This isn't a story about using AI to reduce headcount
It's about why we don't need to add new headcount for the next 1,000 customers we add.
For your company, the numbers may be different, but the underlying thesis is the same:
If AI agents could take on more of the investigation and execution, how much more work could your support team handle?
For most of the history of customer support, the answer was constrained by a simple relationship.
For most of customer support's history
More customersMore support team members
We've spent 14 years building customer support software, and everything we learned building Groove taught us that math well. Agentic Support is what begins to break it.
Before AI agents, the hard part wasn't writing the reply
At its peak, our support team had six full-time people helping 1,300 customers.
We had years of documentation, saved replies, automations, a knowledge base, and people who knew the product inside and out.
But a complicated ticket was still largely manual.
A customer sends a complicated ticket
The hidden workBefore a single word is written
Understand the request
Read the conversation and figure out what the customer actually needs.
Gather customer context
Identify the customer, their account, and relevant history.
Search previous conversations
Find similar issues and see how they were handled before.
Search the knowledge base
Find the documentation that might help explain or resolve the issue.
Investigate the account
Check account and product data to understand what's actually happening.
Bring in engineering when needed
Confirm bugs, technical issues, or edge cases.
Decide what to do
Put all of that context together and determine the right resolution.
The visible work
Write the reply
Seven steps of investigation before a single word of the reply.
The reply might take a couple of minutes to write. Figuring out what it should say was where most of the work happened.
That's why simply using AI to write faster wasn't enough.
First, we tried adding AI to the help desk we already had
Our customers were asking for AI, so we started where you'd expect: adding AI features to Groove.
We built tools to draft responses, summarize conversations, and help agents work faster.
They were useful.
But the workflow hadn't really changed.
HumanCustomer question
HumanHuman investigates
AIAI assists
HumanHuman acts
HumanCustomer gets help
AI made individual steps faster. A human still had to orchestrate the work.
AI was making individual steps faster. It wasn't changing who had to orchestrate the work.
So we started experimenting with something different inside our own support operation.
Instead of asking how AI could help write the answer, we wanted to know how much of the investigation itself it could handle.
That's when we started using Claude.
Then we started using Claude
We started using Claude for straightforward support questions. Then, as MCP became more useful, we connected it to the systems our team normally searched by hand.
Claude could search customer conversations and our Groove Knowledge Base, check Linear for known bugs and feature requests, look at account and billing information, and access internal product data and logs.
We started turning the investigations our best support people already knew how to do into repeatable Claude workflows. Instead of prompting from scratch, we could give Claude a defined process: what to search, which systems to check, what information mattered, and how to reason through the problem.
It was a meaningful improvement.
We estimated Claude made each support person about 22% more efficient, saving roughly eight hours per person per week.
But it wasn't a reliable operating system for support:
Costs scaled with usage
Across our support operation, we were on pace to spend roughly $48,000 per year running these Claude workflows. A complicated investigation could take up to 25 minutes and cost as much as $8.
Workflows were difficult to share and maintain
Everyone needed the right Claude setup, permissions, integrations, and internal access. And when our product or processes changed, the people closest to support couldn't always update the workflows themselves.
Every investigation started from scratch
Claude repeatedly searched the same conversations, account data, knowledge, product information, and internal systems.
Humans still had to run everything
Someone still had to open the ticket, decide what Claude should do, trigger the workflow, wait, review the result, and then go back to the conversation.
Claude made the thinking faster. Humans were still orchestrating the work.
The more capable the models became, the clearer the gap became. Context couldn't depend on a human finding it first. Processes couldn't live only in people's heads. And agents couldn't wait for someone to tell them when to start.
We didn't need another AI feature. We needed to rethink the support platform around AI agents from the beginning.
Customer data was everywhere
Claude also exposed another problem.
The information needed to resolve a conversation rarely lived in one place.
To resolve one ticket
Conversations
Groove
Knowledge
Groove Knowledge Base
Billing
Stripe
Internal context
Slack
Product issues
Linear
Account data
Groove + internal systems
Product usage
Internal product data
The context for a single ticket was scattered across seven systems.
All of the information usually existed somewhere. The hard part was finding it, understanding what mattered, and assembling it into a picture of what was actually happening.
