What happens when AI starts working before your team does

I'm the Co-Founder of Helply, an AI-native support platform. Our agentic support platform includes a modern help desk and ticketing system designed for humans and AI agents to work together, an AI knowledge base that keeps itself up to date, and a powerful MCP that allows you to connect with your tech stack to give Helply full context on every customer.

Think of Helply as the context and orchestration layer that allows AI to take action on every ticket, from responding and investigating to resolving end-to-end.

I'll be speaking more on this at the AI for Customer Support Summit in San Francisco this September. Here's what changed once we stopped asking AI to draft and started asking it to work.

A customer wrote in because their imports kept failing. They had already tried uploading the file again. They didn't include an error message, job ID, or attachment. That ticket used to begin with a human customer support agent opening four tabs.

The Helply AI agent matched the conversation to the customer's account. It found their last three import attempts. Then it checked the relevant knowledge article and identified the problem. The uploaded file had a column mismatch. The investigation took a total of four seconds. The best part was, nobody on our team had to even open the ticket.

The AI agent didn't send the reply automatically. That wasn't within its permissions for this case. It completed the investigation and prepared a customer-ready response. A human agent then reviewed, approved and sent it.

That was the moment the change felt real. Our customer support agent opened the conversation with the work already done. They made the decision – it only took a couple of minutes – instead of spending 20+ minutes gathering evidence manually to eventually make that same decision.

Where we started with AI

Before we got where we are today, we went through two phases. 

First, we added AI to Groove, our help desk at the time. This drafted replies and summarized conversations for us, helping agents work faster while making sure the workflow stayed the same. 

But then we started using Claude with MCP connections. We decided to connect it to Groove, our knowledge base, Linear, Stripe, Slack, product data, and logs.

Now AI could investigate, not just write. But a real human person still had to run the process: a ticket arrived; someone opened it and had to decide what to investigate; then they gave Claude the context, waited, reviewed the output; in the final stage, they then returned to the help desk.

Our process didn't feel broken anymore. Instead, it felt like progress. The problem became clear after we mapped a complicated ticket. Before anyone replied, they had to understand the request, gather account details, search past conversations, check the knowledge base, investigate product usage, involve engineering, and decide what to do.

Writing the reply took two minutes. Figuring out what the reply should say could take up to 20 minutes. AI was helping with the final step, but most of the work still happened before that.

We hit a ceiling

While Claude made the investigation faster, it didn't remove the investigative process from our human agent's plate. We hit three limits.

1. Cost

We were on pace to spend about $48,000 per year on Claude workflows. One complicated investigation could take 25 minutes and cost $8.

2. Context

A single ticket might require information from 7 systems. Someone still had to find and assemble that information for Claude.

3. Orchestration

A person had to open the ticket, start the workflow, wait, review everything, and take the next action.

Those workflows also lived in individual setups. They were hard to share and maintain. Each investigation started close to zero. We estimated Claude made each support person about 22% more efficient. That saved roughly 8 hours per week.

Those were real gains, but we were still limited by human capacity. The bottleneck was the system: for all its wisdom, Claude didn't know when to start, what it could do, or when it needed approval. We had to build the new agentic support platform around this new concept of directing AI work.

The question we had backwards

As neat as it would be, there wasn't “one ticket that changed everything.” We realized we were asking the wrong question. We kept asking, how can AI help our support team work faster? That question assumes every ticket must start with a person.

We started asking, what if our team directed the work instead of executing every step? Some tickets don't need a person. Some need a person after the investigation is complete. Others require human judgment from the start.

The system should determine which path fits. It shouldn't send every conversation to a person by default. We couldn't get there by adding more AI features to Groove. Claude gave us the reasoning engine. We still needed shared context, reusable skills, clear guidelines, repeatable processes, integrations, actions, permissions, and escalation rules.

We needed AI agents that could start working when the ticket arrived. That became Helply.

What actually changed for our customer support agents

Our human support agents eventually stopped starting every ticket from zero. The AI agents now handle repetitive tickets, gather context, search, investigate, follow our processes, and take approved actions.

Our people direct the work. They make judgment calls, handle exceptions, manage relationships, and improve the system. The results showed up in four weeks. 

Median first response fell from 18.2 minutes to 5.8 minutes. That was a 68% improvement. Responses within our one-hour target increased from 61% to 82%. Agents performed substantive work – checking accounts, gathering info, validating data – on 64% of conversations. Escalations to product and engineering fell by 40%.

We established pretty early on that capacity matters most to us. We now support about 1,300 customers and 1,800 monthly conversations with two full-time support people. At one point we had six. More importantly, we're growing, and we don't expect to hire a third customer support agent for the next 1,000 customers.

The biggest surprise was how quickly this spread beyond support. We're a 15-person company. About seven people use Helply each week across support, product, engineering, customer success, and leadership. They can start with the context already assembled. They don't need to find the person who knows where everything lives.

Nobody misses copying account IDs between tabs. Removing that work didn't make the job smaller. It created space for the creative, more meaningful work that only a human can do.

Earning the right to act

If every ticket still waits for a person to begin, AI hasn't truly changed your capacity – it’s only made part of the existing workflow faster. I get so many support teams ask me whether I trust AI. I’ll be honest, that question is too broad to be helpful. Instead, I like to ask what AI has earned the right to do.

Use evidence. Measure resolution accuracy, escalation rates, customer satisfaction, policy compliance, and human correction rates. Then create three clear paths.

1. Resolve repetitive work when the agent has enough knowledge, confidence, and permission. 
2. Collaborate when the agent can complete the investigation, but a person should decide. 
3. Escalate when the case requires human judgment from the start.

Every time an exception happens – the AI agent gets something wrong, or a human support agent has to step in – treat it as a diagnostic, not just a one-off fix:

• Was context missing? 
• Was the knowledge out of date? 
• Was the process unclear? 
• Did the AI agent lack permission to act? 

Fix the actual cause, and the system handles that type of case correctly next time.

AI should own as much work as it can handle reliably. No more and no less. The gap between AI assisting and AI acting is the system that lets the model do real work, and lets people audit and improve it in real time.