High support costs killing your margins? AI agents can fix that, without the risks associated with unpredictable AI chatbots.

Top companies are deploying AI agents that strictly follow company workflows and guidelines, like digital assets that never go off-script.

For instance, Bilt Rewards saves hundreds of thousands of dollars monthly at a 70% resolution rate. It's results like these that drive Gartner’s prediction that AI will handle 80% of interactions by 2029 – cutting service costs by 30%.

If that’s what you want to do too, this guide provides the framework to understand how you can choose the best AI approach, manage the risks, and build your own playbook that demonstrates immediate value.

What is AI chatbot customer service?

AI chatbot customer service is software that uses generative AI to understand and handle customer inquiries through automated conversations. This allows the system to provide immediate answers to common questions, which reduces response times and frees up human agents to focus on more complex or sensitive issues.

It’s crucial to understand that these modern AI agents are far more advanced than simple question-and-answer bots. Think of a first-generation chatbot as an interactive FAQ page that could only point to information, whereas a modern AI agent can converse with customers and follow specific instructions to complete a given set of tasks. They operate with clear workflows, guardrails to stay on-brand, and defined escalation paths to a human when a problem is outside their scope.

Their capabilities are built on a few key distinctions:

• Action-oriented platforms. These integrate directly with your internal systems (like your CRM or order database) to perform tasks such as processing a refund, updating an account, or scheduling a delivery. This is different from purely conversational platforms that can only provide information.‍
• A unified knowledge graph. Instead of static training documents, AI agents pull real-time data from your help center, APIs, and databases, ensuring every answer is accurate and current.
• Secure personalization. Agents can use data from past interactions to personalize the current conversation without storing sensitive personal data, thereby ensuring customer privacy and compliance with regulations such as GDPR.
• Omnichannel and multilingual support. AI agents can be deployed across all touchpoints (including chat, email, and voice), providing consistent service in multiple languages to support a global customer base.

This integrated approach allows companies to be far more efficient at handling queries while keeping customer satisfaction high, and drastically lowering costs.

What companies use AI agents for customer service?

AI-powered customer service isn't limited to one type of business. Companies across all verticals – including finance, travel, hospitality, retail, and technology – are using AI agents to manage customer interactions, improve efficiency, and provide better experiences.  

Here are a few examples of how different companies effectively use Decagon's AI agents.

Fintech: Chime 

Chime is a fintech leader redefining banking with fee-free services to help millions of members build a healthier financial future. Facing rapid growth and increased support volume, Chime needed a way to scale its customer service efficiently while maintaining its member-first values. Decagon provided a unified AI automation platform for both chat and voice that resolved member issues end-to-end, delivering greater efficiency and a stronger customer experience.

“Collaborating with Decagon allowed us to scale member support efficiently while staying true to Chime’s member-first values. Their partnership supported our commitment to operational excellence and continuous improvement.” – Janelle Sallenave, Chief Experience Officer at Chime

• The result: Chime reduced customer support costs by 60% while doubling member satisfaction scores.
• The impact: The AI-powered platform achieved a 70% resolution rate across both chat and voice channels.

Retail: Curology

Curology provides personalized, prescription skincare treatments to customers. Their previous manual support system was unsustainable, with chat handling only 5% of tickets and every inquiry requiring human intervention. Decagon centralized all support channels into a single AI-led system that automates tasks like shipping updates, order replacements, and subscription management.

• The result: Curology reduced its customer support operational costs by 65%.
• The impact: Their chat channel, now powered by the AI agent, fields 80% of all support tickets, a massive increase from the previous 5%.

SaaS: Notion

Notion is a productivity platform that combines notes, documents, and project management in one workspace. Handling one million customer inquiries annually, they needed an AI solution that could deliver high-quality user experiences while integrating deeply with their existing systems. Decagon provided native generative AI capabilities that immediately improved efficiency and freed their human agents to focus on upskilling and specialization.

