Start with a repeated workflow whose inputs and results can be inspected and reviewed by an owner. Fixed calculations or approval rules may suit deterministic automation; classification and summarization of varied text may be candidates for AI. Record the current time and errors, then compare a narrow trial against that baseline.
[01]How do you decide whether AI is needed?
| Example task | Approach to assess first |
|---|---|
| Calculate quotes using fixed prices and tax rules | Explicit calculation and validation rules |
| Extract intent and fields from inquiries | AI classification and extraction, followed by format validation and review |
| Draft answers from internal policy documents | Permission-aware retrieval, source references and answer review |
| Synchronize approved order status | API integration and retry handling |
[02]What data and access should you prepare?
- Real input examples with human-checked expected results.
- Test material covering errors, missing information and exceptions.
- Data owners, current versions and update responsibilities.
- Per-user access boundaries and permitted transmission to external providers.
- Provider APIs, service accounts, request limits and cost checks.
Retrieval-augmented generation (RAG) does not automatically make internal-document answers correct or secure. OWASP notes that RAG does not fully mitigate prompt injection. Assess application-enforced permissions separately from model output, and consider approval steps for consequential actions.
[03]How do you design a narrow pilot?
An illustrative pilot could extract a company name and request type from an inquiry and show a draft to an operator. Let the operator compare it with the original and edit it before extending the workflow to automatic sending or contract changes. This is a hypothetical scope example, not a customer outcome.
- Scope: one inquiry type and one operating team.
- Review: record omissions, misclassifications, unsupported outputs and correction time.
- Fallback: route missing evidence or required fields to an operator.
- Decision: use agreed test results to expand, revise or stop the pilot.
[04]How should you measure value and running costs?
Compare handling time, review and correction time, errors and exceptions for the same task type. Time saved can be calculated as baseline effort minus automated handling, review and correction effort. Convert this to money only using observed workload and an appropriate hourly cost.
Running costs may include hosting, storage, retrieval, integration tools and maintenance as well as model or API usage. Estimate volume and the exceptions operators must handle. Plan to recheck the same test examples when models or source documents change.
[05]Which workflows can you discuss with BOYD?
Discuss inquiry classification, document processing, internal knowledge retrieval and integration with business systems. A consultation starts with your workflow, data and account readiness, separating AI tasks from deterministic processing. Feasibility and cost depend on provider APIs, permissions and agreed scope.
Sources and references
- AI automation
- Readiness
