Smarter AI agents keep failing at real customer service — not from bad reasoning, but from fumbling paperwork and handoffs between systems.

This explores what AI agents need to be able to do, beyond raw intelligence, to actually deliver services that work for the people receiving them. The collection keeps returning to one point: the capabilities that matter most for human-centered service mostly sit outside the model. A smarter agent alone doesn't get you there.

Start with the evidence that capability isn't the bottleneck. In a test of frontier agents on 135 healthcare administrative workflows, the best one got 82.8% of individual steps right but finished only 36.3% of whole tasks. Most failures came from handling documents and coordinating across systems, not from reasoning Can AI agents reliably complete healthcare administrative workflows?. A historical look from GPS onward reaches the same conclusion at a larger scale. Agents stall when five conditions are missing: they must create value, adapt to the person, be trustworthy, be socially acceptable, and fit common standards. Missing intelligence is rarely the problem Why do capable AI agents still fail in real deployments?. Several notes explain where reliability comes from instead. Reliable agents move memory, reusable skills, and interaction rules out of the model and into a surrounding software layer (often called a harness), so the model doesn't have to work them out again each time Where does agent reliability actually come from?. Dubberly describes a similar stack: capability layers (chatbot, then system prompt, then RAG), plus layers of supervisor agents that watch over the others How should agent capabilities be organized into functional layers?. Some service work also needs several agents. Tasks that call for different kinds of expertise, parallel work, or independent checking go beyond what one agent can organize Do single agents always hit organizational limits?.

The human-centered part is mostly about who holds control and responsibility. Designers are settling on an "autonomy slider" that people set per task: suggest only, draft, or execute Should agent autonomy be a continuous slider or fixed levels?. A less obvious idea: some agentic systems may need to hand work back to humans, assigning them tasks and responsibilities. That reverses the usual direction of delegation Can AI systems delegate tasks back to humans?. Accountability has to be built into how work actually moves, too. In multi-agent tests, labeling one agent as "accountable" changed outcomes only when the workflow actually passed the deliverable through that agent Does accountability placement actually shape multi-agent outcomes?.

Service also involves more people than the user. A service agent affects staff, other customers, and third parties, so one line of argument says alignment should go beyond loyalty to a single user. It should balance different interests, keep actions visible and traceable, and leave developers and deployers liable for harms Should AI agents serve one user or balance many interests?. That depends on infrastructure that traces actions back to specific agents, shapes how agents interact, and detects and fixes harm What infrastructure must exist beyond training agents?.

One surprise: the environment can adapt to agents without hurting people. Adding agent-friendly cues to interfaces raised agent task completion with no measurable loss in usability for humans Can interface design help agents complete computer tasks better?. This changes the design job. Salesforce's Kat Holmes names experience architecture, agent collaboration, conversation design, and context maps as the skills teams now need What design skills do teams need for the AI era?. Taken together, the notes suggest that a human-centered service agent depends as much on the system and the design team around it as on the model.