AI Agents in the Enterprise: Why the Hard Part Is Not the AI

The Request Sounds Familiar

We want an AI agent to handle customer service. Queries, transactions, account changes, the works."

It is a reasonable ask. Customers want instant responses. Businesses want lower operating costs. AI has clearly moved on from simple chatbots. Why not bring all of that together?

The answer is not "you can't" - it is "you have to be precise about what you are asking for." There is an enormous difference between an AI that answers questions and an AI that takes consequential actions on behalf of your customers. That distinction determines your architecture, your risk profile, and whether the project delivers the return you are expecting.

The Market Has Already Validated This

This is not experimental territory. Across industries, organisations are running production AI agents that handle real customer transactions at scale.

In financial services, banks are deploying agents that process account changes, initiate transfers, and handle dispute queries without human involvement - achieving resolution times that human agents cannot match. In telecoms, providers are automating plan changes, billing corrections, and technical fault diagnosis end-to-end. In retail, AI agents handle returns, reorders, and loyalty account management across millions of interactions per month. In travel, companies like Hopper have cut customer service costs by 65% while maintaining customer satisfaction scores equivalent to human agents.

The question your organisation should be asking is not whether this works. It is how to sequence the rollout so you capture the value without taking on avoidable risk.

Not All Operations Are Equal

This is where most AI projects go wrong. Organisations treat "customer service" as a single category and try to automate it all at once. In practice, every customer operation sits somewhere on a risk spectrum, and the right level of AI autonomy varies accordingly.

Low risk, high volume. Status checks, FAQ answers, policy lookups, document retrieval. These operations are read-only, they carry no financial or operational consequence if the response is slightly off, and they typically represent 40 to 50 percent of total inbound volume. A retailer answering "where is my order" at scale. A bank answering "what is my balance." A utility company answering "when is my next bill." Automating this tier alone is a substantial cost reduction with almost no downside risk.

Medium risk, structured process. Account updates, refund initiation, subscription changes, appointment scheduling. These involve some financial or operational consequence, but they can be designed so a customer confirms before anything is committed. The AI does the work of gathering information and preparing the action - the customer or an agent confirms before execution.

High risk, time-sensitive or irreversible. Cancellations with penalty windows, regulatory-bound transactions, actions that cannot be undone without significant cost. In financial services this might be an irreversible transfer. In insurance it might be a policy cancellation. In logistics it might be a customs declaration. These operations need careful guardrails and, in most cases, a human confirmation step - at least until the system has demonstrated consistent accuracy over time.

The organisations that succeed with AI agents are the ones that match the level of AI autonomy to the level of acceptable risk, and expand that autonomy gradually as evidence accumulates.

The Cost Reality

A common concern is that AI at scale becomes prohibitively expensive. The reality, when the system is designed properly, is quite different.

At five thousand customer interactions per day across a mix of operation types, a well-architected AI agent system typically costs in the region of two to three thousand pounds per month in AI model costs. That is likely less than the monthly cost of a single experienced customer service agent. The interactions the system handles often represent the workload of dozens of full-time equivalents.

The design choices that keep costs manageable are not complicated, but they do need to be built in from the start rather than retrofitted later.

Using a lightweight model to classify and route each incoming message before deciding whether to involve a more capable model for complex operations. A question about account balance does not need the same model as a complex dispute resolution. Routing correctly means simple queries are handled cheaply, and expensive model capacity is reserved for interactions that genuinely require it.

Caching the system's configuration and knowledge base content so it is not reprocessed from scratch on every interaction. For most enterprise deployments, this single design choice reduces input processing costs by up to 90 percent on the content that stays consistent across requests.

Keeping conversation context concise. Systems that naively carry the full history of every message and system response forward into each new interaction see costs grow exponentially as conversations get longer. A well-designed system summarises completed steps rather than repeating them verbatim.

None of this requires sacrificing capability. It requires treating cost as a design constraint from the beginning, not an afterthought.

What a Sensible Rollout Looks Like

Rather than automating all customer operations at once, the approach that consistently delivers value is phased delivery with clear measurement at each stage.

Phase one targets high-volume, low-risk operations. FAQ responses, status enquiries, document retrieval, basic account information. This phase gives the business a real containment rate - the percentage of inbound volume handled without human involvement - along with customer satisfaction data and confidence in the system before any consequential transactions are involved. Typically this phase alone reduces overall service volume by 40 to 50 percent.

Phase two introduces AI-assisted handling of transactional operations. The agent gathers the required information, validates it, and prepares the action - but a customer or agent confirms before anything is executed. This tests accuracy on write operations without exposing the business to the risk of unconfirmed errors at scale.

Phase three promotes the operations with the strongest accuracy record from phase two to full autonomy. Higher-risk operations remain in the assisted model indefinitely, or are supported by the AI without being fully delegated to it.

This sequencing allows a business to demonstrate measurable return in the first three months while building the evidence base needed to go further safely.

The Part That Surprises Most Organisations

The AI model is rarely where the difficulty lives.

Modern AI is genuinely capable of handling complex, multi-step customer interactions. What takes time and expertise is connecting it reliably to the systems it needs to act on: CRM platforms, billing systems, order management, identity verification, payment infrastructure. Each of those connections has its own data model, its own error states, and its own set of edge cases.

Payment processing is a clear example. An AI agent should never handle payment directly. It orchestrates the conversation up to the point of payment, then hands off to your existing payment infrastructure. The agent and the payment system are separate; the agent's role is coordination, not execution.

The same principle applies across any operation that involves a regulated system, a legacy platform, or a workflow with branching logic. The AI is the orchestration layer. Your existing systems remain the authoritative source of record and execution.

Organisations that underestimate this integration work tend to build impressive demos that cannot make it into production. Those that plan for it from the start ship systems that work reliably under real conditions.

How modernise.io Can Help

We have delivered production AI systems across industries that go beyond question-answering into genuine operational automation - systems that handle real transactions, integrate with existing infrastructure, and scale reliably under production load.

The work is not just about selecting the right AI model. It is about designing the risk tiers correctly, building the integrations that let the agent act on real systems, and sequencing the rollout so the business sees value quickly without taking on unnecessary risk.

If your organisation is evaluating whether AI agents can take on a meaningful share of your customer operations, the answer is almost certainly yes - but the design and sequencing decisions matter as much as the technology. If your organisation is ready to move from the question of whether to the question of how, that is exactly the conversation we are built for.