AI Agents in action: Lessons from tech innovators

AI agents are everywhere right now. But most of them are just chatbots with better branding. The real transformation comes when agents stop just responding and start doing actual work: opening tickets, resetting passwords, coordinating across apps, and getting approvals. But deploying that kind of agent in a real company with legacy workflows, hundreds of SaaS tools, and tight security controls is hard.

We recently met with two forward-thinking tech companies—a fast-scaling B2B sales platform and a fintech operations team—as they evaluated AI agents to solve real operational problems. 

Here are five things we learned from their journey.

1. Talking isn’t enough. Agents have to act.

One company had already rolled out a strong AI search tool. But it wasn’t enough. “We wanted something that didn’t just summarize data but acted on it,” one team member said. That meant Slack agents that could trigger automations, reset passwords, or escalate urgent requests directly from the conversation.

Another team was running a brittle chat-only system to handle Zendesk tickets and escalations. They replaced it with an agent that could validate input, route requests based on time of day, and avoid noisy false positives, all while giving humans the right level of control.

Takeaway: A chatbot that responds is the bare minimum. What matters is whether the agent can act via APIs, automations, and multi-step workflows.

2. The best agent interface is the one your team already uses.

Both companies wanted one entry point for every team, not a sprawl of team-specific bots. For one, that meant a single “company agent” in Slack that could handle everything from IT tickets to HR questions and contract approvals.

The fintech team faced a similar challenge: sales ran on Salesforce, risk operated in Zendesk, and support requests came in through Slack. They needed an agent that could bridge those tools and keep everything moving without toggling between systems.

Takeaway: Centralized doesn’t mean limited. Start with one entry point, like Slack, and let the agent route requests, call sub-agents, and orchestrate across tools.

3. Full autonomy sounds great (until something breaks).

Both companies were excited about autonomous agents, but neither wanted to risk a bot taking irreversible action without oversight.

One team built a semi-autonomous agent that could handle requests like password resets, but only after verifying identity and getting a human OK. Another configured their agent to escalate requests based on business hours, flag anomalies, and block certain actions if criteria weren’t met.

Takeaway: Guardrails matter. Use human-in-the-loop approvals for sensitive flows and automated rules for everything else. Autonomy scales when it’s designed with trust.

4. Start with one workflow. Then grow. Fast.

The B2B platform started with a single IT use case: automating password resets. That small win opened the door to broader workflows such as summarizing support tickets, automating contract intake, and prepping CS teams before calls. Their goal? Save 10 hours per employee per week.

The fintech team did the same: replaced one form, added logic, verified user permissions, and built in escalation paths. Within weeks, they had a reliable agent running in Slack, backed by Tray.

Takeaway: Don’t try to deploy the perfect agent on day one. Nail one use case, then expand. 

5. Flexibility is the difference between a toy and a system.

Both companies brought complex environments to the table: hundreds of apps, specialized tools, and internal systems nobody else uses. One team said it best. “We use weird stuff but it just works.”

They chose Tray’s Merlin Agent Builder because it let them do exactly that: build agents using low-code workflows, swap LLMs depending on the task, plug into Okta or Notion, and add their own business logic as they go. Out of the box, they had access to hundreds of connectors and a toolkit to build anything else.

Takeaway: Building your first agent is great, but you need to plan for the next fifty. Build agents on a foundation that will let you scale. Flexibility makes that possible.

What’s next?

Both companies are now rolling out their agents across departments. They’re building agent-of-agent architectures. Routing tasks dynamically. Moving from AI that responds to AI that gets work done.

That’s what Tray.ai was built for: composable agents on top of a real integration backbone. If your current “AI strategy” stops at chat, it’s probably not a strategy.

Ready to build an agent that actually does the job? Explore Merlin Agent Builder