"Decagon stood out across the board – not just in these core areas, but also through their close collaboration with our technical team and their ability to meet our stringent security and compliance standards." – Emma Auscher, Global Head of Customer Experience

• The result: Ticket resolution time improved by up to 34%.
• The impact: The AI agent successfully handles inquiries with an average "ask for human" rate of just 3.4%.

Which AI solution is best for your customer service?

Approach 1: Traditional AI platforms

These are the traditional chatbots that many people are familiar with. They can be programmed to answer common questions and explain company policies from a knowledge base. However, their primary limitation is that they can talk about a policy but cannot execute an action. For example, a conversational bot can tell a customer about your 30-day return policy, but it cannot actually process the return because it isn’t integrated into your backend order management system.

Approach 2: Bolt-on AI features

These tools are legacy customer service solutions with bolted-on AI solutions, often found as an add-on to existing helpdesk systems. They are designed to assist human agents and can suggest replies, summarize long ticket histories, or classify incoming requests. While helpful for improving agent efficiency on a per-ticket basis, they don’t fundamentally reduce the number of tickets that require human intervention.

Approach 3: Developer-first AI platforms

At the high end of the market are powerful, developer-centric platforms or fully custom in-house builds. These solutions are highly capable but come with a significant drawback: they require significant investment of engineering resources. 

Additionally, the core logic is written in a complex configuration language, which means every time your business needs to update a policy – even something as simple as extending a return window – you have to file a ticket with the engineering team and wait for them to implement the change. 

Approach 4: Business-user-controlled, Gen AI platforms

A new generation of platforms solves this bottleneck by utilizing natural language to shape AI agent behavior. These systems are designed for non-technical teams who are experts in customer service.

At Decagon, this is made possible by Agent Operating Procedures (AOPs). AOPs are workflows written in everyday language that are automatically compiled into code that the AI agent executes.

The CX team writes a simple, clear rules: "If an order was placed within the last 30 days and the item is unused, process a refund." The engineering team maintains control over guardrails and establishes secure connections to the necessary systems in the background. This way, both teams handle the parts of the process in which they have expertise.

With Decagon’s AI agents, you get a system with clear advantages:

• Speed. Implementation takes weeks, not months.
• Transparency. You have complete clarity about the instructions being given you your AI agents. Nothing is buried in layers of code.
• Quality at scale. The efficiency of AI agents keeps pace with even an exponential increase in support requests, due to faster iteration and the ability to have human agents train the AI using their years of support experience.

This approach delivers proven results, with Decagon clients improving resolution rates, reducing support costs, and increasing customer satisfaction.

Why you should focus on resolution rates

When comparing platforms, the sticker price is less important than the resolution rate. What matters most is how many cases the AI closes on its own, because even tiny percentage gains add up fast at scale.

For instance, say you have 100k support cases:

• If 50% are resolved → 50,000 still go to your team.
• If 55% are resolved → 45,000 go to your team (5,000 fewer).
• If 70% are resolved → you have only 30,000 going to your team (20,000 fewer).

As a general rule of thumb, with 100,000 cases, every +1% in resolution rate is ~1,000 fewer tickets for human agents. At ~10 minutes per case, that’s ~333 hours, or over 8 work weeks, given back to the team. Because the gains apply to your whole volume, small percentage lifts translate into big chunks of work removed.

Prioritize platforms that consistently push this number up.

Finding your automation opportunities

To identify the best starting points, you can use a simple framework to help you classify customer inquiries based on two factors:

• Complexity: How many steps, variables, or pieces of information are needed to resolve the issue?
• Risk: What is the financial, legal, or reputational impact on your business if the interaction is handled incorrectly?

Using this matrix, you can sort all of your customer inquiries into three main categories:

Category 1 – Low-medium complexity, low-risk: Automate fully

These are the repetitive, predictable questions that make up a large volume of your support tickets but carry very little risk. Automating them provides the biggest and fastest return on investment.