The problem wasn't intelligence. It was context.
What if our team could direct the work instead?
We had spent 14 years building support software around a world where people executed nearly every step.
Now the technology was changing.
What if our support team didn't have to personally execute every step between a customer question and a resolution?
What if agents could gather the context, search our knowledge, investigate the account, follow our processes, take approved actions, and bring in a human when judgment was needed?
Our team would still own the customer experience.
They'd just spend less time executing every step and more time deciding what should happen next.
We could keep retrofitting AI into the model we'd built years ago.
Or we could rebuild the platform around the assumption that humans and AI agents would work together from the start.
We chose to rebuild it.
Claude had given us the reasoning engine.
What we needed was the system around it:
Persistent context
Shared Skills
Editable Guidelines
Repeatable Processes
Integrations
Actions
Permissions
Escalation rules
Agents that could start working automatically
That became Helply.
Agentic AI moves support teams from doing the work to directing it.
Then we put Helply to work in our own support operation
We didn't want to build Helply around demos or hypothetical workflows.
We already had the perfect proving ground: our own support operation.
Real customers. Real conversations. Years of history and knowledge. Account data. Product issues. Bugs. Feature requests. And more than a decade of edge cases.
So we started running that operation on Helply.
The sequence of work changed.
Before
With Helply
A ticket waited for someone to start working on it.
Helply starts working as soon as the ticket comes in, gathering context and beginning the investigation before a human gets involved.
Our team searched across conversations, knowledge, account data, and internal tools.
AI agents gather the relevant context automatically and bring it into the investigation.
We used Claude manually to help investigate and reason through difficult issues.
Claude's reasoning happens inside the workflow, with the context, knowledge, and instructions it needs already available.
The person handling the ticket had to know how to investigate the issue.
Skills and processes capture how we handle recurring problems so that knowledge can be reused by our team and AI agents.
Complex issues often required pulling in someone with deeper product knowledge.
Helply investigates first and escalates with the work already done, so specialists start with context instead of starting from scratch.
Useful knowledge often disappeared inside resolved conversations.
Resolved issues expose knowledge gaps, helping us improve what customers, humans, and AI agents know the next time.
Helply didn't remove our team from the process. It changed where their time and judgment were needed.
Instead of gathering every piece of context and executing every step, our team could start with work already completed and decide what should happen next.
The human no longer has to start from zero.
What happens when a ticket comes in
Today, when a customer contacts us, Helply starts first.
It identifies what the customer is asking, determines what it needs to investigate, gathers the relevant context, and begins working before either of our two support people touches the conversation.
Then what happens depends on the issue.
If the Agentic AI is highly confident in the solution, it can close it out with guardrails.
If the ticket is complex, our support person can review the work and approve or adjust the response.
Not every ticket needs the same amount of human involvement.
Resolve
Tier 1 / repeatable
Helply has the knowledge, confidence, and permission to handle the issue end-to-end. It takes the appropriate action, responds to the customer, and resolves the conversation.
Investigate together
More complex
Helply gathers context and investigates first. A person reviews the work, redirects the agent if needed, and decides what happens next.
Escalate
Judgment required
Helply recognizes that the issue needs a person and hands it over with the relevant context and investigation already completed.
Here's what the middle path looks like.
Let's say a customer emails us about an import that keeps failing. It's not something Helply can confidently resolve immediately, so it begins investigating before our team gets involved.
Customer“Hey, I'm having trouble with imports. I've tried re-uploading, but it still isn't working. Can you take a look?”
1
Understand
Identifies the issue and what the customer is trying to accomplish.
2
Gather context
Pulls relevant customer, account, and conversation history.
3
Search
Finds relevant articles and previous resolutions from the knowledge base.
4
Investigate
Checks the systems and data required to understand what actually happened.
5
Follow the process
Applies the appropriate support guidelines and workflow.
6
Act
Takes the available action, or prepares the response for review.
7
Resolve or escalate
Closes the loop with the customer, or hands off with the work already done.
Assist AgentInvestigating
What the evidence shows
10:02:14Matched conversation to account · Northwind Labs (Pro)
10:02:15Pulled last 3 import attempts from product activity
10:02:16Found related article · "Why imports fail to process"
10:02:18Import job #48213 failed: column mismatch on upload
Root cause identified — suggesting next steps and drafting a response.
