AI Customer Service Automation: What Should You Automate?

AI can now do much more than answer scripted customer service questions. It can classify cases, retrieve customer information, process routine transactions and coordinate actions across systems. Pressure to use it is also growing. In a Gartner survey, 91% reported pressure from executive leadership to implement AI in 2026.

A refund that falls within a clear policy and spending limit is a straightforward automation candidate. A disputed charge, or a complaint from a customer in financial hardship, however, may need judgement. Treating both the same way can make service worse even as fewer interactions reach a human.

The key question is how much authority AI should have over each task, given its risks and likely effect on the customer. Routine work is often suitable for automation, while complex cases usually need human judgement and consequential decisions need particular care.

Traditional automation, such as rule-based bots, follows predefined rules and workflows. AI can interpret less structured inputs, retrieve relevant information and generate responses or recommendations. When connected to business systems and given the right permissions, it can also take actions.

Five questions can help determine whether a task is suitable for automation:

1.       Is it predictable? Does the task usually follow a recognisable pattern?

2.       Is the information available? Can the AI reliably access the data, knowledge and permissions it needs?

3.       Is the risk manageable? How severe would the harm be if the AI gets it wrong, and can the error be reversed?

4.       Does it improve the outcome? Is service faster, easier, cheaper, more accurate or more available, for the customer as well as the business?

5.       Can it fail safely? Can the system detect when it is out of its depth, stop before taking an inappropriate action, and escalate with context so the customer can continue without starting again?

A weak result on any test may indicate that AI should assist with the task rather than automate it fully.

Teams implementing AI in customer support often start with tasks such as these:

•         FAQs and information retrieval: strong where answers come from current, approved information.

•         Case classification: a good use case, though urgency and sentiment labels need monitoring for errors.

•         Conversation summaries: useful, but they can omit context, so people should be able to check and correct them.

•         Routine transactions: appointment changes, order tracking and subscription changes, provided identity, eligibility and limits are enforced.

•         Simple troubleshooting: appropriate when the process is well defined and the AI can verify the problem is resolved.

•         Post-interaction administration: CRM updates and follow-ups have consequences, so accuracy and permissions still matter.

Suitability depends on the interaction itself. A high-value customer may have a routine request, while a low-value one may have a consequential dispute. A human should stay in control when a case is:

•         Ambiguous or disputed: conflicting information, unclear policy or facts that need investigation.

•         Sensitive: distress, bereavement, financial hardship or vulnerability.

•         Consequential: a decision that materially affects finances, access to services or eligibility.

•         An exception: one that needs discretion or an authorised departure from normal rules.

AI can still gather information, acknowledge a complaint, prepare the case and summarise it before a trained agent steps in. Gartner's August 2026 survey found that 87% considered it essential for companies using generative AI for customer service to offer an option to reach a human agent.

The NIST AI Risk Management Framework offers a useful basis for this thinking. Its emphasis on reliability, safety, accountability and fairness supports a risk-based approach in which oversight reflects the potential consequences of an AI system's actions.

This is a practical framework rather than an industry standard. It describes five levels of delegated authority:

•         Level 1, human only: AI has little involvement.

•         Level 2, AI assists: AI offers suggestions, retrieval, summaries and recommendations.

•         Level 3, AI acts with approval: AI performs the task and a human confirms it.

•         Level 4, AI acts within guardrails: AI completes a task autonomously within predefined limits.

•         Level 5, AI coordinates a workflow: AI carries out multiple connected actions autonomously within an authorised scope, with defined escalation rules.

Level 5 combines multi-step execution with delegated authority. Workflow complexity and autonomy are separate considerations. An AI system that needs human approval for every action remains at Level 3, however many steps it handles. A system that autonomously coordinates multiple actions but escalates defined exceptions can operate at Level 5. Copilots and agents sit along this range. Copilots typically assist people, while agents can be given authority to act, though vendors do not always use these labels consistently. The more useful question is how much authority has been delegated.

Four actions turn the five tests into a decision:

1.       Establish the expected customer and business benefit.

2.       Assess readiness and risk, including data quality, process reliability and whether an error can be reversed.

3.       Set the level of autonomy and human oversight.

4.       Measure outcomes against the existing process.

A refund workflow shows how this works. A routine refund within policy and spending limits may be predictable, supported by reliable records and straightforward to reverse. If those conditions are met, it can be a Level 4 candidate, with appropriate limits and controls. A disputed refund involving suspected fraud, conflicting records or a vulnerable customer calls for a different approach: AI can prepare the case at Level 2, leaving the decision to an agent.

Sign-off typically benefits from more than one function. Customer service operations, risk or compliance, IT and the business owner each see different ways a task can fail. Tasks with potentially serious consequences need stronger oversight, even when their business value is high.

Alongside good data, the system needs consistent business rules, appropriate permissions and reliable connections to the systems where actions happen. Automating a bad process with AI doesn't fix the process. It scales the problem.

Guardrails should match the risk:

•         Low-risk information request: approved knowledge sources and a way to correct answers.

•         Routine transaction: authentication, eligibility checks, spending limits and confirmation.

•         Consequential or disputed decision: human approval, an audit trail and reversal where possible.

A model's stated confidence is not proof an answer is correct, so thresholds should not carry the load alone. And as AI does more across a customer journey, the decisions and handoffs between actions matter more. McKinsey's analysis of agentic customer experience argues for redesigned workflows, clear decision ownership and defined escalation rules as AI takes on more of the journey.

A conversation is resolved only when the customer's issue has been addressed, whether or not a human was involved. Customers may give up, contact the company again or switch channel, so an interaction that ends without a human can still be unresolved. Prioritise:

•        Verified resolution

•       Repeat contact and reopened cases

•        Customer effort and satisfaction

•        Cost per successfully resolved issue

•        Error rates, inappropriate actions and escalation quality

Compare results against the existing process and, where feasible, a control group. This helps distinguish genuine improvements from work simply moving elsewhere. A well-designed pilot should establish a baseline before deployment, with evaluation continuing after launch. NIST's draft TEVV-Athlon framework offers a useful starting point for designing ongoing assessments of AI system performance and impact.

Effective automation focuses on tasks where AI can improve customer outcomes and operational efficiency. Routine, predictable, low-risk work is usually the strongest candidate, while in complex or consequential cases AI supports human judgement. The best automation reduces unnecessary human work without creating unnecessary customer work.