s AI interaction.
It should be to understand whether AI is changing the capacity and performance of the organization.
That is a very different proposition.
A Better Measurement Model for AI Democratization
For Category 3, I would therefore use three levels of measurement.
Level 1 — Adoption
Are people using AI?
Examples:
- Active users
- Usage by role
- Adoption by business unit
- Use-case patterns
This tells us whether the capability is being adopted.
It does not tell us whether it is valuable.
Level 2 — Productivity Signals
Is AI changing how work is performed?
Examples:
- Self-reported time saved
- Task completion time
- Rework
- Quality
- Employee experience
These are useful leading indicators.
Level 3 — Organizational Evidence
Is the released capacity translating into meaningful improvement?
Examples:
- Throughput
- Backlog reduction
- Cycle-time improvement
- Capacity increase
- Service-level improvement
- Cost avoidance
- Revenue capacity
This produces a more credible progression:
Adoption → Productivity → Capacity → Outcome
Not every AI interaction will reach the final level.
That is okay.
The objective is to measure at the highest level that is practical and meaningful for the use case.
Measuring Category 1: Learning Rather Than Consumption
Exploration presents a different challenge.
When teams are experimenting with AI, there may be no immediate business outcome.
Trying to calculate ROI for every experiment would discourage innovation.
The better question is:
What did we learn, and did that learning create a credible path to future value?
The exploration funnel therefore becomes:
Useful metrics include:
- Experiments completed
- Hypotheses tested
- Time to prototype
- Reusable assets created
- Validated use cases
- Production candidates
- Conversion to production
- Estimated opportunity value
- Realized value
Token consumption remains important here, but as the cost of experimentation.
The goal is not:
“Use fewer tokens.”
It is:
“Generate more useful learning and validated opportunities for the investment we make.”
A failed experiment can therefore have positive value if it prevents the organization from pursuing an expensive or unsuitable idea.
Measuring Category 2: Business Outcomes and Unit Economics
Production AI should have much stronger economic discipline.
Instead of:
“This application consumes 300 million tokens.”
we should be able to say:
“This application costs €X per successful outcome, achieves the required quality, reduces processing time by Y%, and creates €Z of annual business benefit.”
This allows intelligent optimization.
If a more expensive model improves the business outcome sufficiently to justify the additional cost, use it.
If a cheaper model delivers the same outcome, use the cheaper model.
The objective becomes:
Optimize business value subject to cost, quality and risk constraints.
This is also where autonomous model routing becomes particularly powerful.
The production platform can continuously learn which model provides the best economics for a particular task.
A model that is expensive but substantially better may be justified for high-value decisions.
A cheaper model may be sufficient for routine work.
The optimal model is therefore not a fixed enterprise-wide decision.
It can be workload-specific and dynamic.
Risk Is a Constraint, Not Always a Price Tag
There is another complication in the economics of AI: risk.
It is tempting to write:
AI Value = Business Value − AI Cost − Risk Cost
The problem is that risk is not always easy to price.
For some risks, expected loss can be estimated:
Expected Risk = Probability × Impact
For example, an organization might estimate the expected cost of a certain type of operational error.
But other risks are much harder to quantify.
A serious compliance failure, security incident, safety issue or reputational event may have consequences that are difficult to estimate reliably in advance.
Some risks should therefore not simply be treated as another number that can be traded against AI savings.
Instead, the Value Plane should operate with risk and quality constraints.
Conceptually:
Maximize business value subject to acceptable cost, quality, risk and compliance constraints.
This is a more realistic way to think about AI economics.
The objective is not to assign a euro value to every possible risk.
The objective is to ensure that economic optimization does not override risk boundaries.
In other words:
Value optimization happens inside a risk envelope.
A More Realistic AI Value Equation
I would therefore avoid treating AI economics as a simple equation where every component can be reduced to a precise financial number.
A more practical formulation is:
AI Value = Measured Outcome Value − Total AI Cost
subject to:
- Required quality
- Acceptable risk
- Compliance requirements
- Security requirements
- Operational constraints
This distinction is important.
AI cost should absolutely be measured.
Business outcomes should increasingly be measured.
Risk should be assessed rigorously.
But we should not pretend that every risk can be reduced to a precise financial number.
The Value Plane Is a Journey, Not a Switch
Organizations will not suddenly go from token dashboards to perfect AI ROI measurement.
