Don't tie your business to a single AI provider

AI workflows are tied to model behavior, prompts, tool calling, embeddings, safety filters and output pricing. Switching models can change results even when the APIs look similar. A company without evaluation can't tell whether quality has improved or been damaged, so it tends to stay with the existing vendor out of fear of retesting.

Use small, fast models for classification, extraction and simple drafts. Use reasoning models for complex analysis, specialized models for images, audio or specific domains, and on-prem or private models when the data or latency calls for it. Don't make the most expensive model the default.

Route by risk, complexity, language and budget. Don't let agents pick models for themselves without a policy. Have a fallback for when a provider goes down or hits a rate limit, and record which model was used for each outcome.

Separate the business workflow, prompts and instructions, model gateway, knowledge, tools and observability. Use a common interface only for the capabilities you need. Hiding every special feature can cost you their benefits, so contain vendor-specific features within an adapter instead.

Keep source documents, metadata, evaluations and business rules outside the model system. Use open formats for export, and put prompts and agent configuration under version control. Credentials live in a secret manager and are never embedded in the workflow.

Build a model gateway for authentication, routing, rate limiting, logging, redaction and cost. It lets you switch providers piece by piece, but watch that the gateway doesn't become a new lock-in: make sure its configuration can be exported and that you have an internal API standard.

Build a dataset from real cases covering standard cases, exceptions, Thai language and risk. Define metrics such as accuracy, completeness, citation, policy compliance, latency and cost. Use human review for the parts that need judgment and automated checks for the rules.

Before switching, run the candidate model in shadow mode alongside production. Break the results down by segment instead of relying on a single average score. Run a canary on a small share of traffic and keep a rollback version. Every change needs a decision record explaining why it was chosen.

A cheap model may need so many retries or reviews that it ends up costing more. Count the model, tools and APIs, infrastructure, human review, errors and downtime. Calculate cost per successful outcome and cost per business unit.

Use a budget per workflow, caching, batching and prompt and context optimization. Set alerts for drift in cost per outcome. Don't lower the quality of important work to save a little money. Model a scenario at 10 times the volume, because agentic workflows may call models several times per transaction.

Keep credentials and billing in the company's name, never in an account the development vendor fully controls. Check model and open-source licenses and any commercial restrictions. Run an exit drill once a year for critical workflows.

Tier your work by criticality. Low-priority work can wait. Customer-facing work needs a fallback provider or a manual route. Financial work must fail closed and never release answers from a model that didn't meet the criteria. Preserve the queue and idempotency for when the service comes back.

Test the failure modes: timeouts, partial output, wrong tool calls, price spikes and regional outages. The dashboard must separate provider issues from data or workflow issues. Be open with customers if the service is running at a reduced level.