Internal AI agents can help employees find company knowledge, propose sales record updates, investigate incidents, draft content, develop software, analyze data, and monitor research. Start with a recurring task where you can give the agent a narrow set of sources and check its work before anyone acts on it.
These seven examples are hypothetical designs that your team would need to implement and test. Use an agent when the work requires choosing what to investigate or do next based on earlier findings. Tasks with a fixed sequence may suit conventional automated workflows.
Copy link to heading1. Answer company questions in Slack
Employees could ask a company knowledge agent how to request equipment or where to find the current travel policy. Give it approved internal documents to search, then have it return an answer with links to the passages that support it. If sources conflict, the response should identify the disagreement and direct the employee to the document owner.
If an employee asks whether they can book a hotel above the usual rate for a conference, the agent may need to consult both the travel policy and the exception process. Its answer should explain which conditions apply and any approval needed before booking, with links that let the employee check the guidance.
Start with material the intended audience can already access. Membership in a Slack channel does not establish permission to see every document an agent retrieves. Check both the requester's access and the audience for the reply before sharing an answer. During the pilot, ask policy owners to review answers for missing conditions and outdated guidance.
Copy link to heading2. Prepare sales operations updates for approval
After a customer call, a sales operations agent could compare the approved meeting notes with the account’s records in your customer relationship management (CRM) system. It would then propose field changes for the account owner, showing each change next to the current value and the notes that support it.
If the customer asked for revised pricing but didn’t give a decision date, the agent’s proposal could look like this:
The owner can approve the next-step change and follow up on the date separately. The review flag keeps the unchanged date from appearing newly confirmed.
Limit the pilot to selected accounts and fields. Before writing the proposed changes, get the owner's approval and check that the record has not changed since the proposal was prepared. Count the corrections the owner makes.
Copy link to heading3. Collect evidence for incident response
When an alert reports failed customer requests, an incident response agent could inspect monitoring data and recent deployments. Design it to follow the evidence, such as checking whether errors began before or after a release, and return a timeline with links to the underlying records. The on-call engineer should be able to distinguish observed events from the agent's explanation of their cause.
During a read-only pilot, the agent can gather evidence without restarting services or changing production settings. Keep those actions with the incident responder while you evaluate the investigation quality. Review whether the agent found relevant evidence and missed contradictory signals. Confidence alone does not establish that the agent identified the cause.
Copy link to heading4. Draft content from approved source material
Give a content agent a writer's request and approved source material, including the brief and the team's editorial rules. Let it find relevant material and ask for missing facts before drafting. The writer should receive a complete draft with sources for factual claims and clearly marked questions for a subject expert.
Give the agent access to the source collection and a draft destination. Reviewers should check factual scope and wording before approving publication through the existing editorial process. An approved source may contain information intended only for employees, so a draft for an external audience still needs that review. Track factual corrections separately from stylistic edits.
Copy link to heading5. Propose a software change for engineering review
Assign a software development agent a narrowly scoped issue, such as a form validation bug, with a clear description of the expected behavior. Have it inspect the relevant code and prepare a proposed change with tests. The agent can use test failures to decide which part of the implementation to revisit.
Start in a separate development environment with access to the repository the task needs. Give an engineer the code changes, test results, and any unresolved assumptions to review before merging or deploying. Passing tests supports the review but does not establish that the change meets every requirement. Include the engineer's review and repair time when assessing whether the agent helped.
Copy link to heading6. Analyze a dataset with inspectable code
Ask a data analysis agent to investigate a question such as why support volume changed. With an approved dataset, it could write and run code to compare request categories across the relevant periods. Require it to return the code and calculations with its findings so an analyst can reproduce the result. It should identify missing data and define categories such as repeat contacts.
For that support question, compare total tickets with tickets per active customer. More customers could explain an increase in tickets even if each customer needs less help. The output should state the comparison periods and show how reopened tickets affect the count. If customer activity data is unavailable, the agent should identify that gap before drawing a conclusion about why demand changed.
Begin with a limited dataset in a controlled execution environment, without credentials that allow changes to operational systems. Have an analyst check the definitions and calculations before the findings inform a business decision. Correct arithmetic can still produce a misleading result if the analysis counts the wrong group of customers. Compare the results against a known analysis during the pilot.
