Over 40% of agentic AI projects are expected to be canceled by 2027. Considering that approximately 72% of AI projects are currently in production, this estimate is causing justified concern among enterprises planning or actively implementing agentic AI adoption. What are the reasons for this discrepancy in expectations and outcomes? How can they be avoided? This article provides a comprehensive view of the state of agentic AI and how it fits into enterprise structure.
Implementing AI agents: What can go wrong?
Adopters who are canceling agentic AI projects name the following challenges as the causes of their projects’ failure:
- Escalating costs
Implementing and training artificial intelligence across the enterprise is already a costly investment. However, these expenses can increase exponentially with agentic AI if its problem-solving and decision-making routine isn’t fine-tuned and tailored to optimize resources instead of using them mindlessly. Psychological factors also play a role: as managers made active use of AI tools one of their main criteria during employee reviews, many employees started actively consuming AI tokens to demonstrate their productivity, which led to overall cost escalation.
Tokenomics is a very sensitive issue for many enterprises adopting AI. Token spend among employees alone adds a lot to a company’s bill. Stories of a single user costing an organization around$150,000a month don’t happen by accident. Instead, they are a direct result of how people use AI, how adopters expect people to use AI, and an unhealthy imbalance between expectations and productivity. Now, if we add multi-agent systems to the mix, the expenses skyrocket because most enterprise tokens are consumed by agents interacting with other agents.”
- Vague business value
Gartner analysts describe the majority of agentic AI projects as proof-of-concept, hype-driven campaigns, where adopters pursue the promise of cost reduction and efficiency gains rather than diving deep into their own specific business problems and building from that point. For that reason, only 11% of agentic AI projects were documented to successfully leave pilot purgatory and reach production.
Many agentic AI pilots start with a bang—but when it comes to scaling andactually interactingwith a specific enterprise context, they fizzle out. This happens because the issue is not with the model or algorithm, but with how AI impacts roles and processes within your organization. Some processes will be inevitably changed—some workflows may even disappear as redundant. But unlessit’sall mapped out in advance, your teams and employees will be the ones figuring everything out, which willultimately resultin frustration and lackluster outcomes
- Outdated governance
In 2025, only 8% of enterprises were reported to have a reliable and in-depth AI governance framework accompanying their AI adoption efforts. Meanwhile, around 75% of organizations are still in the early stages of outlining and implementing new governance for artificial intelligence. The latter implies that a large number of adopters apply rigid, traditional governance rules to agile and less predictable systems, which is bound to create multiple issues and increase risks along the way.
Traditional software governance is very linear and rule-based, much like non-AI software itself. When a system encounters an error, it doesn’t try to solve it on its own—it signals the user and ceases all activities until the error is fixed. AI, agentic AI, in particular, will try to solve the error on its own, without informing users. However, since it lacks critical thinking and human perspective, it’s more likely to generate more errors and quietly sabotage an entire workflow. The worst part is this issue will only become visible when critical processes start falling off.”
- Lack of evaluation
A less obvious but very impactful reason agentic AI projects rarely evolve past the experimentation stage is the way adopters assess outcomes—or, to be more precise, what they don’t assess. In many cases, the focus is on the demo, not on the data. Therefore, judging model performance gets complicated because adopters lose sight of what impacts the change, how the model makes decisions, and when reliable context starts drifting. This problem also connects to the issue of vague objectives and expectations—when an AI pilot isn’t developed to resolve a certain pain point, there are no clear KPIs for tracking progress.
These challenges are part of a larger and less obvious issue—lack of a structured, established operational vision for AI in the enterprise.
Operational models: Before and now
Defining the processes, employee roles, governance, and technology, the operational model is the blueprint of every enterprise. The clarity of decision-making, employee efficiency, and technology performance mainly depend on how well the operational model is built, detailed, and structured. Previously, the approach to operational models had been well-mapped and familiar to experienced enterprise leaders, allowing them to navigate change and adapt to new variables. However, with the emergence of artificial intelligence, an entirely new terrain appeared, putting existing operational models into perspective.
It’s worth noting that in this particular case, we’re talking about an operational model used in organizations working with automation tools and automated protocols—and how new trends have impacted Centers of Excellence (CoE) in terms of implementing technology and fostering its adoption across the organization.
Organizations leveraging RPA to optimize repetitive processes that don’t involve human intervention use a centralized operational model. It’s a classic CoE that covers ownership, building, and running RPA solutions. It also handles quality control and governance, ensuring consistent observability and performance monitoring.
