Almost one third of organizations are creating custom CX tools they can deploy for specific use cases, rather than buying from third-party vendors. When CX Network conducted its annual research into the state of CX in 2026, one third (33 percent) of respondents said they mostly acquire new AI capabilities for CX through in-house builds.
Traditionally, this has taken months of development work as coders collaborate with different stakeholders to build, test, and roll out tools at scale. Today, with sophisticated AI tools, that doesn't have to be the case.
One case study comes from Carparts.com, which coordinated expertise from across its CX, data science, technology, and marketing teams to build three new, AI-powered tools: a chatbot called Spark, a ticketing system, and a catalog.
Speaking to CX Network, director of CX Aurellia Pollet says: "Because of the complexity of our business, it was really hard to buy solutions like a catalog, for example. So, we built our own. Before AI, every time we wanted to buy new software to do something, we had to consider how it would plug it to all our custom systems. Finally, this has been levelled up with AI and we don't have as many constraints anymore. It's so freeing."
This article explores the reasons organizations are turning to in-house builds, the tools available to support their ambitions, the solutions they are developing, and how to get started.
Why are organizations building their own solutions?
No-code tech development – whether in CX or beyond – is commonly known as vibe coding and it's growing in popularity. It democratizes the ability to develop custom tools for specific use cases using natural language prompts, rather than code. As the Carparts.com example illustrates, business complexity is one reason organizations choose to develop their own tools – but it isn't the only reason.
"Many enterprise builders feel trapped between stiff legacy chatbots and generic LLM wrappers," says Nikola Mrkšić, co-founder and CEO of PolyAI. "They've resorted to building in-house because they want the reliability and precision of the former and the fluency and flexibility of the latter."
AI specialist Mrunal Gangrade says that in regulated industries in particular – banking, finance, insurance, and healthcare – the main reasons to build in-house concern security, trust, and model maintenance, particularly when it comes to customers.
"If you build in-house rather than buy, customers may trust your organization more," she says. "It is also easier to maintain and further enhance a model if it is developed in-house, rather than acquired from a vendor. Finally, the cost perspective. In the long run it is beneficial for any organization to build AI rather than take a third-party solution because the contract renewal and other outlay and operational costs add up. However, if it is in-house then it is easier to maintain, develop, and further enhance."
What tools are available to help organizations build their own AI?
While many use LLMs such as Claude Code, there are a growing number of tools on the market to help organizations that have the ambition – and in-house talent – to see such projects through.
Google has Opal for mini web apps and also allows developers to build apps in its AI Studio. OpenAI has Codex, a dedicated coding agent. Base44 – the viral platform whose founder is also known for talking about high token costs – was acquired by Wix in 2025 for US$80 million.
There are also CX-specific tools.
In May 2026, PolyAI added to this list by opening up its proprietary dialog model, Raven, to allow anybody to build a production-ready dialog agent that is capable of handling complexity, in less than 10 minutes. The tool is called the Agentic Dialog Platform and it makes it possible for anybody to build dialog agents for customer conversations. Trained on the "hundreds of millions of conversations" running through the PolyAI platform every year, Mrkšić says its capabilities really shine when applied to service edge cases and complex customer needs.
"Dialog is very challenging to get right. It's much more than responding to a customer in a way that sounds natural," he says. "It's really about your agent participating in a high-stakes conversation and delivering a resolution. Very few AI systems can handle dialog for arranging medical transportation or helping you through a gas leak, to name a couple of examples."
Mrkšić says there are two groups who may benefit from using the tool to develop thier own agents: non-technical builders, "who can now create production-grade dialog agents quickly with our Agent Builder", and developers who want to use the Agent Development Kit who want to build dialog agents using their own IDE, their own coding assistants, and their own workflows.
What are organizations vibe coding for CX?
Alex Levin, CEO and Co-founder of Regal says that in his experience, engineering teams are the earliest movers with coding agents – and CX teams are a close second, mostly due to the financials.
Levin says: "The financials are simple – engineering and CX are both seen as cost centers, and every efficiency from AI goes straight to the bottom line. What's interesting is that we aren't seeing people do mass layoffs in those two teams. Instead, they are slowing or stopping hiring, and using AI to take over more and more of the level 1 tasks so the humans can focus more on planning and differentiation.
"For instance, our customers use Regal to build their own AI voice agents and quickly iterate without needing to hire and re-train humans in the contact center," Levin continues.
How to get started with your custom CX tools
While it may be attractive to swerve the capex and opex costs associated with vendor-developed AI, building custom tools in-house is not free; time, resources, and testing all come with a price tag.
On the business side of things, Levin's top tip for getting started is to understand how much these costs are likely to run to. "First, measure token usage per person, and treat high usage as a good sign right now," he says. "Then start communicating a revenue per employee goal and how that would impact company growth, margin, and how much more you can pay staff, so employees are aligned on why this transition to add leverage is so important."
