Build, buy, or both? How to choose AI agents without locking yourself in.

Data from Plain’s customers on what support teams run today, and how to keep your options open as your support changes.

ON THIS PAGE

What teams like yours are doing

Your options

Decide by type of work

How to compare agents honestly

What your platform has to handle, whichever agent you choose

Your first 30 days

What changes for your team

What nobody knows yet

Where Plain fits

Share this guide

Among Plain’s customers, the share of support teams running an AI agent in their inbox doubled in under a year. More support leaders are now deciding which agent to use: the one built into their support platform, one from a vendor, or one they build themselves.

AI-native companies

60%

35%

All support teams

31%

64%

Built their own

Built into the support platform

Other

64% of teams with an agent use the one built into their support platform, and 31% built their own (some teams run more than one). AI-native companies are the exception, with 60% of teams using agents they built.

Teams that build give similar reasons. They want an agent they control, one that can reach their own data, take actions in their own tools and follow their own policies.

Building is no guarantee of better results. In our data, teams that build and teams that buy see similar resolution rates on average, but results vary widely within each group. In a quarter of teams, the agent resolves 21% of its conversations or fewer. In another quarter, it resolves 59% or more.

Whether you build or buy tells you little about how well your agent will work.

You may also want a different agent later, and some teams already run more than one. If changing agents means changing support platforms, you move your conversations, history and workflows each time. So choose a support platform that can run any type of agent, not only the one you start with.

Section 01

What teams like yours are doing

Figures come from Plain’s customer workspaces, December 2025 to October 2026. Most are AI-native and developer-tool companies. See About the data.

Most teams start with the agent they already have. 68% of teams started with the agent built into their support platform, and 64% of teams with an agent run it today. It’s the fastest way to start, because it already works with your channels and your help center.

AI-native teams are the exception. Among AI-native teams with an agent, 60% built their own and 35% run the built-in one. In developer-tool companies the split is closer: 38% built and 52% built-in. If you’re AI-native, your peers are building.

Teams running the built-in agent · by type of company
All teams with an agent
64%
Developer-tool companies
52%
AI-native companies
35%

Bigger, busier teams build or bring in a vendor. Among teams with 26 or more people replying to customers, 73% of those with an agent run one they built or bought from a vendor. Among teams with one or two people replying, 87% run the built-in agent. Volume shows the same split, and nearly all teams above 2,000 conversations a month run a built or vendor agent. The setup that works for a small team often stops fitting a large one.

Teams mostly keep the agent they start with. Only 8% of teams that started with the built-in agent later built their own. Most teams stay with the agent they start with, so measure it early on your own conversations (section 4). If it isn’t working, your support platform shouldn’t be what stops you from changing it.

Teams get one agent working before they add a second. 91% of teams with an agent run exactly one that talks to customers.

Agents do most of their work on chat. In the median team with an agent, it replies in 97% of chat conversations, 22% of email and 1% of Slack. If most of your support happens over email or Slack, expect an agent to cover less of it.

Agent replies in, by channel · median team
Chat
97%
Email
22%
Slack
1%

The most common reason for a handoff is a gap in knowledge. It accounts for a median 17% of handoffs with a recorded reason. A customer asking for a person accounts for 7%. A teammate stepping in unprompted is nearly as common as a knowledge gap, at 15%.

Section 02

Your options

You have three options: the agent built into your support platform, one from a specialist vendor, or your own. Most teams start with the built-in agent and add another when the work needs their own data or tools.

The built-in agent · 64% of teams with an agent

  • Good for: getting started fast, and routine questions on chat.

  • What it costs: usually the least, because it’s part of the platform you already pay for. The real cost is keeping your help content current.

  • When you’ll want more: customers ask about their own account, or need an action taken in your systems. Help content can’t answer those.

  • Example: at Laravel, the built-in agent is the first to pick up 97% of chat conversations, about two thirds of the team’s support volume. Since it went live, the share of chat customers giving the top satisfaction rating has risen from 73% to 90%.

The vendor’s agent · 10% of teams with an agent

  • Good for: teams that want more than the built-in agent offers, without building and maintaining their own.

  • What it costs: vendors often charge per resolution, and each defines a resolution differently. Compare definitions before you compare prices (section 4).

  • When you’ll want more: your needs become specific to your product, or you don’t want to open your systems to another company.

  • Why some teams like it: they pay per result. As one customer put it: “We don’t want to commit to pay something when we don’t know the results.”