Copy link to heading7. Monitor research and report meaningful changes
Schedule a research agent to check an approved set of sources and compare new material with its previous reports. For an internal product team, that might mean reviewing standards updates and investigating which changes affect an existing integration. The report should link to the source material, distinguish new findings from repeated coverage, and explain the connection to the team's work.
Give the agent a record of source versions it reviewed and findings it already reported. If several articles describe the same standards change, it should connect them to the existing finding. It should report again when the underlying requirement changes or new evidence changes the earlier conclusion. Keep the source history separate from the agent's summaries so reviewers can check what changed.
For example, a report could identify a revised authentication requirement, link to the changed section, and name the integration that needs review. If the publication date changes but the requirement stays the same, that alone would not warrant another notification. Let the team define which changes deserve an immediate report and which can wait for a digest.
Define the source list and how far the agent may follow additional references. Have a subject expert review consequential interpretations before the team acts on them. Scheduling alone does not make a task agentic. If the job only fetches the same pages and reports changed text, a fixed workflow may be sufficient.
Copy link to headingHow do you choose and evaluate a pilot?
Choose work that recurs often enough to evaluate and produces an answer or artifact a colleague can check. Prefer a task with accessible source material and a clear stopping point, such as a proposed account update or a draft. Start with read-only access where possible. If the deliverable requires writing, keep the destination narrow and the result reviewable before it affects other work.
Run representative tasks through the current process to establish a baseline, then compare the agent-assisted process on similar work. Agree on acceptable results before the pilot begins.
Investigate failed cases before broadening access or adding users. When a pilot saves drafting time but creates more review work, it may need a narrower task or better source material.
Before expanding a successful pilot, define how you will govern the employee-built agent, including who maintains it and reviews changes to its access.
Copy link to headingWhich Vercel examples can you build from?
Vercel's eve incident response template gathers observability information from Datadog, GitHub, Vercel, and other tools. An investigation can start with a Slack mention, a message in a watched channel, or a request from an external system. The agent checks possible explanations against live signals and attaches source links so the responder can inspect the evidence behind its findings.
Investigations are read-only by default. Saved runbooks let the team describe procedures for recurring incidents that the agent can load on later matching requests. Saving and loading those runbooks requires private Vercel Blob storage. Use a recurring alert to evaluate whether the agent follows the procedure and returns evidence the responder needs before adding actions that change a system.
The eve content agent starts with a writer's request in Slack and reads Notion source material under the writer's own permissions. It loads style instructions for the requested format, such as a blog post or newsletter, and checks the draft for banned words. The writer can request revisions in the same Slack thread, where the session retains the conversation context.
The agent creates a page in a Notion drafts database and returns its link only after the writer explicitly approves the draft. Saving the draft does not publish it to a public site. When adapting the template, decide which editorial checks belong before saving the draft and which remain with the person preparing it for publication. Passing a word check does not establish that a claim is accurate or suitable for an external audience.
Copy link to headingFrequently asked questions
Copy link to headingHow do you choose a first internal agent?
Choose a recurring task whose result a colleague can judge against clear requirements. Favor a limited set of approved sources and an output you can review before use. Avoid expanding the pilot to several departments before you understand its errors and review workload.
Copy link to headingDoes every internal agent need write access?
No. An agent can investigate a problem or prepare a recommendation using read-only connections to business systems. Add write access when a defined task requires it, and limit that permission to the records or destination involved.
Copy link to headingCan an internal agent run on a schedule?
Yes. Research and monitoring tasks can start on a schedule without an employee sending a request. Define which sources the agent may inspect, what warrants a report, and when a person needs to review its findings.
Copy link to headingHow should a pilot be measured?
Compare the agent-assisted process with the existing process on similar tasks. Include review and correction effort alongside completion time, output quality, and operating cost. Examine failed and rejected results so the pilot's assessment includes work that never reached completion.
Copy link to headingWhen is a fixed workflow sufficient?
When you can specify the steps and decision rules in advance, a fixed workflow may be sufficient. Consider an agent when intermediate findings determine which source to inspect or action to take next. Running a workflow on a schedule does not change that distinction.