The Value Plane should mature progressively.
Level 1 — Consumption Visibility
We know:
- Who is using AI
- Which models
- How much
- What it costs
Level 2 — Use-Case Attribution
We know:
- Why AI is being used
- Which application
- Which process
- Which project
Level 3 — Performance Measurement
We know:
- Quality
- Task performance
- Human intervention
- Productivity signals
Level 4 — Outcome Measurement
We can connect AI-enabled work to:
- Process outcomes
- Operational metrics
- Customer outcomes
- Revenue
- Cost
- Capacity
Level 5 — Value Optimization
We can dynamically decide:
- Which model to use
- How much intelligence to apply
- How much to spend
- Whether to automate
- Whether to scale
- Whether to redesign the workflow
- Whether to stop
This is where the AI Value Plane becomes truly powerful.
It moves the organization from:
“What did our AI cost?”
to:
“What did our AI enable?”
and ultimately:
“Where should we deploy more or less AI intelligence to maximize enterprise value?”
From AI Consumption Management to AI Value Management
The enterprise AI journey should therefore evolve.
Stage 1 — AI Access
Who can use AI?
Stage 2 — AI Consumption Management
What are people using?
How much does it cost?
Stage 3 — AI Performance Management
Is the AI working?
Is the quality acceptable?
Stage 4 — AI Value Management
What work is AI enabling?
What productivity is being created?
What business outcomes are changing?
Stage 5 — AI Value Optimization
What is the optimal level of AI intelligence to deploy for each task, given its value, cost and risk?
That is where I believe enterprise AI management ultimately needs to go.
The Enterprise AI Architecture We Need
The emerging architecture therefore has three connected layers.
The AI Value Plane
Answers:
Why are we using AI and what value is it creating?
The AI Control Plane
Answers:
Who can use AI, which models can they use, and what are they consuming?
The Model Intelligence Layer
Answers:
What level of AI capability should be deployed for this task?
Together, these form a more complete enterprise AI operating model.
What Should Enterprise AI Budgets Actually Manage?
This framework also changes how AI budgets should be designed.
Different categories need different economic units.
Democratization
The economic unit is the:
Workforce / role / process
The focus is overall productivity and capacity.
Exploration
The economic unit is the:
Project / experiment
The focus is learning and future opportunity.
Production
The economic unit is the:
Application / business process
The focus is business outcome and unit economics.
This is more meaningful than applying one generic AI budget policy to every form of AI usage.
A per-user spending limit may be useful as a guardrail.
A project budget may be appropriate for exploration.
A cost-per-outcome target is much more meaningful for production.
The important distinction is:
A budget controls consumption. A Value Plane determines whether that consumption was worthwhile.
The AI Value Management Loop
Once the Control Plane and Value Plane are connected, AI management becomes a continuous loop.
The system continuously asks:
Are we using the right AI capability?
Are we achieving the required quality?
What does the outcome cost?
Is the value sufficient?
Should we scale, optimize or stop?
That is fundamentally different from simply monitoring tokens.
The Shift We Need to Make
The enterprise AI conversation often starts with:
“How much AI are we using?”
It should evolve to:
“What are we using AI for?”
Then:
“How well is it performing?”
Then:
“What outcome is it creating?”
And finally:
“Is the value created worth the AI investment and risk?”
That is the shift from consumption management to value management.
Token management is not going away.
FinOps is not going away.
Model gateways are not going away.
Budgets are not going away.
They are all necessary.
But they represent only one side of the equation.
The Control Plane tells us how AI is being consumed.
The Value Plane tells us why it is being consumed and whether it is creating value.
And the two need to be connected.
The Hypothesis
Perhaps the next stage of enterprise AI is not about giving everyone access to better models.
It is not even primarily about reducing the cost of tokens.
It is about building the organizational capability to answer a much harder question:
Where should we spend AI intelligence, at what level, for which work, and with what evidence that the resulting value justifies the cost and risk?
That is the hypothesis behind the AI Value Plane.
We have spent the first phase of enterprise AI building access.
Then we built consumption controls.
Now we are learning how to measure performance.
The next step is to connect all of these to outcomes and value.
Because ultimately, the enterprise does not need to maximize tokens.
It needs to maximize the value of the work that AI enables.
Don’t just build an AI Control Plane. Build the Value Plane that tells you where AI is worth deploying.