He adds: "Companies will gain more leverage by using AI to do more with their same team, rather than cutting the workforce. They'll be able to serve more customers, close more deals, and innovate more products by combining AI with the team's expertise."
It's also important to note that vibe coding custom tools will not eliminate token costs.
On this point, Gangrade says: "Try to build your knowledge locally, so there are fewer interactions that will incur token costs. If I search for ABC one day, then the next day I also search for ABC, that will be two token hits going out. However, nothing changed. So, if you have static data like that, build a repository and maintain that knowlege internally. That way, when I repeat the search, it doesn't utilize more tokens."
Employee training on how to get the most out of any tool – acquired or developed – is also important, particularly when it comes to something as potentially costly as tokens.
"Internal employees should be trained on how to maximize the use of AI by usage of less tokens. It is like building up your own agents locally and so that you can fetch the information from your local agents rather than utilizing the tokens which is going to exponentially increase the cost," she adds.
Finally, Mrkšić advises designing for "resolution instead of response".
He explains: "A good dialog agent goes beyond just answering questions. It has access to the information and systems it needs to actually close the loop."
Organizations can also select their build tools to match the capabilities of those developing them. Mrkšić says: "For example, your CX managers know the customer journey better than any engineer. On the other hand, your developers know how to solve truly complex problems. Give both groups the right tools: natural language for the non-technical builders, real code and control for the true developers."
Last but not least, Mrkšić says organizations should "stop brute-forcing dialog with wildly complicated prompts on general-purpose LLMs".
"It's slow, expensive, and falls apart under pressure. Use models built on real dialog to have real dialog with your customers," he adds.
However, vibe coding is not for everybody. At Carparts.com, dedicated technical teams were involved throughout. For those who take the vibe coding route, organizational culture and oversight must be sharp to avoid "technical debt", the DIY drawback, whereby real-world vulnerabilities and mistakes occur because the human instructing the AI tool, does not have the technical expertise to avoid or fix major errors.
"I would not recommend a small company to go into a venture like [ours]," Pollet says. "First of all, you need a team of data specialists. We had three data scientists working full-time on this project. You cannot teach yourself these skills on your computer. Also, sometimes an off-the-shelf solution will work perfectly if the business is not complicated. Maybe you sell t-shirts, for example. You may have a lot of SKUs, maybe lots of different colors and sizes. But for us, a single model of car can have multiple types of parts that differ by the year the car was made," she explains.
Grounding new AI tools in secure data
One of the most popular tools to build in-house is the chatbot. However, whether voice or written, developing and training these tools securely takes more than a few natural language prompts.
To be contextually relevant, conversational interfaces such as chatbots and voice agents must be grounded in the data organizations hold. One of the richest and most comprehensive sources of data is closed tickets from the contact center: these include the edge use cases many will oversee, and capture how customers typical converse.
However, according to Brad Murdoch, CEO of AI-help desk platform provider Deskpro, this is also a source of personally identifiable information (PII), including customer names, email addresses, account details and credentials, and sensitive case histories, for example information customers shared with the expectation it would stay private.
"Indexing this content without the right controls in place introduces two specific risks," Murdoch says. "The first is inadvertent disclosure where your AI surfaces information from a resolved ticket that the customer you're talking to now was never meant to see. There is no malicious intent, but this is a major compliance issue," Murdoch explains.
"The second is prompt injection. This is when a bad actor hijacks a prompt going from your CX or help desk platform to an AI foundation model running in the cloud and reframes the prompt to extract sensitive indexed data," he says.
In light of this, Murdoch says there are four primary controls that need to be in place before you the ticket archive is used as an AI knowledge source. "Skip any of them and you're taking on risk you may not be aware of until something goes wrong," Murdoch says.
The four primary controls are:
- PII redaction: Strip personally identifying information from ticket content before it reaches the model. The original ticket stays intact, while the model only sees a cleaned version.
- Establish access scoping limits: This means the AI is only drafting responses from tickets within their own department or line of business. Managers should have the ability to manually exclude specific tickets from indexation entirely. These controls keep the knowledge source relevant and limit what the model can surface.
- Establish a ticket timeframe: This ensures only index resolved tickets from within a recent window are used, keeping the knowledge source current and reducing the volume of older data the model is drawing on.
- Understand where data is held: Is your help desk deployed on a public cloud? If so, your data is likely being sent to public cloud-hosted AI foundation models over the internet. For most companies this is fine, but if you work in a regulated industry or need to comply with data sovereignty rules, you need a help desk and AI models that run in a virtual private cloud, sovereign cloud, or in your own data center. This keeps AI processing within your organization's security perimeter to meet compliance requirements.
Murdoch says: "If you set up the right data foundation, your ticket archive will become one of the most valuable assets in your AI arsenal but get this wrong and you're one prompt injection away from a big problem."