Core concepts
An employee gives a job a name and a place on your team. Its application brings together the agents that do the work, the knowledge they use, and the people who can work with them. You prepare the employee, test it against real examples, then publish it for chat. Channels and routines determine where and when it works.
Employees
An employee is an Assistant application with an employee profile: a portrait, a job, and a selected agent for working chat. Its name, category, and persona come from the same application. The employee ID and application ID identify the same record. An agent has its own ID.
Start with a role, describe the work, attach a brief, or record one. A role supplies detailed instructions and checks. Your own notes add the policies, sources, and requirements that make it useful for your team. The preparation conversation helps build the workflow, connect its tools, and save expectations for testing it.
For example, Robin might review invoices against your purchasing policy. Robin is the employee. The application holds the finance knowledge and access settings. The agent checks required fields, calculates totals, and decides when a person should review the invoice.
Employees and Applications show two views of that same setup. Employees focuses on preparation and readiness. Applications exposes the underlying workspace and its other features. An existing Assistant application does not become an employee automatically. Its employee profile must exist for it to appear in the employee list.
The shared company category catalog organizes both views. Categories such as Finance or Legal help people find related work. They do not grant access. A persona controls tone and voice, while the workflow and instructions define the job.
Start here: Create and prepare an employee.
Applications
An application is the workspace that connects knowledge, agents, members, settings, and channels. An employee uses the Assistant type. Applications can also serve other purposes:
| Type | What people use it for |
|---|---|
| Assistant | Ask questions and work with agents in a conversation. |
| Search | Find documents and relevant passages without a conversation. |
| Document Insights | Extract fields from many documents into a structured table. |
| Recommendations (beta) | Find related items using your knowledge and criteria. |
An Assistant application can answer questions from connected knowledge before you add a custom workflow. An employee adds a preparation experience and a named job to that application. One employee can use several cooperating agents, and an ordinary application can also have several agents.
Application membership controls who can use or change the workspace. Connected collections define its knowledge sources. Check both when someone cannot find a document or use an employee.
Start here: Create an application. For a document Q&A walkthrough, use the Quickstart.
Agents
An agent is a workflow that carries out a task. It can retrieve evidence, ask for missing information, make a decision, call a tool, create an output, or hand work to another agent. You can build it in an employee's preparation conversation, inspect its steps in the visual builder, or maintain it in an agent project.
Evaluation criteria, also called reasonings, help decision nodes evaluate an answer or condition. An invoice workflow might check whether the purchase order is missing before choosing its next step. Model-based decisions can vary, so test the paths that matter with representative inputs.
Several agents can cooperate inside one application. A main agent might ask a specialist to check contract dates and another to verify amounts, then combine their results. The employee's selected agent is the entry point for working chat.
A saved draft is still preparation work. Publish for chat makes the selected agent's published version available for working chat when the agent is enabled. Passing expectations does not publish it. Review and test later draft changes before publishing those changes too.
Start here: Create an agent and test its behavior.
Libraries and collections
A collection holds related files, such as purchasing policies or signed contracts. A library groups collections. Connect the collections an application needs, then wait for file processing to finish before expecting their contents in answers.
Collections support documents, spreadsheets, presentations, images, video, and audio. You can upload files or sync them from Google Drive, SharePoint, Dropbox, OneDrive, and Box using connectors. Processed spreadsheets can also become datasets that agents query.
A creation brief explains the job. A message attachment supplies context for that conversation. Use the knowledge area for sources the employee should reuse in future work. These inputs serve different purposes, even when they contain the same file type.
The knowledge summary shows connected files, processing status, and known page counts. These totals describe available sources. Check retrieval and citations to verify that the employee uses the right evidence. Permission to search a collection and permission to send its original files are separate application settings.
Start here: Create a library, upload and process documents, or search knowledge from code.
