When an agent is useful
Rule-based workflows work well with structured forms. Agents help when emails, attachments, policies or contracts are written differently each time and need several retrieval and drafting steps. A chatbot mainly answers a conversation; an agent completes a bounded task with approved tools.
Capabilities and approval boundaries
Classify requests, retrieve a policy or price list with citations, draft quotations and summarize conflicting sources. Do not let the agent decide payments, customer commitments, price changes or deletion. Grant minimum access: a quotation agent may read a price list, not edit it. If information is absent, say so; consequential outputs wait for a person.
Implementation and evaluation
Choose one repeated task and an operator who knows the correct result. Gather only the documents and permissions it needs. Build a test set with real situations and expected answers. Run in parallel with people, agree an acceptance threshold and adjust until it is met. Hand over the model account, instructions and test set; rerun tests when documents or models change.
Illustrative everyday tasks
A sales agent can prepare quotes from a supplied price sheet for human review. An HR agent can answer documented leave and expense questions and escalate missing policies. A reporting agent can assemble a draft with sources for each point. These examples assume usable written policies and records; they are not customer outcomes. Our own reporting agent first ran in week 39 of 2026.
Data handling and costs
The selected document passages go to an external AI provider. We use paid APIs from providers such as Anthropic, OpenAI or Google and share the selected provider’s terms before connecting data. Paid API data is generally not used for training by default under the applicable terms, but temporary abuse-monitoring retention may apply; free-tier terms can differ. Exclude sensitive data that the task does not need and log inputs, outputs and reviewers. Usage depends on reading and output volume. Set spending limits; enforcement may lag and small overruns can occur.
Tools, packages and preparation
Agents can work with Google Drive, Docs, PDF, Word, Notion, Gmail, Outlook, Sheets, Zalo OA and Telegram where access permits. Run them in n8n or a dedicated small service. Start with one task and test set, expand to multiple tools with logs and approval gates, then maintain through periodic evaluations. Bring a concrete task, existing documents and a reviewer who can judge quality.