One conversation is dominating my feed and conversations with senior leaders – the RoI or lack thereof from the AI investments that they are making in their organizations. An article claimed that only one in five dollars spent on AI tokens can be tied to a quantified financial outcome. And this is a problem. Not just a financial problem, but a deeper systemic problem.
Read the data together and a sharper picture emerges:
- 12% of CEOs report both revenue AND cost benefits from AI (PwC, 2026).
- In the The state of AI in 2026: On the road to ROI report, McKinsey states that 80 percent of survey respondents say it (AI) has improved their productivity and half say it helps them make better decisions. Yet, only 37 percent of organizations report any positive EBIT contribution.
- Only 6% of companies derive measurable value from AI at all (HubSpot/Forbes, 2026)
This is a significant blind spot. Everyone is responding by investing in better observability tools, implementing financial controls or setting up governance guardrails, hoping that this will solve the problem.
AI Isn’t Like the Technologies That Came Before It
One thing that almost everyone forgets when thinking about the returns on investment is that the AI models and the impact they can have is fundamentally different from all the technologies that have preceded them in one very distinctive way – all the tech that came before was deterministic. You knew exactly what their capabilities were, whether an organization had successfully implemented the tech and if the adoption followed. It was also easier to benchmark and compare their usage and impact across organizations.
However, the current generation of AI models come with probabilistic capabilities. One can’t simply know the exact capabilities that they bring to the organization (as this depends on the context provided, the quality of the data used for training, the skills of the people using the tech – prompt engineering, the governance guardrails set up and many more). The same model deployed across departments within the same organization can have very different results, in terms of the value they bring in.
The number of variables that can derail the value one can derive from the deployment of AI models is one too many. And this is a risk that not many people fully understand or plan for. To truly derive meaningful or impactful value from these models, we need to address all these issues. We need to clean up our dataset that we will train the models on. We need to have a clear governance policy in place (what can you do with AI, who can access what data, clearly defined roles and responsibilities, people trained on how to interact and work with these models effectively, etc).
Cultural Readiness in Addition to Data Readiness:
On top of this, we need to address the cultural aspects of deploying AI models in our workflow. Everyone talks about data readiness, when it comes to deploying these AI models. People and cultural readiness is as much or even more important and gets limited or no attention.
The fear that our employees have about AI replacing them is real, whether voiced or not and needs to be addressed before large scale, enterprise level deployment of AI models.
Do we have a culture which expects and rewards the experimentation with technology, which also means that there will be failures and mistakes along the way.
Do we have the social and emotional bandwidth that allows our teams to reimagine how work gets done.
Do we have the autonomy to decide where we bring in AI models in our workflow?
Before we talk about how to fix this, we need to be precise about what we’re trying to measure, because most organizations are chasing the smaller of two very different prizes.
Most organisations are chasing the wrong Metric
Bain’s modelling of the $4.7 trillion in corporate profit at stake splits into two pools: $1.1 trillion from productivity gains, and $3.5 trillion from competitive shifts, the restructuring of who wins as AI changes what’s possible at sector level. This begs the question – do our teams even know what valuable work looks like?
We tend to measure things that are easier to measure and take it as a proxy for the more important measures but which are more difficult to measure. Its easier to measure how less people we need to do the exact same work, so we end up measuring them. It is easier to measure how much faster a part of the process has become, so we measure that. It is easier to measure the increased output from a team, so we measure it. But these improvements do not necessarily translate into high impact at an enterprise level.
What we forgot about Dr Eliyahu Goldratt’s Theory of Constraints
We are still stuck in optimizing locally, every process, every output and every individual team’s performance. This is usually counterproductive when it comes to optimizing for the entire organizations output. Dr Eliyahu Goldratt had long ago proved that improving every process locally, only makes the overall process globally worse.
He also gave us the solution to the problem.
Identify your biggest bottleneck right now. For some organizations, it could be in product development, for some it would be in sales, for others, it could be in manufacturing. Once we have identified the bottleneck, we need to get everything in that value chain work towards improving the capacity of this bottleneck, until it is no longer remains the bottleneck and something else becomes the bottleneck.
In his books, The Goal and Critical Chain, he shows how can one go about finding these bottlenecks in your organization. Simply put, it is the place where most work gets stuck before it can capture value in the marketplace.
For example, significant RoI shows up in whether your drug discovery cycle is faster than your competitor’s; whether your engineers can model 100 experiments where they previously ran five, which improves the success rate of the product; whether your customer insight compounds over time in ways that aren’t visible quarter to quarter. These are bottlenecks that are provide great opportunity to use AI models to address them.
These are proxies for competitive position that precede financial returns by months or years, and organisations that accept only financial proof dismiss them as “soft benefits” before the compounding has had time to work.
When we start using these powerful AI models to help us identify these bottlenecks or opportunities and remove them, we unlock significant value. This is simple. But it is not easy.
This requires judgement. This requires restraint. This requires us leaders, to know our business inside and out. This requires the organization to work as one unit. This requires a culture where everyone understands their role in the larger picture and is willing and able to work together as a team and optimizing the enterprise rather than their department.
Who is responsible for the RoI on the AI Investments – The Entire C-Suite
This has always been the job of a leader. What has changed now, is that with the introduction of AI models across the organization and the expectation from leaders to deploy and infuse AI in every process has super charged the behavior that optimizes the local efficiency, which is directly worsening the overall impact that we seek to make as an organization.
The organisations that are outperforming by infusing AI in their organizations got there by answering three important questions that most organisations skip:
- What organizational bottleneck or competitive capability are we trying to improve using AI and how do we measure progress? And do we have the patience to wait for the time it requires to show up?
- Who is accountable for the answer to the previous question, with decision authority, not just a reporting line?
- Do our employees experience, know and understand this? Do they feel that the deployment of AI models is giving more to them than it takes from them?
Without honest answers to all three, better dashboards will simply make the blind spot more visible without moving the needle.
So, instead of asking everyone to use AI models in everything that you do and optimize every little task, the better approach is to work together to identify the biggest bottleneck that prevents the organization achieve its stated objectives and use the power of these AI models to address them as one organization – striving for global optima instead of local optima.
I would love your thoughts on this..
PS: This post was originally published on SAP Community Network and has been republished here.
