View Demo
How to get an AI Agent to handle multi-step, complex queries
Practical guidance for getting AI Agents to take on more meaningful, complex work.
Many support teams start out using AI to handle straightforward work.
Like basic FAQs or password resets. These types of queries are easy to automate, feel safe, and can reduce some support costs, but they’re not going to move the needle.
The real payoff comes when you start automating the complex, high-effort, multi-step tasks that require things like judgment, sequencing, and system interaction. The ones that tie up your team and slow down resolution.
That’s where AI Agents deliver asymmetric value. The more you trust your AI Agent with meaningful work, the more your return compounds.
This guide covers:
What “complex” actually means in support
What constitutes a “complex query” is often misunderstood. Teams assume that it means “rare,” when in practice, complexity comes in many different forms.
You'll find complexity in:
Work that unfolds across multiple steps
Identity checks, verifications, eligibility decisions, confirmations, follow-ups. Any situation where the resolution happens through a chain of actions rather than a single reply.
Work that depends on rules or dynamic reasoning
Refund thresholds, billing adjustments, policy-based decisions, location-specific rules, account-level nuances – anything shaped by who the customer is and what state they're in.
Work that requires the agent to navigate tools or systems
Order lookups, subscription validations, record updates, workflow triggers. These are the moments where humans routinely lose time and where AI can return it.
Work that needs clarification before it can proceed
Disambiguating intent ("cancel order" vs. "cancel subscription"), gathering details, or confirming meaning before taking action.
Even simple questions can require complex workflows behind the scenes. Questions like “Where’s my order?” may seem straightforward, but resolving it means checking multiple systems, verifying shipping data, and piecing together what actually happened.
All of that is complex, and it’s the kind of work your team handles every day. With the right training, your AI Agent can too.
How to get your AI Agent to do more work for you
To get your AI Agent to take on more complex work (and lots of it), you need to teach it how your support operation actually works and give it access to the systems it needs to take action, test how it performs at scale, deploy it on multiple channels, and watch how it performs so you know where to fine-tune.
At Intercom, we have a framework for managing this called the “Fin Flywheel.” It’s a continuous improvement loop that lets you train, test, deploy, and analyze Fin on an ongoing basis so any changes you push to production are controlled, and you can use performance insights to inform future iterations.
Whether you’re using Fin or another AI Agent, the steps can be universally applied.
Let’s break them down.
1. Train
Start by mapping the work you want the AI Agent to take on.
Training an AI Agent to handle complex work requires a different approach than for informational queries. You need to teach it your policies, SOPs, decision paths, and tools your agents use.
The first step is mapping the types of complex queries you want it to handle; things like damaged order claims or account troubleshooting from start to finish.
When you have that list, start mapping the actual workflows your team follows to solve them today.
Capture:
  • What triggers the query
  • What questions human agents ask
  • What decisions they make
  • What systems they use
  • What causes failure
  • What outcome results in an issue being resolved
This is important; you’re not documenting for the sake of documentation, you’re painting a clear picture of how the work actually gets done. By doing that, you can create a roadmap for your AI Agent to follow.
While simple queries can be handled by pointing the AI Agent at content, complex ones need operational steps written out so the AI Agent knows how to move from intent to outcome.
AI Agents solve this with instruction frameworks. In Fin’s case, these are called “Procedures” (think of them as the modern equivalent of teaching a human agent how the job works).
With Procedures, you can use natural-language instructions and deterministic controls to train Fin to follow your SOPs carefully and exercise experience and judgment. These enable it to handle even your most complicated queries in a controllable, predictable way and still deliver a natural, agentic customer experience.

fin.ai

Procedures | Complex Queries

Fin automates the most complex customer queries—like refunds, transaction disputes, and technical troubleshooting—with speed and reliability. Give Fin detailed, step-by-step instructions, and it will follow them exactly as expected—reducing time to resolution and improving the customer experience.