The agent investigates the moment a ticket arrives: matching the account, pulling product activity, and finding the root cause.
By the time someone on our support team opens the conversation, Helply has already gathered context, investigated the issue, and identified the likely root cause.
If that's enough, our support person reviews the work and approves or adjusts the response.
Instead of searching for information, they're deciding what should happen next.
Suggested next stepsDraft · reply to customer
No teammate needed unless the customer wants the exact internal detail. A customer-ready draft is staged below.
Hi there,
Thanks for the patience! Your last import failed because of a column mismatch on upload. The file had an extra column the importer didn't expect.
I've re-mapped the columns and re-run the job. It completed successfully, so your records should be in now. Let me know if anything still looks off.
Prepared by Helply from import job #48213 and past resolutions.
It then stages a customer-ready draft, so a person reviews the work instead of starting it.
If our team wants more context, we tell Assist to keep digging.
“Check whether we've seen this before. Look through previous support conversations, known issues, and relevant product information. Tell me what happened in similar cases and how we resolved them.”
Our support team directs the investigation. The Agentic AI does the legwork.
The agent handles more of the investigation. Our team directs what happens next.
Agents do the work
Everything we learned using Claude shaped how we built Helply. We stopped expecting humans to act as the bridge between AI and all of our systems, and started expecting AI to take ownership and deliver results.
By the time an AI agent starts working, it can already access:
What the customer is trying to do
Account and product context
Previous conversations
Similar issues
Known bugs and feature requests
Relevant knowledge
Processes and Guidelines
If the agent needs more, it can investigate further.
And if a human needs to step in, they don't start over. They get the context, investigation, and recommended next step.
Every difficult conversation can also expose a gap:
Was context missing?
Was knowledge outdated?
Was a process unclear?
Was there an action the agent couldn't take?
Those lessons help us improve the Knowledge, Skills, Processes, Guidelines, Actions, and Context available next time.
The team isn't only directing conversations. It's improving the system that handles them.
We taught Helply how our team works
Context wasn't enough.
Our support people had spent years learning how we handle customer issues: when to escalate, what information to trust, how to investigate different problems, and which actions were safe.
If agents were going to handle more execution, that knowledge couldn't stay inside people's heads. So we made it explicit.
Guidelines
How should the agent behave?
Skills
What should the agent know how to do?
Processes
What steps should it follow for a type of issue?
Actions
What is the agent actually allowed to do?
The more of our operating knowledge we put into Helply, the more work our team could direct instead of manually execute.
The impact went beyond support
Support
68%
Faster median first response18.2 → 5.8 min
64%
Conversations with substantive AI-agent work
Our two-person support team moves through harder issues faster because much of the investigation has already happened. Shared Skills, Guidelines, and Processes mean less time repeating investigations and more time on conversations that need human judgment.
Product & Engineering
40% ↓
Fewer escalations to Product & Engineering
~160
Investigations performed by Helply over four weeks
Helply can check known issues in Linear, search product information, inspect logs and product data, and bring that context back into the conversation. So Support resolves more technical questions on its own, and Product and Engineering spend less time acting as an extension of the support queue.
Engineering gets an investigated issue, not another question to investigate.
Across our team
We're a 15-person company, and customer context matters well beyond our two-person support team.
Roughly 7 of our 15 people use Helply outside the core support workflow each week, including Product, Engineering, Customer Success, and leadership.
Instead of finding the person who knows where to look, they can start with the context Helply has already assembled.
Here's what changed in the numbers
Response
Before Helply
18.2 min
Median first response
With Helply
5.8 min
Median first response
Before vs. after median first-response time.
61% → 82%
Within our one-hour target
12.4 min
Sooner, on median first response
AI work
64%
Conversations with substantive AI-agent work
40% ↓
Product & Engineering escalations
Those metrics tell us Helply is changing how the work gets done.
But there's another number we care about even more: capacity.
Capacity
1,300
Customers
~1,800
Conversations / month
2
Full-time support people
0
Planned support hires
At its peak, our support team had six full-time people managing support for 1,300 customers. We now have two managing the same workload.
More importantly, we won't need to increase headcount as we add the next 1,000 customers.