The centralized operating model proved to be highly efficient for RPA-powered workflows. However, when applied to AI implementation, the results are less encouraging.
Essentially, sinceagentic AI is more about independent problem-solving rather than following a repetitive pattern and a set of rules, a centralizedCoEends up with a lot more work on its hands. Sinceit’sthe onebuilding, reviewing, running quality control, implementing, and deploying each new AI agent, each initiative takes a considerable amount of time. Due to this, an enterprise runs into a bottleneck instead ofmaintaininga desired pace. Naturally, organizations started working on ways to bypass this bottleneck—which resulted in what we call a federated operational model.
Within a federated operational model, the ownership doesn’t belong to CoE alone. Instead, every team and department can build, own, and run agents. The intent behind such a model is to ensure greater flexibility and power workflows with agents trained in the specific context and materials provided by teams. This intent works at the initial stage, when the model demonstrates a good understanding of its objectives, connects to the knowledge base, and works with templates or documents provided by the teams.
However, in a federated model, those who build and train agents also provide KPIs and impact the way they interact with information.
The outcome is that the agentic governance starts varying from department to department, and quality assessment gets messy because everyone uses their own unique KPIs to establish value.
A federated model is an attempt to enable easier and more natural interactions between human employees and agents, allowing employees tobenefitfrom agents faster and letting agents learn from interactions. However, without strict, uniform governance, things get chaotic fast: teamsdon’testablishguardrails or privacy policies, which, in turn, leads to major security issues and, once again, compromises productivity instead of increasing it.
In our practice, when enterprises try to switch from a more rule-based and rigid model to a more AI-friendly one, they end up with a federated model. The trap is that it shows some promisinginitialresults—and gets complicated later, when the first governance issuesemergeand ownership becomes vague.
This juxtaposition demonstrates how more structured, RPA-oriented approaches don’t work when it comes to implementing and scaling agentic AI across the enterprise, while newer and less polished practices fall off due to the lack of structure. That prompts a question: Is there an operational model that is the perfect fit for implementing agentic AI?
Before
Now
Centralized model
- CoE builds, reviews, and runs products, handles governance and quality control across organization
- Governance is covered and controlled
- Working with AI (especially multiple AI agents) leads to bottlenecks
Federated model
- Different teams build and run agents for their goals
- Agent governance and evaluation standards differ based on department
- More prone to governance challenges and data security issues
Here’s the biggest revelation: no, there isn’t. At least, there is no publicly known and commonly practiced operational model for agentic AI that enterprises can replicate. Why is that so? First of all, because of agentic AI agility. Agentic AI is what you make of it. It is tailored specifically to your organization, your industry, your workflows, and context. So, with so many variables, you will end up with an operating model that works for you specifically.
Similarly, successful adopters who cracked the code and gained value from their agentic AI implementation won’t share their full journey, since it’s their competitive secret—and every step they took is closely connected to their internal business processes. But that’s not all bad news: it just means that enterprises get to build their own agentic AI operational model from the ground up. If they succeed, this model will work for them for years to come.
Agentic AI: Building a path to production through gates and rails
Despite the potential for development and implementation chaos, the positive outcomes of agentic AI are more than wishful thinking. As successful adopters report exceeded ROI expectations and doubled productivity, there is a window of opportunity for organizations. All they need is the right approach to taking the leap.
So, how can adopters make sure their agentic AI projects grow beyond the demo phase and successfully become a dependable service?
The key is to make each step, each transition deliberate. There are many unknown variables when it comes to AI, so adopters must glean everypossible sliverof knowledge from every stage before theyproceedto the next one. This is particularly important when they have to say “no” to the AI pilot at the early phase. The exploration of the idea should be throwaway-friendly, allowing adopters to preserve necessary resources for other potential agentic AI candidates.
The best way to visualize an agentic AI adoption journey is to segment it into four gates:
- Experiment
The very first step is to prove the agentic AI project can potentially work. This should be done by indicating an existing and impactful pain point this project can resolve. Adopters should identify metrics for measuring progress and documenting changes, dissect agentic capabilities, and apply them to barriers caused by the pain point. This is also the stage when it’s safe to scrap the project in case it fails to prove its long-term usefulness. - Pilot
Once adopters prove the project can drive tangible results, the pilot stage can be initiated. Within this stage, the project is developed and designed to work with real users. This is also where the project’s boundaries are defined. Adopters should gain a full view of what goals, tasks, and features will be done, what will be left out, and how efficiently users interact with what’s left. - Production
If the pilot succeeds at providing value to real users and is fully scoped, it can transition to production, where it’s run for its corresponding service-level objectives. Clear ownership, reliable use, and successful coverage of established performance metrics usually signify the production’s success. - Operating and evolving
Once in production, the agentic AI product remains under constant monitoring to prevent context drift, introduce improvements or new iterations, or even retire the product in time.
