The reality, roadblocks and risks behind the hype
The vision of an agentic future is exciting, but yet to materialise. What’s the hold up asks Alan Reed, Head of Technology Innovation at bet365?
For some time, the technology industry has talked as though agentic AI is about to reshape how work gets done. We have been promised autonomous workflows, agents that can connect systems, software factories that can turn intent into working code, and digital workers capable of taking on entire chunks of process.
It is an exciting vision. But in most organisations, that revolution has not arrived yet.
AI is everywhere, but much of it is still being used as an assistant. It helps people write, summarise, search, code, analyse and move faster. That is useful, but it is not revolutionary. It is still largely operating at the level of individual productivity.
The bigger shift comes when AI moves from helping someone complete a task to autonomously taking responsibility for part, or all, of a process. That is the agentic leap. It is also where things become much more complicated.
Dan Shapiro captured this progression well in his piece The Five Levels: from Spicy Autocomplete to the Dark Factory. The direction of travel is clear: from prompt-led assistance to fully autonomous SDLCs.
For me, the key takeaway from Dan’s article is this: the greater the autonomy, the more important it becomes to understand where control actually sits.
That is why the agentic revolution has not arrived at enterprise scale yet. Not because the technology is over-hyped or the opportunity is small, but because moving from assistance to autonomy is not a simple upgrade.
It fundamentally changes the risk profile, the control model, and the relationship between people, systems and decisions.
Why the revolution has not arrived yet
Image
Complex organisations do not run on information alone. They run on context, judgement, history, and trust. Agentic AI is therefore as much an operating model question as it is an economic one.
With traditional software, cost is tied to predictable usage patterns. With agentic AI, cost becomes unpredictable because an agent reasons, searches, calls tools, hits dead ends, and retries.
This also shifts where the operational effort sits. AI moves the human burden from writing code to checking it. Developers become verifiers, asking, “Is this what we meant, is it safe, does it work, and should it go live?”
The cost and workload do not necessarily disappear. The bottleneck simply moves. That is why the first benefit of agentic AI is often latency, not labour reduction. Work moves faster. Teams get time back. But the people who understand the process, the data and the risk become more important, not less.
The risk of competing answers and the unreliable narrator
As agents move closer to operational decision-making, consistency becomes critical. Even using the same model, outputs can vary if the prompt, query sequence, or context changes. This is a significant risk when agents support audited business processes or workflows.
The same issue applies to code. There is an assumption that agents can read a codebase and infer what a system is supposed to do. But a codebase is not a perfect record of intent. It shows what a system currently does, not what the business originally asked for, or which parts carry historical compromises and old bugs.
If an agent treats code as the absolute source of truth without surrounding business context, it is working with an unreliable narrator. Complexity is reproduced and automated.
The real danger is not a sci-fi AI takeover, but fragmentation. Different teams building different agents for different purposes, using mismatched models, prompts, data sources, and permissions, will collectively create an AI estate that nobody fully understands.
That is not transformation. It is fragmentation with a better interface.
Governance without constraint
Image
Teams need freedom to experiment, but live operational access should never be the default.
Live access changes the risk profile. An agent can update records, move statuses and place pressure on transactional systems never designed for automated interaction. If the task is analysis, summarisation or quality assurance, a curated snapshot may be safer than direct system access.
Observability is critical here. One interesting example comes from Lovable, where agents can escalate when they get stuck: first to another agent, then into a Slack channel where a human can help.
That shows both the power and the limitation of the model. The agent can keep working, try different routes and seek help. But someone still needs to understand why it is stuck and whether the path it wants to take is acceptable.
In one instance, an agent identified what it believed to be a bug in open-source software and moved towards patching it. Technically, that may have been logical. Operationally and ethically, it raised much bigger questions.
Agents are designed to make progress. But in business, progress is not always the same as the right decision. Sometimes the right answer is to stop, ask, or do nothing.
Why the revolution will happen
Despite these risks, the agentic revolution will happen because the prize is too big.
Most organisations are slowed down by the middle layer of work that happens between people, systems and stages of a process: handoffs, updates, reconciliations, reviews, quality assurance and impact analysis.
That is where agentic AI becomes genuinely interesting. The opportunity is not just individual efficiency. It is organisational effectiveness.
We are already seeing early architectural signals of this shift. Spotify has discussed agentic development with Anthropic, including the evolution of Backstage from a human-facing developer portal into an agent-first platform.
Anthropic’s Claude Code Routines point in the same direction, allowing saved prompts, repositories and connectors to run automatically on a schedule, via API call, or in response to GitHub events.
This moves AI from something a person actively prompts to something configured to keep working in the background. But the more useful these systems become, the more important it is to define their environment, permissions, triggers and escalation routes.
Give agents the right world to work in
Image
When thinking about the right environment for agents to work in, an analogy from elite sport comes to mind.
Modern teams have extraordinary amounts of real-time data, but the value is in filtering signals; knowing what to ignore, what to escalate, and what to act on. Instead of replacing entire human chains, the better approach is supporting the right points within it, whether that’s assisting analysis, checking work, or preparing the ground before a decision.
Agents are only as useful as the information they can safely access. If pointed at messy data, they will automate that mess. Rather than letting every agent roam across Jira, Git, Confluence, and SharePoint, organisations should create curated, domain-specific knowledge environments.
This gives agents the right context, in the right form, with the right constraints, and then governs how they use it.
This fundamentally changes the human element. The skills needed to manage live agents require a blend of localised business knowledge, AI understanding, and data literacy. The people organisations assume they can replace are often the very people they need most.
They are the ones who know the exceptions, understand where the data is weak, and recognize when something is technically correct but commercially wrong. Their knowledge is the ultimate harness around autonomy.
New tools need new thinking
So where is the agentic revolution?
It is coming. But it is not arriving as a single dramatic moment. It is arriving through decisions organisations are making right now about data, access, autonomy, governance, infrastructure and trust. That is why the current hype is both understandable and dangerous.
Understandable, because agentic AI really can change how work gets done; dangerous, because organisations can mistake activity for progress. More agents do not automatically mean more intelligence. More automation does not automatically mean more control. More access does not automatically mean better outcomes.
Agentic AI will be a major step forward. But it will not reward organisations that apply old thinking to new tools.
If the legacy mindset was to hard-code processes into systems, the new temptation is to push those processes into agents and assume the complexity has disappeared. It has not. It has just moved.
The agentic revolution will not be defined by how much autonomy organisations can create. It will be defined by how much autonomy they can safely understand, govern and trust.