Your own agent · 31% of teams with an agent

  • Good for: work that needs your own data and tools, such as account-specific answers and actions.

  • What it costs: varies based on complexity. You pay for engineering time to build and maintain it, plus model and hosting costs.

  • When you’ll want more: maintaining it pulls engineers off the product.

  • Example: n8n built its own agent after support volume grew twentyfold in early 2025. About 70% of tickets are now handled end to end by AI, and resolution time fell from 5 days to 1.5. Read the n8n story.

Why teams build, and what gives them pause

From Plain’s conversations with customers and prospects.

  • Control. “We see support as something that’s really core to our business. It’s a moat. Because this experience is so core, we want to own it end to end.”

  • Access to their own systems. “I want something that is actually agentic. It can plug into our APIs and take actions, or pull in user information, order information, account information from our homegrown tool.”

  • Keeping systems private. “We don’t want to expose all of our services to loads of third parties.”

  • The cost of maintaining it. “We do have the tech stack to build our own AI agents. But do I want that level of work, maintenance, all that kind of stuff?”

Running two side by side

Some teams run the built-in agent on routine questions and their own agent on the work that needs their systems. 5% of teams with an agent run more than one type today.

Incoming conversations

Built-in agent

Answers from your help content.

Your own agent

Account data and actions in your tools.

Both hand off to your team, with full context

You don’t have to pick one option for everything. Section 3 breaks the choice down by type of work.

Section 03

Decide by type of work

Most teams end up with a different answer for different work, so decide one type of work at a time.

Type of work

Start with

WHY

Drafts and summaries for your team

Team sees it

Start with

Your platform's internal assistant, with a person approving

Move on when

A person checks every reply before a customer sees it

Triage and routing

Team sees it

Start with

Your platform’s internal assistant

Move on when

Routing is low risk, since no customer sees it

Routine customer questions

Customer sees it

Start with

The built-in agent

Move on when

64% of teams with an agent run it, and chat is where agents do most of their work

Account-specific answers and actions

Customer sees it

Start with

Your own agent or a vendor’s

The agent needs your account data and your tools

Escalation to engineering

Eng sees it

Start with

People, with the agent investigating first

Engineering starts with full context

Section 04

How to compare agents honestly

No independent benchmark compares AI agents across vendors. Every resolution rate you see is published by the vendor, using its own definition.

Three terms tell the numbers apart

Involvement

The AI replied in the conversation at some point.

Resolution

The conversation ended with no person involved.

Deflection

The customer didn’t reach a person.

The same agent can look very different depending on which one a vendor reports. If a vendor says its AI is in front of 60% of conversations, that’s involvement. It doesn’t tell you how many of those conversations a person still had to finish.

What real teams see

In the median team with an agent, a reply followed by a handoff is the most common path.

41%Agent replies inof all conversations
32%Resolvedof agent conversations
58%Handed to a teammateof agent conversations
Medians across teams with an agent
How much results vary

Results vary far more from team to team than between building and buying. A quarter of teams see 21% resolved or fewer, and a quarter see 59% or more. Judge an agent on your own data.

Share of the agent’s conversations resolved, by team

0%

21%

59%

100%

Bottom quarter

Middle half of teams

Top quarter

How to measure your own agent

Count the conversations the agent replied in. Split them into resolved and handed off, using the definitions above. Be careful with an agent’s own “handled” label, because the agent sets it itself.

Questions to ask any vendor, or your own team

01

How do you define a resolution, and does a handoff ever count as one?

02

Does a conversation the customer abandons count as resolved?

03

Can I see replied and resolved separately, on my own data?

04

What happens to the conversations the agent hands off, and how long do they wait for a person?

Section 05

What your platform has to handle, whichever agent you choose

Your agent will change as you grow, so these are the things your platform needs to keep steady underneath it. Use this list to evaluate any support platform.

Evaluation checklist

  • Clean handoff to a person

    More than half of the conversations an agent replies in end with a teammate, so the agent has to pass on full context and the customer shouldn’t have to repeat themselves.

  • Approval before the agent acts

    A person can review a reply or an action before it reaches the customer.

  • More than one agent at once

    The built-in agent and your own can work side by side, each on its own work.

  • Change agents without migrating

    You can add or swap an agent without moving your conversations and workflows.

  • Scoped access

    Each agent can reach only the data and tools it needs.

  • Honest measurement

    You can see replied and resolved separately, for each agent.

