Moveworks: interview questions and learning guide

Conversational AI basics: intents versus LLM reasoning

Old chatbots matched keywords like a vending machine, while modern assistants read the sentence like a person.

Older bots used intents: a list of fixed goals trained with sample phrases, with slots to fill. They are predictable but break on new wording. Large language models read free text and reason about meaning, so they handle varied phrases and can plan several steps. Modern enterprise assistants mix both ideas: a reasoning layer interprets the request, plans steps and runs actions, with guardrails for safety. Predictability and flexibility must be balanced.

Interview tip: Explain the trade-off honestly: intents are predictable, LLM reasoning is flexible, and good systems add guardrails.

Enterprise search and grounding on company knowledge

A grounded assistant answers with the company handbook open on the desk, not from memory.

Language models can invent facts, so enterprise assistants ground answers in company knowledge. The assistant searches approved sources such as SharePoint, wikis and knowledge bases, picks the best passages and writes an answer from them, often with a link to the source. Search results must respect who is asking, so a person sees only what they may read. Good grounding depends on fresh, clear, well owned content.

Interview tip: Say that the quality of answers depends on the quality of content, and that grounding plus citations reduces made-up answers.

Connectors, identity and permissions

A connector is a key, and identity decides which doors that key may open for each person.

For an assistant to act, it needs connectors to systems such as ServiceNow, Workday, Salesforce and SharePoint. Each connector uses an authentication setup agreed with the system owner. The assistant must know who is asking, usually through single sign-on with the chat tool, so every request runs with that person's rights. Use least privilege for any service account, log what was done and review access regularly.

Interview tip: Stress that the assistant should never have more access than the person asking, and every action should be traceable.

Building custom agents: triggers, actions and plugins

Building an agent is like writing a job description: when to start, what steps to take and what to hand back.

Builders can create custom agents, also called plugins, in a low-code and pro-code studio such as Creator Studio or Agent Studio. A custom agent has a trigger, the kind of request that should start it, and actions, the steps it can run against systems. You describe inputs it needs, add instructions, connect actions and test. Start with a narrow, useful job and expand later. Always include checks before actions that change data.

Interview tip: Say that you start with one narrow use case, define inputs and failure paths, and test with real phrases before launch.

ITSM ticket deflection with ServiceNow

Deflection is a friendly guide at the entrance who solves small problems before anyone joins the queue.

In IT service management, many tickets are simple: password resets, access requests and how-to questions. An assistant connected to ServiceNow can answer these, run the fix, or create the right incident or request with details already filled in. This lowers ticket volume and speeds up users. Measure deflection carefully, because a closed chat is not always a solved issue. When it cannot help, a clean handoff with context saves agent time.

Interview tip: Quote real measures: deflection, resolution rate and handoff quality, and say you also check satisfaction.

Prompt and instruction design for assistants

Instructions are the briefing you give a new team member before their first shift.

The way you write instructions strongly shapes assistant behavior. Good instructions state the role, the task, the allowed sources, the tone and what to do when unsure. Clear examples help the model copy the right style. Ask for short structured answers and set limits such as no guessing about policy. Test with real, messy requests, change one thing at a time and keep a version history, because small wording changes can alter results.

Interview tip: Show that you treat prompts like code: version them, test them with sets of real questions and change one thing at a time.

Analytics, rollout and adoption

A rollout is like opening a new shop: you start small, watch the customers and keep fixing the shelves.

Launching an assistant is as much about people as technology. Start with a pilot group and a few use cases, announce it where staff already work, and give examples of what to ask. Track usage, resolution, satisfaction and failed questions. Failed questions show missing content or actions, so review them weekly. Share wins with leaders, train support teams for handoffs and grow in waves by team or region.

Interview tip: Use numbers: adoption, resolution and satisfaction, and show a loop of review and improvement after launch.

Security, privacy and responsible AI

A safe assistant is a good employee: it keeps secrets, follows the rules and admits what it does not know.

AI assistants touch sensitive data, so security and privacy come first. Control who can ask what through identity and permissions. Protect personal data, set retention rules for chats and log actions. Be aware of prompt injection, where hidden text tries to make the model ignore its rules, and keep risky actions behind confirmation and limits. Govern the assistant like any system: owners, reviews, testing and a plan for incidents.

Interview tip: Cover people, data and process: permissions, data handling and testing, then mention prompt injection by name.

Moveworks is an enterprise AI assistant for employee support and workflow automation. Employees ask in chat tools or portals and it answers from company knowledge or runs actions in systems like ServiceNow and Workday. It reduces tickets and waiting time.

An intent bot maps messages to fixed goals and breaks on new wording. A reasoning layer reads the meaning, plans steps and runs actions, so it handles varied phrases. It still needs guardrails and clarifying questions for safety.

Ground answers in approved sources, retrieve passages the user may read and show citations. Keep content fresh and owned. If nothing relevant is found, the assistant should say so and offer a handoff.

Pick one narrow use case, write a clear trigger, collect needed inputs and connect actions to the system. Add a confirmation step and failure messages, test with real phrasing and pilot with a small group.

Track adoption, resolution, satisfaction and unanswered questions, then fix gaps in a regular cycle. For security enforce identity and least privilege, log actions, protect personal data and test for prompt injection.

An AI assistant platform that answers employee questions and automates support requests across IT, HR and other systems inside chat tools.

A basic chatbot answers from a script. An action-taking assistant understands requests and completes tasks in connected systems.

Cover happy paths, wrong inputs, unclear phrases, escalation to a person and check each system action.

Intent is what the user wants. Entities are the details, like a date, name or device.

Resolution rate, deflection of tickets, user satisfaction, time to answer and how often people ask for a human.

Access control by role, no exposing private data to the wrong person, logging, retention rules and clear user notice.

Short, direct, one topic per article, plain words, current information and clear next steps.

Log it, find the source article or rule, fix the content or flow, test and monitor.