Customers today expect AI to solve problems, not just answer questions. When someone interacts with a support bot, they want it to handle the task end-to-end, whether that means processing a refund, modifying an order, canceling a subscription, or triaging and diagnosing system issues.
Being passed to another channel or made to finish the job themselves leaves customers frustrated and annoyed.
To meet these expectations, connect AI to the systems where work actually happens. That can be harder than expected—but getting it right is essential for pleasing customers and standing out from the competition.
When agentic AI hits bottlenecks
For support use cases, LLMs have essentially been commoditized. With Anthropic, OpenAI, and others all delivering comparable capabilities, the constraint has shifted from “Is the AI smart enough?” to “Can the AI reach what it needs to?”
When GenAI was the name of the game, that meant connecting a chatbot to a few internal knowledge sources. The requirements for agentic AI go much further. A single task might involve a dozen different integration points, potentially including:
- Billing and payment systems: To process refunds and update payment methods
- Inventory and order management: To check status and modify orders
- Customer relationship management systems: To access customer history and update records
- Shipping and logistics application programming interfaces (APIs): To track packages and initiate returns
- Authentication and identity systems: To verify users and reset passwords
- Internal workflow and ticketing systems: To diagnose and address system issues
Each of these integrations might look different from the others. Some will have clean APIs. Others will need browser automation. Some third-party systems lack modern interfaces entirely and require creative hacks, such as email-based workarounds.
Protocols like the model context protocol (MCP) can simplify how AI systems connect to multiple data sources, but they don’t eliminate integration work entirely. That leaves organizations facing a resource problem.
Without the technical skills to do this work themselves, support teams’ integration projects become just another item in the engineering queue—typically behind a long list of product features. The end-to-end resolutions promised by agentic AI remain a work in progress, moving slowly, if at all.
How AI integrations get done (sometimes)
At a high level, organizations have three options for handling the interoperability requirements of agentic AI.
Option 1: Engineer dependency
In the best case, let’s say your agentic AI implementation gained the executive support necessary to become a top engineering priority. The team performs all your initial integrations right away, and the system works flawlessly.
So far, so good—but what happens at day 90 or day 180, when leadership’s focus has moved elsewhere, and support needs to iterate on the platform or add a new workflow? You’re right back in the queue, fighting a losing battle for resources while customer needs go unmet.
Option 2: Manage service
Outsourcing integration work back to the AI vendor can be an attractive alternative. They know their platform best, and they can handle configurations and connections efficiently. This works well when you’re up against an aggressive implementation timeline and need expert help navigating complexity.
But over time, it becomes clear that you’ve just traded one dependency for another. The changes you need will always compete with the vendor’s other priorities. Do you want to be at the mercy of the vendor’s service level agreements?
Platforms built around a single ecosystem present a related issue. If your tech components live predominantly within one vendor’s architecture, the vendor’s integrations can work flawlessly from day one. But few companies and use cases live entirely within walled gardens. An AI optimized for one ecosystem may struggle to extend beyond its boundaries to other systems within the company and beyond.
Option 3: Support empowerment
Support teams performing their own integrations never have to wait in someone else’s queue. If they can configure and manage their own AI workflows, a support manager who spots an automation opportunity can implement it immediately.
This isn’t a big ask, as leading agentic AI platforms offer integration tools designed for non-technical users. However, this requires ensuring that “simple to use” is more than just a marketing term. As you evaluate options, bring support leaders in for a hands-on integration test-drive.
Identify three high-priority support use cases you’d like to enable, then have non-technical team members try configuring them themselves. Can they do it without the vendor holding their hand? What’s their comfort level with this type of work?
Test drive your integration
Interoperability is an ongoing operational requirement, not a ticket to be closed. There will always be more connections to make and more workflows to configure. Plan your journey accordingly.
Don’t aim for comprehensive integration across all your touchpoints right out of the gate. Identify three to five use cases that combine high customer impact with reasonable complexity, get them working, measure the results, then expand based on what you learn.
Build iteration into your assumptions. Six months from launch, you’ll have new products, different vendors, changed processes, and new insights about what customers actually need versus what you predicted. Systems that made sense initially may need to be replaced.
Organizations that treat deployment as a foundation for ongoing refinement maintain momentum better than those expecting a stable end state. Your AI is only as valuable as what it can reach and act upon.
Al Martin is the VP of customer experience at Forethought.