In 2026, the problem is no longer justifying why to invest in AI. The problem is different: deciding which initiatives deserve capital, which ones only generate enthusiasm, and which ones should be stopped before they continue consuming executive budget.
The data points in that direction. Deloitte reports that AI investment kept growing through 2025, but returns remain difficult to capture for many organizations. IBM is even more direct: in its 2026 outlook, only around 25% of AI initiatives deliver the expected ROI.
That figure should change the tone of any investment conversation. Because it means the problem is not "whether AI works." The problem is whether the organization knows how to invest in it with discipline — and that is exactly where most companies still fall short.
In 2026, the bottleneck is no longer technical. It is now about capital allocation
Throughout 2024 and 2025, many companies funded AI as distributed innovation: area-by-area pilots, tactical copilot purchases, agent PoCs, RAG experiments, partial automations, SaaS tools with embedded AI.
That was useful for learning. But poor for building a portfolio. Because most of those decisions were made without a consistent portfolio logic: no clear economic baseline, no total cost of scaling, no homogeneous success criteria, no comparable data readiness across cases, no explicit rules for killing, pivoting or scaling.
The result today looks less like a strategy and more like a sum of initiatives. That is precisely what a CTO, CDO or CFO can no longer afford in 2026.
The most common mistake is evaluating AI as if all initiatives were equivalent. A company can push completely different things through the same budget pipeline: productivity copilots, document automation, predictive models, operational agents, data stack modernization, data governance for AI-readiness, retraining, observability, security and AI controls.
These initiatives do not compete in the same value category, nor do they share the same risk structure or the same capture horizon. That is why AI investment control cannot be based solely on impact promises. It must be grounded in a comparative framework.
What a serious committee should evaluate before approving an AI budget
By 2026, any AI initiative requesting meaningful budget should be able to answer, with evidence, at least six questions:
Not "what technology it implements." What real problem it moves. If the case cannot clearly express the current cost of the problem, its frequency and its impact on revenue, costs, risk or time, it is not yet ready for serious budget consideration.
Without a baseline, there is no ROI. There is only a narrative. The right question is not "will something improve?" It is: against what current state are we going to measure, and who validates that difference?
Many AI initiatives do not fail because of the model. They fail because the company tries to capture value on top of data that is unready, poorly governed or distributed in ways incompatible with the implementation. McKinsey consistently emphasizes that AI value correlates with integrated data management practices.
Not the cost of the pilot. The cost of turning the pilot into operational capability. This is where many organizations deceive themselves: they fund a small proof of concept, validate it technically, and then discover too late that the rollout requires integration, governance, support, observability, training, security and change management.
Not every project needs immediate returns. But every project needs an explicit horizon.
Every AI initiative should enter the portfolio with three possible paths defined from the outset: scale, pivot, kill. If there is no kill criterion, you do not have investment discipline. You have project attachment.
The wrong metric: measuring "pilot success" instead of "probability of capturing value"
One of the most common mistakes in 2026 is overvaluing the pilot. The pilot matters, but not in isolation. It matters as partial evidence of something far more relevant: the probability of capturing value at scale.
An AI pilot can look "successful" because it works technically, produces promising results or excites the end user — and still be a poor investment if the cost of scaling is excessive, the operational dependency is high, the required organizational change is unworkable, the necessary data is not sustainably available, or impact measurement at scale has not been resolved.
A mature organization should not only ask "did the pilot work?" It should ask: did the pilot reduce enough uncertainty to justify a larger investment?
That is a different standard entirely.
The new financial problem: AI spend without FinOps discipline
There is another structural shift in 2026: AI budgets no longer live in isolation. FinOps Foundation reports that in 2025 most teams were already managing AI spend, and in 2026 that rises to 98%, with AI cost management as a critical priority. Moreover, many organizations are under pressure to self-fund AI investment using optimization savings.
This has two concrete implications. First, the AI problem is also a variable cost problem: inference, embeddings, storage, vector search, retraining, observability, SaaS tooling and cloud consumption are no longer a "side budget." Second, AI investment control needs to speak the language of FinOps: it is not enough to say "this can deliver a lot of value" — you need to answer how much it will cost to operate, what variability the consumption has, what infrastructure commitments it generates, and how it is absorbed financially if it scales.
In other words: AI can no longer be funded like a lab. It must be evaluated as an operational capability.
A useful way to organize the portfolio: the four AI investment categories
Not all AI budgets should compete against each other. A practical way to organize the portfolio is to separate initiatives into four categories, which should not be evaluated by the same criteria:
- Productivity quick wins. Copilots, task automation, document summarization, internal support. Seek short-term returns, low friction and visible adoption.
- Functional value cases. AI applied to areas such as sales, risk, operations, supply chain or customer service. Seek measurable impact on business KPIs.
- Enabling foundations. Data governance, quality, metadata, architecture, security, evaluation, observability. They do not always show immediate direct ROI, but they raise the success rate of the entire portfolio.
- Strategic bets. Agents, proprietary capabilities, sectoral differentiation, high-impact automation, internal platforms. These are longer-horizon, higher-risk, higher-potential-advantage wagers.
The mistake is making an enabling foundation compete against a quick win using the same criteria. They should not be evaluated the same way.
What should be required before approving a pilot
A serious pilot does not enter the portfolio on sponsor enthusiasm alone. It should come with a mandatory minimum brief that includes, at a minimum:
- Defined business problem with a validated baseline
- Explicit value hypothesis
- Required data and readiness status
- Pilot cost and estimated scaling cost
- Executive owner and technical owner
- Success criterion, pivot criterion and kill criterion
- Evaluation window
If an initiative cannot fill out that brief with reasonable precision, it is not ready for budget. It is ready for discovery. And discovery is not the same as investment.
What reveals that an AI proposal is inflated
There are fairly clear signals that an initiative comes with more promise than discipline. Your committee should be alert when a proposal talks a lot about the model and little about the process, promises "transformation" without a baseline, uses projected savings without explaining the formula, avoids discussing data, does not estimate total operating cost, has no real business sponsor, depends on external talent without clear knowledge transfer, assumes a successful pilot equals viable scaling, or does not define what would happen if it fails.
In an executive committee, these proposals should not enter the capital pipeline as "advanced cases." They should be returned to maturation.
What changes when a company matures: from approving projects to managing a portfolio
The companies advancing in 2026 no longer treat AI as a sequence of isolated cases. They treat it as a portfolio with discipline. That means reviewing portfolio performance quarterly, reallocating budget based on evidence, cutting projects without drama, funding foundations even when they lack demo appeal, linking AI spending to optimization and operational capacity, separating exploration from scaling, and building organizational memory of what worked and what did not.
AI investment control is not a spreadsheet. It is a governance capability. And that cultural shift matters more than any scoring framework.
Yaripo's position
Most of the market still sells AI as a sequence of promising use cases. That is no longer enough. A serious organization needs a different language: capital discipline, comparable prioritization, scaling cost control, data readiness, kill/pivot/scale rules and the relationship between FinOps, Data Governance and value capture.
That is the Yaripo perspective: not viewing AI solely as a technological possibility, but as a portfolio of decisions where every dollar invested should have a logic of value, risk and sustainability.
In 2026, the problem is no longer securing AI budget. The problem is demonstrating that your organization knows how to allocate it better than everyone else.
Most companies no longer need to convince themselves that AI matters. They need something harder: a system for deciding what to fund, what to stop and what to scale — before enthusiasm turns into spend without return. That system is not built with slogans. It is built with discipline. And that is the difference between organizations that collect pilots and organizations that truly convert AI into competitive advantage.