Pre-built AI agents are quickly becoming the next commercial layer of enterprise software. Vendors are packaging conversational agents, task-specific bots, and workflow assistants as ready-to-deploy components, promising faster rollout and immediate productivity gains.
On the surface, this looks like progress. Organisations no longer need to assemble models, prompts, and interfaces from scratch. They can switch on agents for customer support, internal service desks, sales enablement, or operations with minimal configuration.
But as these agents move from demos into live environments, a familiar pattern is emerging. The technology works. The agents perform their narrow tasks well. Yet once they touch real systems, real data, and real users, friction appears.
Go deeper with GlobalData
Access deeper industry intelligence
Experience unmatched clarity with a single platform that combines unique data, AI, and human expertise.
The problem is not intelligence. It is integration.
From capability to connection
Pre-built agents are optimised for capability. They are trained to answer questions, trigger actions, or guide users through processes. What they are not optimised for is the complexity of enterprise environments.
Most organisations operate across fragmented stacks: legacy systems, cloud platforms, third-party tools, bespoke integrations, and multiple identity layers. An agent that performs well in isolation must navigate this complexity reliably and safely.
This is where many deployments stall. The agent may understand what needs to be done, but the surrounding systems are not designed to support autonomous or semi-autonomous execution.
US Tariffs are shifting - will you react or anticipate?
Don’t let policy changes catch you off guard. Stay proactive with real-time data and expert analysis.
By GlobalDataThe result is a growing gap between what agents can do in theory and what they can do in practice.
The four integration bottlenecks
Across early deployments, four integration bottlenecks are appearing consistently.
1. Data access and boundaries
Agents depend on timely, accurate data. In reality, data is distributed across systems with inconsistent schemas, access rules, and update cycles. Without careful design, agents either see too little to be useful or too much to be safe.
2. Identity and permissions
Agents act on behalf of users or systems, but enterprise identity frameworks were not built for non-human actors. Deciding what an agent is allowed to see, change, or initiate requires more than copying a user role. This often becomes the first hard stop in deployment.
3. Workflow orchestration
Triggering a single action is easy. Managing a sequence of actions across multiple systems, with fallbacks and exceptions, is not. Many agents end up constrained to advisory roles because orchestration layers are missing or fragile.
4. Monitoring and correction
Once an agent is live, teams need to know when it fails silently, produces degraded output, or requires human correction. Without clear monitoring, problems surface only after users notice inconsistent results.
None of these issues are new. What is new is the speed at which agent deployments are exposing them.
Why pre-built does not mean plug-and-play
The appeal of pre-built agents lies in speed. Organisations want results without long build cycles. But speed at the surface often hides complexity underneath.
Pre-built agents assume a level of standardisation that rarely exists. They expect clean APIs, stable data models, and predictable workflows. In many enterprises, those conditions are aspirational rather than real.
This creates a mismatch between vendor expectations and operational reality. The agent functions as designed, but the environment does not.
Technology teams then face a choice. Either they constrain the agent’s scope so tightly that its impact is limited, or they invest in integration work that was not originally planned.
Integration becomes the real product
As more agents enter production, integration itself becomes the differentiator. Organisations that treat integration as a first-class capability move faster and with fewer surprises.
This means investing in:
- consistent data access patterns rather than one-off connectors
- clear service boundaries that agents can rely on
- identity models that accommodate machine actors
- observability that covers agent behaviour, not just system uptime
In these environments, agents can evolve from assistants into reliable components of daily operations.
In less prepared environments, agents remain novelties. They work well in controlled scenarios but struggle at scale.
The shift technology leaders must make
For technology leaders, the shift is subtle but important. The question is no longer “Which agents should we deploy?” but “What must be in place for agents to operate safely and consistently?”
That reframing changes priorities. It moves attention away from feature comparison and towards foundational capability.
Teams that succeed with agents tend to ask practical questions early:
- How does this agent authenticate itself across systems?
- What happens when upstream data is delayed or incomplete?
- Where do we see and measure agent-initiated actions?
- How do humans intervene when outputs degrade?
These questions are not about AI performance. They are about system design.
A predictable next phase
The next phase of the agent cycle is likely to be consolidation around integration frameworks. Just as early cloud adoption exposed gaps in identity, monitoring, and cost control, agent adoption is exposing similar gaps in orchestration and oversight.
Vendors will respond with better tooling. Platforms will mature. Standards will emerge.
In the meantime, organisations that treat agent deployment as a technology integration exercise rather than a feature rollout will move ahead.
What this means for technology strategy
Pre-built agents are not a shortcut around technical discipline. They accelerate value only when the underlying environment is ready.
For many organisations, the real work sits below the agent layer: simplifying data access, clarifying system ownership, and strengthening integration patterns.
Those investments are less visible than launching new agents, but they determine whether agents become dependable contributors or ongoing sources of friction.
In that sense, pre-built agents are not just new tools. They are stress tests. They reveal how well modern technology stacks are designed to support autonomous action.
The organisations that pass that test will not be the ones with the most agents, but the ones with the most coherent integration underneath.
Gulzar Singh, Senior Fellow – Banking & Technology; CEO, Phoenix Empire Ltd