Common examples include:

• Checking the status of a shipment.
• Updating account information, like a shipping address.
• Requesting a refund for a recent, simple purchase.
• Resetting passwords for clients.

In some instances, high-complexity issues can be fully automated if there's a clear-cut procedure to follow, like multi-step warranty claims with defined criteria.

Category 2 – High complexity, high-risk: Always escalate to a human

At the opposite end of the spectrum are inquiries that should always be handled by a person. While these situations are low in volume, they are often emotionally charged, legally sensitive, or technically complex, and the cost of an error is simply too high. Forcing a customer through AI in these moments leads to frustration and can cause lasting damage to your brand.

Examples of these situations include:

• A customer threatening legal action.
• A user reporting a security breach or compromised account.
• Handling queries that fall under strict regulatory guidelines, such as certain medical-related questions.
• A customer expressing extreme frustration or anger.

Category 3 – The mixed quadrants: Use a hybrid approach

Many inquiries fall somewhere in the middle – they might be complex but carry low risk, or be simple but carry high risk. For these situations, a hybrid approach where the AI and a human agent work together is most effective.

The AI agent can act as a data collector, handling the initial steps of the interaction. It can identify the customer, pull up their account history, and ask clarifying questions to understand the issue. The AI agent can then hand off the entire conversation, with all its context, to the right human expert for the final resolution.

Using technology to manage the matrix

You don't need to manage this matrix with manual oversight alone. Modern AI platforms enable you to integrate rules and guardrails directly into the system, allowing for automatic escalation of conversations based on predefined triggers or keywords. 

Decagon takes that even further. Our agents do not need to rely just on keywords to assess the severity of the situation, but can understand the context and take the necessary next steps. 

For instance, say a customer writes, "My account is locked and the payment for my son's school trip is due today. Can you please help me?" Instead of triggering a standard "locked account" workflow, our AI understands the emotional and time-sensitive context of the request. It immediately escalates the ticket to a high-priority queue for urgent human intervention, ensuring a family doesn't miss an important event.

Additionally, look for platforms with a monitoring feature that provides a layer of human-in-the-loop oversight. This allows managers to review conversations in real-time to ensure the AI is behaving as expected and that escalations are working correctly.

Regardless of the query type, building customer trust is the primary goal. You can maintain it by following three universal principles:

• Bot disclosure. Always be transparent with customers when they are interacting with AI.
• Clear escalation paths. Make it easy and obvious for a user to ask for a human agent at any point in the conversation.
• Cite your sources. When an AI provides a specific piece of information, like a policy detail, it should be able to cite the official source, preventing any kind of confusion.

The legal case that changed AI chatbot rules

The 2024 ruling in Moffatt v. Air Canada is a critical lesson in AI liability. After Air Canada’s chatbot incorrectly promised a customer a retroactive bereavement fare discount, the airline was held responsible for the misleading information and forced to honor it. This set a clear precedent: Your company is liable for what your AI says.

The issue here is that the chatbot operated without essential guardrails, as it was not constrained to official policies, failed to cite its sources, and lacked clear audit trails. 

These are entirely manageable risks. 

A platform using Decagon’s Agent Operating Procedures (AOPs) would have forced the bot to provide only the approved policy text. Tools like Watchtower add another layer of safety by allowing teams to monitor conversations and flag errors for immediate review.

To comply and operate safely, businesses need critical safeguards like decision logs, version control for the AI's rules, and automatic escalation paths to human agents for high-risk queries. This is exactly what enterprises get when they partner with Decagon.

Next steps: Implementation and calculating ROI

Once you have a strategy for automation, the final steps are to map out the implementation, calculate the return on investment, and prepare your team for the change. A successful deployment requires clear metrics, a realistic timeline, and strong internal adoption.

Implementation timeline for enterprises

With Decagon, the implementation process can be surprisingly fast. An enterprise can go from initial discovery to a full deployment in about two months, not two quarters.

