As customer experience (CX) teams race to deploy AI agents, leaders are faced with the question: Do we build our own custom AI agents from scratch, or do we purchase an off-the-shelf platform?
Recent industry research highlights a striking disconnect. While a significant majority of CX leaders express a desire to build their own AI agents to maintain control, market data tells a cautionary tale. According to widely-cited MIT research, 95 percent of in-house AI initiatives fail. The hidden costs of custom development are high, with organizations shelling out on model API calls, infrastructure, prompt engineering, testing and data management.
Gartner forecasting indicates why the build-versus-buy question is becoming more urgent. As enterprises increase their use of GenAI models already embedded in software, alongside new AI agents operating across multiple workflows, Gartner expects this shift toward multi-step, agentic automation to drive a sharp rise in short-term AI model usage. The analyst firm has increased its 2026 outlook for AI model growth to 110 percent, adding an estimated $6BN in spending this year.
That points to a market where buying AI capabilities through established platforms will become increasingly common, even as organizations continue to assess where custom-built agents can deliver genuine differentiation.
The Allure of Building: Trust and Specificity
If the costs and failure rates of building are so high, why do so many CX teams still want to take the DIY route? According to Dhwani Soni, Global VP of Product Management and Design at 8×8, it comes down to two factors: trust and domain specificity. As Soni told CX Today:
“Off-the-shelf agents are built for the average workflows. However, most companies don’t have standard average workflows.”
“They have very specific terminology. They’ve got escalation paths, regulatory constraints… They are not looking for off-the-shelf, but something that’s curated for their need and something that they can trust with their infrastructure.”
8×8 offers AI Studio, a tool that lets contact center teams build and deploy AI agents on the platform they already operate.
When businesses evaluate standard SaaS AI solutions, they often find them lacking the nuances they need to apply to their specific industry. But attempting to build a custom solution from the ground up to solve this problem introduces operational risk and delays.
From Off-the-Shelf Bots to Prebuilt Agents
There is a shift in the way vendors are packaging AI agents. Salesforce has launched Agentforce Help Agent, a pre-packaged autonomous service agent built on the Agentforce 360 Platform. Unlike earlier agent-building approaches, where companies had to connect their own knowledge sources, define actions and wire up channels themselves, Salesforce says the Help Agent comes with guided setup, Salesforce Knowledge grounding, pre-packaged actions and deployment across voice, web portal and messaging from a single screen.
The release indicates that the “buy” side of the market is heading away from one-size-fits-all bots toward configurable agents that come with the core plumbing already in place.
The Pragmatic Middle Ground: Buy the Commodity, Build the “Secret Sauce”
For enterprise leaders deploying these systems, the “build vs. buy” debate is rarely a binary choice. Instead, it’s about strategic allocation of resources.
Simon Ellis, Head of AI Transformation and Enterprise Architecture at Pets at Home, views the dilemma through the lens of a technology maturity curve: don’t spend money building what someone else has already perfected.
“I’m going to buy the commodity if the platform’s there and it’s good, and you’ve got partners who have spent hundreds of millions, if not billions, investing in building something. You take that, and then you build your secret sauce—what makes you different.”
At Pets at Home, this means relying on established platforms like Salesforce for standard customer service and veterinary B2B support. But when it came to the company’s highly specific consumer-facing digital pet care platform, the clear choice was to build.
“We built our own digital pet care platform, so that is where we have built differentiation, because that’s unique to us. We can’t go out and buy a pet care platform. It doesn’t exist,” Ellis told CX Today. “If we wanted to build all that [foundational infrastructure], you’re talking hundreds of millions and years, so buy the best and bring them together.”
The Hidden Risk of Building: The Compliance Trap
While building custom agents offers control over the “secret sauce,” it also places the burden of legal compliance squarely on the enterprise’s shoulders. And as recent research from Aithos shows, this is a burden most companies are not prepared to carry.
In testing leading AI models against European regulations (including GDPR and the EU AI Act), European research non-profit Aithos found that every major model failed compliance checks in realistic workplace scenarios. Nadia Kadhim, Executive Director at Aithos told CX Today that the assumption that using a “name brand” model guarantees compliance is a dangerous one.
“Where teams are most exposed without realizing it is the application layer. So that’s the part that the business actually controls,” Kadhim said. “There’s an assumption that if you pick a reputable name-brand model, that the compliance should be sorted; it’s probably fine. Assumptions like these can be really, really dangerous… even the best performing model broke the law in over a third of cases under realistic pressure.”
Kadhim emphasized that whether an organization builds or buys its AI agents, it retains the liability in deploying the AI. But building a custom agent from scratch means engineering compliance guardrails that established platforms have already spent millions developing.
“In the last 10 years in privacy and compliance… what I’ve seen mostly is an assumption that we can outsource responsibility and therefore we can outsource liability. That is not the case,” Kadhim noted.
“It is always the organization that deploys a system, builds an agent, or even just buys an agent from another company; there is still responsibility to make sure that its behavior is legal.”
The “Buy” Caveat: Native vs. Bolted-On Infrastructure
When an organization chooses to buy a foundational AI platform, the key to success lies in integration. Soni warned against “bolted-on” AI solutions that sit on top of legacy infrastructure.
“You add an AI layer on top of a platform and you immediately start inheriting the integration debt,” Soni said. “You’ve got transcription intermediaries, API latency; you’ve got context loss, especially when a call transfers. Most of these agents demo fine in a controlled environment, but when you start deploying them at scale, they start to show the break points.”
To avoid this, Soni advocated for native infrastructure, where the AI is embedded directly into the platform, giving the large language model (LLM) direct access to real-time voice data and the full interaction thread.
When to Build vs. When to Buy
How should CX leaders make the final call on where to draw the line? The decision ultimately comes down to whether the AI agent is a core product, or simply a vehicle to deliver better customer service.
When to Build (The “Secret Sauce”):
- The agent is a competitive advantage: The AI’s proprietary reasoning patterns and data signals represent core intellectual property.
- Deep domain-specific logic is required: Dealing with highly specialized workflows like financial risk modeling, complex clinical protocols, or bespoke supply chain optimization.
- Highly unique workflows: Operational processes (like a bespoke digital pet care ecosystem) are entirely unique to the organization and cannot be mapped to standard industry templates.
When to Buy (The “Commodity”):
- Time-to-value is critical: Teams need rapid deployment and must demonstrate business impact in weeks, not years.
- The workflow is common, even if the business is not: Use cases such as service requests, account inquiries, case management, appointment scheduling, order updates and FAQs are increasingly supported by pre-packaged platform agents.
- Teams want the plumbing already built: Platforms such as Salesforce’s Agentforce Help Agent bundle knowledge grounding, workflow actions, omnichannel deployment, testing tools and escalation into the product.
- Lack of dedicated AI infrastructure teams: The organization does not have the internal resources to sustain, maintain and optimize LLM infrastructure long-term.
- Compliance and security: The business operates in a regulated industry, requiring a platform to enforce constraints and compliance out of the box.
Keeping the Human at the Helm
Whether an organization chooses to build or buy, the technology must ultimately serve the human experience, Ellis said.
“Technology done well is absolutely seamless… Our colleagues are the most important part of our business. The empathy, the care, the showing up… we’re very much focused on how we can use AI to help our colleagues elevate, free them up from some of the grunt work, so that they can focus on the pet and the customer.”
Soni echoed this sentiment, advising teams to design their AI workflows with the human handoff in mind from day one. “Most early failures honestly do not happen in the AI interaction, but it happens when AI says it can’t help.”
