The PEX Report 2025/26 shows agentic artificial intelligence (AI) investment intent racing ahead of governance and process visibility - and that gap is where most projects will quietly fail. An analysis of the structural weaknesses that stall deployments before production, and what the Klarna experience should have taught everyone.
There is a revealing asymmetry buried in the PEX Report 2025/26. Well over half of surveyed businesses - 59 percent - plan to invest in agentic AI or AI agents within the next 12 months, yet fewer than half currently have an AI governance policy of any kind, and less than a quarter use process intelligence to support their transformation work.
In other words, the majority of organizations preparing to hand decisions to autonomous systems have neither a framework for governing those decisions nor reliable visibility into the processes the systems will be asked to run.
That gap between appetite and infrastructure is where projects go to die, and the analyst community has already priced it in: Gartner predicts that more than 40 percent of agentic AI projects will be canceled by the end of 2027 as costs escalate, business value fails to materialize and risk controls prove inadequate.
For anyone who lived through the RPA hype cycle, the diagnosis will feel familiar - the binding constraint is rarely the technology itself, but the organization asked to absorb it.
Here are five gaps that show up most consistently, and, in one well-documented case, publicly.
1. Your processes aren't fit for autonomy
An agent is only as good as the process it has been asked to run, and in most organizations that process exists as a blend of tribal knowledge, outdated standard operating procedures and workarounds that nobody has written down because everybody already knows them. This is the process excellence community's oldest lesson restated for a new technology - automating a broken process simply produces a faster broken process - but agency raises the stakes considerably, because an agent doesn't just execute a flawed workflow at speed. It makes autonomous decisions inside that workflow, at scale, without a human pausing to ask whether the exception it just encountered is evidence that the process itself is wrong.
The readiness test here is unglamorous: process mining coverage, visibility of exception rates, and documentation that reflects how work actually happens rather than how it was designed several reorganizations ago. On that measure the industry has some distance to travel - the PEX Report 2025/26 found that only 23 percent of surveyed firms use process intelligence to support business transformation, which means roughly three-quarters of the organizations queuing up to deploy agents cannot yet see the processes they intend to hand over. Organizations that skipped this foundational work could still muddle through with RPA, because bots failed loudly and predictably when reality diverged from the script. Agents, by contrast, tend to fail quietly and plausibly - which is a far more dangerous combination.
2. Your data can't support agent decision-making
Most enterprise data was structured for humans to query, not for autonomous systems to act upon, and the difference matters more than it first appears. Customer records fragmented across five systems, field definitions that drift between regions, and unstructured documents with no retrieval layer are all familiar inconveniences for an analyst, who can spot the inconsistency and route around it. For an agent, those same flaws are decision inputs.
The compounding effect is what makes this a readiness question rather than a hygiene one. A dashboard built on messy data produces a misleading chart that a sceptical human might interrogate; an agent built on the same data produces an action - a reorder, a refund, an escalation - and then another, and another, each one inheriting the original error and extending its consequences.
An honest audit of whether your data architecture can answer, in real time and with appropriate access controls, the questions your agents will need to ask is a prerequisite, not an optimization. If your last data governance initiative stalled in a steering committee, you already have your answer.
3. You have no accountability model for autonomous decisions
There is a question most leadership teams cannot yet answer, and it will define the next two years of deployment: when an agent makes a bad decision, who owns it?
The governance frameworks in place at most organizations were built for systems that recommend, not systems that act, and the difference is structural rather than semantic - where frameworks exist at all, that is, given the PEX Report 2025/26 finding that fewer than half of businesses currently have an AI governance policy. There is typically no defined escalation path for when an agent's confidence drops, no audit trail standard for autonomous actions, and no clarity on whether the process owner, the IT function or the vendor is accountable when something goes wrong.
The logic for fixing this before deployment rather than after is straightforward: an organization that cannot assign accountability for an autonomous decision has no basis on which to grant autonomy in the first place. Treating governance as an afterthought is, in effect, drafting the business case for your own project cancellation.
Case study: Klarna and the failure the dashboard didn't show
No company illustrates the readiness question more publicly than Klarna. In February 2024, the Swedish fintech announced that its OpenAI-powered assistant had handled 2.3 million conversations in its first month - the equivalent workload of roughly 700 human agents - while cutting average resolution time from 11 minutes to under two and projecting around $40 million in annual savings. It was, for a period, the most-cited AI deployment in customer service, and CEO Sebastian Siemiatkowski leaned into the framing enthusiastically.
By May 2025, the same CEO was telling Bloomberg that the company had gone too far, conceding that an overemphasis on cost had produced lower-quality service, and committing to a model in which customers would always be able to reach a human. Klarna began hiring human agents again under a flexible, remote "Uber-style" model, while maintaining - with some justification - that it was not abandoning AI but rebalancing toward a hybrid architecture.
The instructive detail is where the system broke. On simple, high-volume queries such as order status and payment schedules, the AI performed at or near human level; on complex disputes, fraud claims and emotionally charged interactions, resolution quality deteriorated in ways the headline metrics initially concealed. Klarna's failure was not technological but organizational: the company measured averages when the risk lived in the distribution, scoped autonomy by volume rather than by complexity, and had not modeled the cost of unwinding the strategy if quality slipped. Every one of those is a readiness failure that predates the technology - and every one was avoidable with the disciplines described above.
4. You're retrofitting agents onto legacy workflows
The most common failure mode in early deployments is organizational-technical rather than purely technical: integrating agents into legacy systems is complex, disrupts existing workflows and often demands costly modification, which is why rethinking workflows from the ground up is frequently the better path than bolting agents onto processes designed around human handoffs.
Transformation leaders should recognize the pattern, because it is the same one that defined the last decade. The organizations that "did digital transformation" by digitizing paper forms are now "doing agentic AI" by inserting an agent into a workflow whose approval chains, batch cycles and system boundaries all assume a human is performing the work - with the predictable result that the agent inherits every constraint and delivers a fraction of the value. If your agentic AI roadmap amounts to a list of existing tasks with "agent" written beside them, what you have is an automation backlog in costume.
5. You can't distinguish real capability from agent washing
The vendor landscape is actively working against your due diligence. Of the thousands of vendors claiming agentic capabilities, Gartner estimates that only around 130 are genuinely agentic; the remainder are rebranding existing chatbots, assistants and RPA tools in a practice its analysts have labeled "agent washing." An organization that cannot articulate the difference for itself - autonomous reasoning toward a goal, orchestration of tools and systems, persistent context across a task - cannot meaningfully evaluate vendors, scope pilots or set realistic expectations with its board. It will buy an assistant, call it an agent, and report disappointing results that reveal nothing about the technology and everything about the procurement.
There is a related discipline in value definition, because many use cases positioned as agentic today do not require agency at all - a deterministic workflow would be cheaper, faster and considerably easier to govern. Genuine readiness includes the organizational confidence to say so before the pilot begins, rather than in the post-mortem.
What unites these five gaps is that none of them concerns model capability, and that is precisely the point. The models will continue to improve regardless of anything your organization does, which means the differentiator over the next two years will be the least fashionable work in the transformation portfolio: documenting processes as they actually run, repairing data foundations, writing accountability frameworks before they are needed, and pressure-testing vendor claims against a definition of agency you can defend.
Klarna paid publicly for lessons that are now available to everyone else at no cost - and the organizations most likely to succeed with agentic AI are the ones willing to teach them second-hand.