Our read
Start with a slice of one story, not the whole story. AMD's HR agent spans SAP SuccessFactors, ServiceNow, SharePoint and Teams approvals, so borrow one lookup and one write action from it. Autodoc's order tracking and address updates are a smaller model. Both are vendor-published accounts, not validated scopes. Test the permission model and the handoff trigger in your own pilot.
Start with the action, not the chat window
Kore.ai describes its platform as a place to build and deploy AI agents for service and for work. Its homepage frames service as resolving customer cases from start to finish across channels and systems, and work as completing employee requests without handoffs or delays.
That framing makes the useful question not whether an agent can answer, but which system it acts in. The two customer stories answer differently. AMD uses the agent for HR transactions and approvals. Autodoc uses it for order tracking, address updates and refund checks. Treat them as sources of ideas, not templates: take the closest story and cut it down to one read action and one write action before widening anything.
Connections: the systems that are named
AMD's story names three systems: SAP SuccessFactors, ServiceNow and Microsoft SharePoint. It also describes an API-based orchestration layer for cross-system actions and process automation. The company has about 30,000 employees and roughly 15 HR contact-centre staff.
Kore.ai's marketplace lists building blocks for plugging agents into other systems, including ServiceNow, Zendesk, Freshdesk, Salesforce, HubSpot, Gmail and Outlook. It also lists templates such as a support ticket categorizer, which tags tickets in an ITSM tool for routing, and an escalation predictor, which reads ticket conversations for sentiment and risk cues. For a team on Zendesk or Freshdesk, those listings are a concrete place to start.
Autodoc's story does not name the order or payment systems behind its agent, so the connection list for a similar setup has to come from your own stack.
Guardrails: what ABL and the trace give you
The homepage describes Agent Blueprint Language as a typed, schema-driven language in which enterprises define agent behaviour, tools, guardrails, orchestration and handoff logic. The practical appeal is that the tools an agent may call and the conditions for passing a case to a person sit in one structured definition, not in prompts scattered across flows.
The same page describes StateTrace, which captures the execution path across handoffs, state, tool calls and context. For a security or HR-operations reviewer, that trace is the artefact to ask to see in a pilot, because it shows which tools the agent called and when it transferred.
On data, Kore.ai's privacy notice says customer data is used only to provide the services and not to train generalised models. If a customer picks a third-party model, integration or data source, information may go to that third party as configured, under its own terms. Model choice is therefore part of your data review. AMD's story adds that its deployment ran on AWS, with encrypted storage for interaction logs and Amazon Bedrock for the conversational layer.
Execution: what the agents actually do
At AMD, employees and managers complete high-volume HR transactions and get answers to common questions without waiting for a person. Answers are specific to role and region. Managers can approve transactions inside Microsoft Teams through the universal AI agent, which brings together approvals from HR, travel and expenses, and finance systems. The story also lists a roadmap to expand onboarding and live agent support.
At Autodoc, the agents handle order tracking, address updates, refund checks and product compatibility questions, and contain routine inquiries through self-service across 27 countries and several languages. The story says refund checks, not refund approvals, so do not assume the agent moves money.
The contrast matters for scoping. AMD's agent spans several back-office systems and includes an approval action, so it is the widest example here and a poor first scope. Autodoc's is described in terms of customer-facing lookups and routine requests. A sensible first flow takes a slice of either: for instance an order-status lookup plus an address change, or an HR answer plus one transaction.
Handing over to a person
AMD describes a moments-that-matter escalation: critical issues are escalated automatically to HR professionals, preserving human contact where it matters most.
Autodoc puts the weight on what the person has after the handoff. Its Agent AI module gives human agents real-time guidance, next-best actions, guided playbooks, live coaching and automated summaries. Autodoc reports that instant knowledge retrieval cut research time by 10 percent, a vendor-published figure. Handoff is two things: the trigger that sends the case, and the tools the person has when it arrives.
Limits, and a pilot shape worth trying
Both stories are published by Kore.ai. They describe what the agents do, not how permissions are set, which actions need approval, what counts as a critical HR issue, or what context a person receives at handoff. The figures are vendor-reported: AMD cites an 80 percent reduction in HR resolution time and a 50 percent self-service deflection rate; Autodoc cites a 30 percent reduction in agent workload and a 20 percent decrease in handling time within six months. They come from different companies and metrics and do not compare. Ten other named companies, including Eli Lilly and Deutsche Bank, are listed on Kore.ai's own pages; workflow detail here comes only from the AMD and Autodoc stories.
Editorial suggestion, not a vendor capability: pilot one read action and one write action, and include at least one case built to force a handoff. Then read the execution trace and check that the agent used only the tools you listed. Kore.ai says its Proof Concierge proves a use case before you commit, which is a natural setting for this test.
Method and limits
2026-10-20
This dated snapshot counts distinct companies in public customer evidence. CONFIRMED means a case study, named testimonial or press source; DECLARED means the company itself told usedby; OBSERVED means a vendor listing; SIGNAL means a job post. All four count, but their strength differs. Removed and unlevelled lines are excluded. Examples are a capped selection verified within 90 days. Vendor facts are sourced records checked within 60 days. Coverage is uneven and missing evidence is unknown. These figures measure neither market share nor product performance; capture dates are not adoption dates.