Here is a sample timeline:

• Week 1: Run technical discovery, set up the sandbox and communication channels, and document existing support workflows.
• Week 2: Hold the project kick-off, begin drafting Agent Operating Procedures and core integrations, and define success criteria and pilot use cases.
• Week 3: Configure the agent, routing rules, and escalation paths, and start internal testing of core workflows.
• Week 4: Test integrations and edge cases in parallel and refine AOPs based on early results.
• Week 5: Merge testing insights, finalize configurations and escalation protocols, complete compliance documentation, and train the team on monitoring.
• Week 6: Launch with a controlled rollout, monitor performance in real time, make rapid adjustments, and scale to full deployment.‍
• Post-launch: Monitor daily, refine AOPs weekly, expand to additional workflows, and drive continuous improvements from conversation data.

Measuring the ROI of AI agents

The return on investment from an AI agent is highly measurable and goes far beyond simple cost savings. To get a complete picture of its impact, you should track a combination of primary and secondary key performance indicators (KPIs).

Primary Metrics:

• Resolution Rate: Percentage of inquiries fully resolved without human intervention.
• Customer Satisfaction (CSAT): Satisfaction scores for both automated and escalated interactions.
• Cost per Resolution: Total platform cost divided by automated resolutions.
• Escalation Rate: Percentage of conversations handed off to human agents.
• Average Handle Time (AHT): Time for human agents to resolve escalated conversations.

Secondary Metrics:

• Agent Satisfaction: Human agent satisfaction with AI support.
• Knowledge Gap Identification: Insights from conversation logs about system gaps.
• Customer Effort Score (CES): Measure of customer effort required.
• First Contact Resolution (FCR): Rate of issues resolved on first contact.

Success isn’t a one-time setup. It requires a cycle of continuous improvement, including weekly refinements of the agent's rules and procedures, monthly performance reviews against your KPIs, and quarterly strategic decisions about which new workflows to automate.

Change management for internal adoption

Introducing a powerful automation tool can create anxiety for your customer service team if they view it as a threat rather than a tool. A thoughtful change management strategy is essential for getting your agents on board and ensuring a successful adoption.

The key is to demonstrate how the AI agent is a partner that is there to make their jobs better, not to replace them. Here is a simple strategy for encouraging internal adoption:

1. Start with champions. Select your most tech-savvy and respected agents for the pilot program. Their endorsement will drive team-wide adoption more effectively than any top-down mandate.
2. Show immediate value. Position AI as handling repetitive, mundane tasks so agents can focus on complex, engaging challenges that utilize their expertise.
3. Transform roles. Evolve agents from ticket handlers to higher-value roles, such as problem solvers and product experts.

Execute this approach correctly, and you'll achieve both operational efficiency gains and a more engaged, skilled workforce, creating a win-win scenario for your business and your team. 

Here’s a review from one of our clients that bears testimony to this approach:

“Decagon has proven that with the right technology, AI can help scale a customer service team and keep CSAT high. They truly allow you to create a unique user experience, where you're adding in AI, but not losing the personalized touch of a human agent. Their team has become an extension of us, and we are very grateful for their flexibility and partnership.” – Daniel B., Source: G2 reviews

Your competitive advantage starts now

The debate about whether companies should adopt AI for customer service is over. The real question is how to do it correctly and how quickly you can get started. AI agents are rapidly becoming essential business infrastructure, as fundamental as a CRM or payment processor. 

This simple shift, from typical conversational bots to action-oriented AI agents, can transform your support department from a cost center into a strategic asset. When your AI can execute your business policies, including higher-value tasks like processing refunds, updating accounts, and managing shipments, it significantly improves the customer experience and provides invaluable data insights back to your product and operations teams.

So, what are the next steps? 

1. Begin by auditing your most common customer inquiries.
2. Build a coalition of support from leadership.
3. Schedule demonstrations with platforms that will empower your business teams.

Seeing this technology applied to your specific use cases is the fastest way to understand its potential. If you're ready to see how a policy-driven AI agent can transform your customer service operations, you can get a demo of Decagon today.