Connections and channels
Connections give an employee access to sources and tools. Channels give people a place to use it.
| Connection | Purpose |
|---|---|
| Knowledge connector | Sync files from a document service into collections. |
| Website access | Let an agent use a website with the configured access and login. |
| MCP provider | Give the application tools from another service, such as a CRM. |
| Channel | Let people work with the application through a web widget, WhatsApp, Slack, or Teams. |
Use Website access or MCP connections in employee preparation to set up those tools. Credentials belong in the connection form. Saving a provider and completing its sign-in are separate steps. Verify the connection before relying on its tools.
MCP also works in the other direction: a connected AI client can use Docana's tools to create or manage employees. That is different from adding an external MCP provider to an employee. The MCP documentation explains both connections.
Choose the Web Widget, WhatsApp, Slack, or Microsoft Teams for the people who need access. Your code can also call the API directly.
Routines
A routine is a saved prompt with a trigger. It can run an agent on a schedule, when documents arrive, or when another system calls it through the API.
For Robin, a routine might check new invoices every Friday at 9 AM. The saved prompt defines the work, the schedule includes a time zone, and the routine's enabled state determines whether it runs. Publishing for chat does not by itself activate a routine or connect a channel.
Check the saved routine when you want to know an employee's recurring responsibilities. Check work history to see what actually ran.
Expectations, work history, and performance
These views answer different questions about an employee:
| View | What it tells you |
|---|---|
| Expectations | Whether the tested workflow met the requirements in saved examples. These use the agent's test cases and evaluation results. |
| Work history | Which executions ran, their status, duration, and available execution details. Open the logs to inspect their steps. |
| Performance | Measures extracted from observed work, such as invoice amounts, review outcomes, or cases needing help. |
A finished execution means it completed without a recorded runtime error. Inspect its answer and evidence to determine whether the work was correct. Likewise, a passing expectation describes a tested example. Run tests again after changes to the workflow, knowledge, or requirements.
Conversation insights powers the performance measures you define. Its settings, including sampling, determine which conversations produce data. Document Insights is a separate feature for extracting fields from files.
Start here: Inspect an employee's results or run agent tests in CI.
Skills and templates
Skills give an agent instructions for a capability, such as creating a PDF report or chart. New employees start with skills and sandbox code execution enabled for these outputs. The deployment must have the required sandbox credentials. Test that the employee creates a real file, checks its content, and returns a working download link.
A document template provides reusable content and structure, such as fixed contract text mixed with prompts that fill in details from knowledge. Email templates do the same for messages. Use a template when the output needs an agreed format.
A role starter serves a different purpose: it supplies the employee's initial job instructions. Selecting an invoice review role does not select a document layout.
Start here: Create a document template.
How it all connects
Robin's invoice review brings the pieces together:
- Choose an invoice review role and add your team's purchasing rules during employee creation.
- Connect the policy collection and any tools Robin needs. Prepare an agent that checks fields, totals, and review thresholds.
- Test complete invoices, missing information, and cases needing a person. Inspect the result and any generated report.
- Publish the selected agent for chat. Set up channels or activate a routine when those are part of the agreed job.
- Use work history and performance to inspect actual work, then improve the draft and test it again.
You can manage this setup through the Employee API, CLI, and MCP. Employee export and application export use the same portable package. It preserves definitions and references to the original sources. File contents, credentials, conversations, and access grants stay in the source workspace. An imported employee starts as a draft so you can reconnect sources and test it before publishing.
If you came from docana.com
An AI workforce is the employees and automations you manage in Docana. The product uses more specific names for their parts:
- Super employee describes an employee whose work can span several agents and tools.
- Corporate Knowledge refers to libraries, collections, and the sources connected to applications.
- Workforce Builder describes creating and preparing the work. The employee preparation conversation and the agent's Build tab provide two ways to do that.
- LLM evaluations includes criteria used in workflow decisions and tests that check model-generated results.
Use Employees for the full preparation walkthrough. The feature pages in this documentation explain the individual parts, and the API Reference gives their request and response contracts.