AI Agents in Customer Support: Build vs Buy and What to Automate First

Support is the function where AI agents have the clearest business case and, at the same time, the widest gap between vendor marketing and what teams actually experience once an agent goes live. This article covers realistic performance benchmarks, where automation actually works well versus where it doesn’t, and how to think through building versus buying. For the outbound side of the house, see how AI agents are changing sales prospecting.

Adoption Is High, But the Numbers Hide a Gap

Salesforce reports 66% of service organizations were running AI agents in 2026, up from 39% in 2025, with 91% of CX leaders under executive pressure to deploy AI faster. Contact center AI usage is even broader at the surface level, with 88% reporting some form of AI in use, but only 25% saying it’s fully integrated into daily operations.

The gap shows up clearly once you compare vendor-published numbers to independent, enterprise-wide benchmarks. Individual vendors report strong deflection rates, Decagon around 80% across its customer base, Ada in the 70-80% range, Sierra near 70% at a specific customer. Zendesk’s enterprise-wide median across all CX programs, by contrast, is 41.2%, with a top quartile of 58.7%.

That gap isn’t necessarily vendors lying. It’s usually a difference between deflection, whether a query was contained without a human, and resolution, whether the customer’s actual problem got solved. A platform can show 90% deflection with only 40% true resolution, since those are two different measurements being reported as if they were one.

Where Automation Actually Works, and Where It Doesn’t

Performance splits sharply by the type of request, not by which vendor or model is behind the agent. High-structure intents with a clear backend system to check, like order status or password resets, deflect in the 65-80% range consistently. Sentiment-heavy and dispute-style intents, the kind involving frustration or a judgment call, stay in the 19-34% range regardless of vendor.

Structured, data-backed questions are the safest starting point. If the answer lives in a system the agent can query directly, order status, account balance, appointment scheduling, the agent’s job is retrieval and formatting, which is exactly what current agents are reliable at.

Emotionally charged or ambiguous requests are the wrong place to start. Refund disputes, complaints, and anything requiring judgment about an exception to policy consistently show the weakest automation performance, no matter which platform is running it.

The build-vs-buy decision follows from this split. Vendor platforms tend to be tuned for the common structured cases and get you to a working deflection number quickly. Building in-house makes more sense when the differentiator is a highly specific structured workflow tied to proprietary systems that a generic vendor agent won’t integrate with cleanly.

What to Automate First

Given the intent-level performance split, the practical starting point is narrower than “automate support.” Start with the highest-volume, highest-structure intents specifically, the requests where the answer already exists in a system of record and the agent’s job is lookup and formatting, not judgment.

Resist the instinct to prove value by automating the hardest tickets first. The data consistently shows dispute and sentiment-heavy intents underperforming across every vendor and model, which means starting there risks a bad first impression of the technology instead of demonstrating what it’s actually good at.

How Phinite Fits Into Build vs Buy

Phinite sits closer to the “build with managed infrastructure” end of the spectrum than either a narrow point-solution vendor or a fully custom in-house build. Multi-channel deployment across Slack, WhatsApp, email, and web chat is handled as a platform feature, so support agents can go live across whatever channels customers actually use without a separate integration project per channel.

Governance and observability – audit logging, role-based access, execution tracing – are built in from the start, which matters specifically for support, where an agent might be taking real actions like issuing a refund or updating an account. That combination lets a team build support-specific logic tuned to its own systems, without also having to build the deployment and governance layer that a narrow vendor platform would otherwise leave out entirely.

Key Takeaways

Support automation adoption is high, but enterprise-wide resolution numbers (41.2% median) are meaningfully lower than the deflection numbers vendors publish (70-90%), largely because the two measure different things. The clearest path to a good outcome is starting with high-structure, data-backed intents rather than the hardest emotional or disputed cases, where every platform currently underperforms. Whether to build or buy comes down to how differentiated your support workflows are and whether you need the deployment and governance layer a narrow vendor tool won’t provide.