Regardless of the instruction framework you’re using, here are some principles for building good instructions:
2. Test
When an AI Agent is handling high-stakes, complex work, you need a way to test how it behaves at scale. A robust testing approach will enable you to:
  • Identify mistakes before they impact customers.
  • Validate new or updated instructions (or Procedures, if you’re using Fin).
  • Check how it behaves across different customer types and data states.
  • Ensure tone, behavior, and policy alignment.
  • Confirm you’re happy with how it responds across channels and customer types.
The important thing here is to test for real-world conditions. Conversations rarely follow the “happy path,” so you shouldn’t test for the ideal state if you want to get a true read on how the AI Agent will act in a live environment.
Real-world processes, like refunds, quickly expand into combinations of customer states, policy thresholds, and edge conditions, creating dozens of possible paths through a single flow. You need a way to test these things at scale.
Test complex workflows end-to-end
Whatever AI Agent you use, run end-to-end tests with the complex queries and workflows you’ve scoped to ensure it’s responding and behaving as expected:
1
Start with a realistic customer message
2
Run the flow end-to-end
3
Observe which steps are selected, what data is retrieved, and how decisions are made
4
Evaluate both the outcome and the path taken to get there
Some AI Agents provide a built-in testing suite for this. In Fin’s case, it’s called “Simulations.” It enables you to test Procedures in end-to-end simulated customer conversations and see how Fin follows each step and reasons through them before setting them live.
3. Deploy
Once you’re confident in how your AI Agent performs in complex scenarios, you can set it live.
Start with a defined scope
Based on your testing, decide:
  • Which workflows the AI Agent should handle entirely end-to-end.
  • Which customers or segments you’ll deploy it to (e.g., certain regions, languages, or tiers).
  • Which channels it will operate on (e.g., chat, email, voice, social).
  • When it should hand over to a human, and how those handovers should look.

Deploy across the channels your customers actually use
There are two main levers for getting AI Agents to do more work for you. One is training them to handle complex work, the other is deploying it across as many channels as you can.
Remember that complex work doesn’t just reach your team through chat, it arrives over the phone, in messaging apps and community forums, and anywhere else your customers look for support.
Looking to scale support over the phone? Check out our guide that can provide a radically better phone experience.

fin.ai

How to get started with a Voice AI Agent

AI enables natural, reliable voice support that resolves issues fast. This guide shows why voice is unique and how to deploy AI systems that outperform IVRs.

Deploying your AI Agent to handle complex work across multiple channels is a fast way to increase the total amount of high-effort, complex work it can take on. ​​It also gives customers a consistent experience, no matter where they contact you.
4. Analyze
Once you’ve set your AI Agent live to handle the tough work, you’ll need to keep a close eye on how it’s performing. The early data will give you insight into where it’s resolving queries well, and importantly, where it’s falling short and needs to be improved.
You need to measure what reflects real automation value. Here are some metrics to focus on:
  • Resolution rate: How often the AI Agent fully resolves queries.
  • Automation rate: How much total volume the AI Agent is absorbing.
  • Customer effort: How much back-and-forth is happening in conversations, and whether customers experience less friction.
  • Escalation rate and quality: How often the AI Agent hands over to humans, where that happens, and the quality of the handovers themselves.
Most AI Agents will have built-in analytics you can use to track performance. Fin, for example, has an Insights product that breaks down how it’s performing and assigns a CX Score to conversations so you can monitor customer sentiment.
These types of insights should inform the next set of training improvements you make. Where things are going well, expand the AI Agent’s coverage and give it harder work. Where things aren’t going so well, use the learnings to fine-tune instructions or behavior, re-test, and then re-deploy.
Treat complex automation like a compounding flywheel
Getting your AI Agent to take on more complex work is an ongoing commitment, which is why we recommend using a continuous improvement flywheel. The more you invest in each step – train, test, deploy, analyze – the better your AI Agent gets.
As your AI Agent takes on more complex work, the role of your human teammates will change too. In the AI-first model, your team’s primary focus becomes optimizing the system, refining logic, and improving AI’s performance over time.
Aim higher than you think you need to
The biggest barrier to automating complex queries is underestimating what AI can handle. Many teams start small by using AI to handle simple, informational queries, but meaningful impact requires giving it the hardest, most complex work.
Remember that the work you automate determines the value you unlock. If you want your AI Agent to take on more, give it more, and give it the framework to do that work well.
Across a wide range of industries, we’re seeing companies lean into this and trust Fin to manage complex work end-to-end that used to require human skill and judgment.
Topstep
This has been huge because Fin Voice is able to not just answer the basic questions that traders have. But even if the traders start getting into some more intricate questions, they want to know, 'Hey, if my account balance is X and I lost this much, what is my consistency target for next month going to be?' Fin is able to go ahead, do the math in its head, and come up and give them an answer.
- Dennis O'Connor, Director of Support
Underdog Fantasy Sports
We got probably 7,000 contacts within two days, just from people asking where their money is and what’s the status of their withdrawal. Fin was able to pull up the customer’s profile, check their recent withdrawals, and give a personalized update with an 80% deflection rate. That alone saved us tons and tons of time.
- Andre Gamboa, Director of Customer Support
Anthropic
Fin moved beyond FAQs and transactional support – it started to deeply participate in the support experience.
- Isabel Larrow, Product Support Operations Lead at Anthropic
Ready to see how Fin can automate your most complex workflows?