Agentic AI helps your team do more with less, deliver better work, and grow without adding headcount.
The new math of customer support
We've spent 14 years building customer support software. Our years building Groove taught us the math:
The old math
Customers ↑Tickets ↑Headcount ↑
Growth eventually meant hiring more people.
The new math
Customers ↑AI leverage ↑Human capacity ↑
Growth doesn't have to mean proportional headcount growth.
We went from six full-time support people to two.
The lesson isn't that every company should have two support people, or even that you should reduce support headcount. It's that people shouldn't have to personally execute every step between a customer problem and its resolution.
Agentic Support makes that possible. And the future of Agentic Support is here with Helply.
Move from doing the work to directing it.
Frequently asked questions
Is Helply the new version of Groove?
Helply was built by the team behind Groove and incorporates more than a decade of what we've learned building customer support software.
Groove was designed for an era when humans handled nearly every step of customer support.
Helply was built from the ground up for a world where humans and AI agents work together.
Today, Helply is the platform we’re building for that future.
Couldn't we just build this with Claude?
You can build pieces of it. That's how we started.
The hard part was turning individual experiments into something our entire support operation could depend on: shared Skills, editable Guidelines, repeatable Processes, persistent context, permissions, Actions, escalation rules, and agents that start automatically.
Claude gave us the reasoning engine. Helply became the support system around it.
What happens when an agent gets something wrong?
We control what agents are allowed to do and when a human needs to be involved.
For workflows that require review, our team can see what the agent investigated, which information it used, what it found, and what it recommends doing next before taking action.
When something isn't right, the human can correct it, continue the investigation, or take over.
Those exceptions also show us where our Knowledge, Context, Skills, Processes, or Guidelines need to improve.
What can Helply actually see and do?
Helply works with the systems and information we give it access to. Inside our operation, that can include conversation history, the Groove Knowledge Base, account information, Stripe billing, Linear issues, product activity, and logs.
Agents use that to understand a request, gather context, investigate, follow our processes, recommend next steps, and take the Actions we’ve allowed. We control what each agent can access and do.
Does Helply respond directly to customers?
Yes, when we've configured an agent to do so and it has enough confidence to handle the conversation.
Other issues can require human review or automatically escalate to our team.
The point isn't to force every conversation through AI. It's to let agents handle the work they can and bring humans in when their judgment is actually needed.
How long did it take to see results?
We started seeing clear results within about four weeks of putting Helply into our own support operation.
For the results in this report, we compared the four weeks before rollout with the four weeks after, looking at response times, AI involvement, support volume, investigation time, and human workload. During that period, conversation volume increased approximately 4%.
All figures in this report are illustrative placeholders and will be replaced with validated production data before publication.
Did Helply reduce your support team from six people to two?
Not by itself. Our team had six full-time people at its peak and has two today, but that change happened over time and wasn’t entirely caused by Helply.
What Helply has changed is how much support work those two people can handle and how we think about future hiring. Today, two people support roughly 1,300 customers, and we’re not planning to add another support person anytime soon.
Are you trying to replace support people?
No. We're trying to change what support people spend their time doing. Traditional help desks were built around humans executing almost every step. Agentic Support lets AI agents handle more of those steps while people stay in control.
The end state isn't support without people. It's people directing far more work than they could ever do themselves.
Does Helply learn from every investigation?
Every investigation tells us how an issue was solved, what context mattered, where the agent looked, and where it got stuck. We use those patterns to improve Helply’s Knowledge, Skills, Processes, Guidelines, Actions, and Context so similar issues are easier to handle next time.
We do not claim automated learning as a product capability until that behavior has been verified. [VERIFY THIS PRODUCT BEHAVIOR BEFORE PUBLISHING]
Keep reading
See how we put it into practice.
Continue reading to see how agents investigate incoming tickets, take approved actions, and bring in people when needed. The rest of the case study covers:
How we decide when agents resolve, collaborate, or escalate.
How we turn our team’s knowledge into guidelines, skills, and processes.
What changed across Support, Product, and Engineering.
See what Agentic Support could look like for your team.
Helply is the AI-native customer support platform built from more than a decade of experience building support software, and everything we've learned running Agentic Support inside our own operation.