Inventory the work before listing AI features
Many enterprise initiatives begin with a request to build a chatbot that can answer questions. The chat interface, however, is only an entry point. An engineering team needs to understand the underlying job: what triggers it, what information the employee gathers, which judgment is applied, what happens next, and how completion is verified. Without this process definition, even a capable model will produce plausible responses that may not help anyone finish the work.
Use observable work units as the scope of the inventory. Examples include answering a customer inquiry, checking an order, summarizing a sales visit, drafting a maintenance recommendation, or assisting with an internal IT incident. Avoid treating broad functions such as customer service or operations as single processes. Their data, permissions, owners, and exceptions are usually too different to design and evaluate together.
- Trigger: Who starts the process, when does it begin, and does the input arrive through chat, a form, a document, a message, or a system event?
- Current steps: Which systems, fields, documents, and business rules does an employee use?
- Output: Does the process return information, create a draft, update ERP or CRM data, or notify another owner?
- Definition of done: Who verifies the result, and what evidence must be retained for audit purposes?
- Exception path: What happens when data is missing, rules conflict, or a connected system is unavailable?
Prioritize with value, feasibility, and risk
Good early candidates are usually frequent tasks with recognizable inputs, verifiable outputs, and affordable human review. Knowledge retrieval, summarization, classification, draft generation, and read-only queries across systems are often better starting points than autonomous decision-making. A workflow should be approached more cautiously when it depends heavily on undocumented expertise, receives inconsistent inputs, or can create material financial, compliance, safety, or customer impact when handled incorrectly.
Do not assess value only by counting manual steps. First determine whether the real bottleneck is information retrieval or content processing. If delays are caused by excessive approval layers, unreliable master data, or unclear ownership between departments, adding AI may simply move bad information faster. A useful design distinction is to separate four levels of assistance: presenting information, recommending an action, preparing an editable draft, and executing an action. Permissions can then increase only when evidence supports the next level.
For each candidate, record frequency, typical complexity, consequences of error, cost of review, acceptable response time, and the fallback procedure. The best first workflow is not necessarily the most visible one. It is often the workflow that can demonstrate data access, integration, evaluation, and governance within a controlled boundary.
Map authoritative data, system boundaries, and permissions
The quality ceiling of an enterprise assistant is often set by data governance rather than model capability. For every process, identify the authoritative source, data owner, update schedule, version policy, retention requirement, and permitted audience. If the same policy or customer status appears in formal documents, CRM notes, and personal spreadsheets, the project must establish which source takes precedence. Otherwise, a retrieval-augmented assistant can confidently combine outdated or contradictory information.
Separate read access from write access. Permission to check an order does not imply permission to change it. Permission to prepare a quotation does not imply permission to send it to a customer. Each tool or integration should define authentication, least-privilege access, allowed fields, confirmation requirements, audit records, timeouts, and failure behavior. For connections to LINE, ERP, CRM, cloud storage, or IoT platforms, teams should also verify API reliability, rate limits, data residency constraints, and whether the source system supports the required operation safely.
Authorization is more subtle than deciding who may open a document. An assistant can retrieve from several approved sources and infer sensitive information that no individual source reveals directly. Role-based access, data classification, personal-data masking, retrieval filters, and audit logs therefore belong in the initial architecture. They should not be treated as controls to add shortly before launch.
Design exceptions and acceptance criteria as part of the workflow
Because generated output is probabilistic, the inventory must describe more than the happy path. Define when the assistant must refuse, request missing information, show its sources, ask for confirmation, or hand the case to a person. Critical workflows can use deterministic checks for required fields, amount formats, product codes, customer status, or valid state transitions. If a check fails, the assistant should not proceed with a downstream system action.
Acceptance testing should use representative examples from real work and cover normal requests, ambiguous wording, insufficient permission, stale information, conflicting sources, prompt injection, and unavailable external systems. Evaluation should ask more than whether an answer sounds correct. It should verify that citations are traceable, the proper tool was selected, protected data remained protected, actions respected authorization, and a reviewer can understand and correct the result. After launch, capture feedback, failure categories, and human handoff points so the team can tell whether to improve prompts, retrieval, source data, integration logic, or the process itself.
The final inventory should be a process map, not merely a feature backlog. It should connect triggers, sources, decisions, system actions, permissions, exceptions, acceptance criteria, and accountable owners. When those elements are explicit, an integration team can translate them into a phased architecture and validation plan without making hidden assumptions about how the business actually operates.