Experiment to pilot
Pilot to production
Production to operating and evolving
Is this the right problem?
Can it be operated safely?
Can it be run an improved?
Goals
- To identify the right process for transformation
- To establish the autonomy level
- To name the owner who will be accountable after launch
- To agree upon a quality bar
Goals
- To ensure that every step can be traced and explained
- To establish clear evaluation gates for every change
- To set boundaries, such as cost caps per task and loop limits
- To build governance (limitations, observability, audit trails)
Goals
- To monitor the product’s functionality with the named owner
- To implement versioned changes
- To be able to do a rollback if necessary
- To introduce consistent improvements
- To easily retire the model once it’s done
These gates are as important as the transition between them. Before you advance, you must ask yourself: Who’s the owner right now? How is the project monitored? What is the rollback plan? What does it cost to run?
If you don’t have these answers, then you have to take a step back and see what you might have missed. Only after you can answer all the questions, can you proceed with confidence.
However, how should these questions be answered? The information is there, but it should be found and deciphered easily, without wasting time on finding owners, making sense of evaluation criteria, or traceability. For that reason, adopters must establish four gates that will connect them to necessary information at every layer.
Observability
- Tracing every LLM and tool call
- Documenting latency, token cost, and errors per agent and task
- Ensuring explainability of any failure
Evalutation
- Using evaluation datasets to gate every change
- Establishing clear release versioning for clean rollback
- Using production failures as feedback for new tests
Cost governance
- Implementing per-task unit economics
- Outlining gateway budget caps and iteration limits
- Preventing accidental token spend by making runaway loops error out, not bill out
Security and governance
- Scoping least-privilege tools
- Securing immutable audit trails
- Calibrating human-oversight gates to risk
Starting agentic AI roadmap right: What should adopters keep in mind?
Every step outlined above just scratches the surface of adopting agentic AI for enterprise. There is still much to uncover ingreat detail. However, right now we described a starting point and a structure that helps adopters navigate and stay on course without chaos. At the end of the day, successful AI implementationisn’tabout technology and algorithms, but about people, processes, and the mindset. So, what else should adopters do to ensurea viableand efficient strategy?
- Don't make agentic AI the endgame
It’s important to keep in mind that sometimes the outcome of the experiment phase can be the discovery that there is no need for an AI agent.
This is not an error or flaw—not all processes and workflows automatically benefit from agentic AI. Identifying such a process means that it can be improved through easier and more cost-effective technology or that it doesn’t need to be improved at all, and the bottleneck is elsewhere.
Adopting agentic AIdoesn’tgive anyone an instant edge. At the end of the day, every tool only works well whenit’sapplied to the right goal and with the right intention. So, if you realize that your workflowdoesn’tactually needAI, you are already in a better position than the competitors who waste resources and time on an agentic AI pilot for a similar workflow at their company.
- Lead with the problem
Markets-wide hype makes a poor advisor for a specific enterprise strategy. No matter how actively the innovation is promoted, it should have a very defined place in the organization. For that reason, adopters need to explore the most common and persistent pain points affecting their company or departments, and start their experiment with such questions “Can this particular problem be resolved through agentic AI? What outcomes will it deliver?” - Know where the company stands today in terms of AI
AI adoption takes time—and the waiting period depends on the enterprise’s previous experience with the technology. Adopters should have a clear picture of their organizational relationship with AI initiatives. If they have tried implementing AI POCs before, what was their visibility into the project? How did they work with observability or evaluation? What went wrong, and why? - Work with an external partner
Deep enterprise knowledge and stellar AI expertise can come in pairs. However, sometimes enterprise executives find themselves in need of agentic AI engineers to meet the project goals. They can either nurture ones internally, investing a considerable amount of time and resources, or they can partner with professional teams and combine their unique insights with vetted experience. The latter proved to be a viable strategy as 67% of companies who partnered with external teams successfully sent their projects into production. Among organizations who chose to build internally, only 33% made it to production.
If you are looking for a full view of your enterprise’s potential with agentic AI and a detailed pilot-to-dependable service roadmap, let’s chat! At Trinetix, we have been driving and fostering innovation for industry leaders and Fortune 500 enterprises for 15 years—and counting. With our diverse technology talent pool and domain expertise, we will help you nurture and drive the change your organization needs to break barriers and cross new thresholds.