Why scoped access matters: in IBM’s 2026 Cost of a Data Breach Report, roughly 1 in 5 organizations had an AI-related breach, and 92% of those lacked proper AI access controls.

Section 06

Your first 30 days

Give work to an agent in stages, one type of request at a time.

Stage 1

Internal note

The agent leaves a note for a person to review. The customer sees nothing.

Stage 2

Reply, stays open

The agent replies to the customer, and the conversation stays open for a person to check.

Stage 3

Reply and close

The agent replies and closes the conversation on its own.

A type of request moves up a stage only once it has a track record at the stage before.

If you’re starting out

  1. Week 1
    Fix your knowledge

    A gap in knowledge is the most common reason agents give for handing off. Fix the top gaps in your help content before you switch anything on, and pick one person to keep answers current.

  2. Week 2
    Start inside your team

    Turn on an internal assistant for drafts and triage, with a person approving its work. Note which types of request it gets right.

  3. Week 3
    Switch on one agent, on one channel

    Pick the channel where your routine questions arrive, and have a person check each reply. Teams in our data start on chat (45%), email (37%) or Slack (17%).

  4. Week 4
    Measure, then decide what moves up

    Track replied and resolved separately. Expect more than half of the agent’s conversations to reach a teammate at first.

If you already run an agent

  1. Week 1
    Read your handoffs

    Sort last month’s handoffs by reason: help content gap, needs account data, action in your systems, asked for a person.

  2. Week 2
    Measure honestly

    Split your numbers into replied and resolved, for each agent. Compare them with what your vendor reports and with the ranges in section 4.

  3. Week 3
    Fix what your help content can fix

    Close the knowledge gaps from week 1, then move the types of request that are now reliable up a stage.

  4. Week 4
    Decide whether to build

    If most remaining handoffs need account data or actions in your systems, that's the work for your own agent or a vendor’s. Pick one type of request and run it alongside the agent you have.

5 steps to build a support agent you can trust

Resend built two agents this way and brought them to Plain. They wrote down how.

Section 07

What changes for your team

Agents have not shrunk teams. In most teams that turned an agent on, the number of people replying to customers rose over the following three months, and so did conversation volume. Plan for an agent to absorb growth.

The work per person stays about the same. Replies per person barely moved after teams turned an agent on. Teams with no agent show the same flat pattern.

A handoff saves less than you’d expect. Within the same team, a conversation the agent hands off takes about as many replies from a teammate as one no agent touched. The customer writes about 50% more messages along the way. The value sits in the conversations the agent resolves alone, so measure those and keep improving the handoff.

Section 08

What nobody knows yet

Our data answers a lot about how teams use agents, but not everything. These are the questions it can’t settle yet.

• What does an agent cost to build and maintain over time?
Published estimates are opinion.

• Does building get better results than buying?
Teams differ from each other far more than agent types do.

• Why do teams differ so much?
It could be their setup, the agent itself, how hard their questions are or how quickly they close conversations. Our data can’t separate these.

• Why do some teams stop?
Some teams that turned an agent on no longer run one.

• How does AI outside the inbox compare?
A chatbot on a website or in product docs never reaches this data.

That’s why we recommend starting with what works today, on a platform that lets you add or change agents later without migrating.

Section 09

Where Plain fits

Many AI platforms tie your agent choice to your infrastructure choice. On Plain you can start with our agents, connect a vendor’s or build your own, and change that mix without moving platforms.

Sidekick

Drafts and summarizes, with a person approving.

Customer Agent

Answers routine questions across your channels.

Bring Your Own

A vendor’s agent or one you built, connected to Plain.

All three run on Plain, so your conversations and workflows stay put when agents change.

See how AI-native teams run their own agents on Plain.

About the data

  • Who. Plain’s customers, December 2025 to October 2026, excluding demo and test accounts.

  • Who they are. Mostly fast-growing AI-native and developer-tool companies. This is a view of Plain’s customers and should not be read as the whole market.

  • What to know when reading it. Plain’s built-in agent is included on every plan, so the share of teams using it reflects price as well as preference. Teams that build differ from teams that buy in size and volume, so comparisons between them are not cause and effect.

  • How we count. A team has an agent in a month when an AI agent replied in 10 or more conversations that month. Custom agents are counted only when their behavior shows they are AI.

Get your support running on Plain in minutes.

Get your support running on Plain in minutes.

Get your support running on Plain in minutes.