Platform · AI Workflows
Agentic Workflows,
Human Direction
You decide where agents are used, and what they can and cannot touch. You determine what counts as reference, and the tone and terminology they follow. They run the repetitive passes, and only what genuinely needs a person reaches your team.
What Can Be Run
Three Workflows, Two Kinds of Agent
Two write into your content. The third only annotates it. Knowing which is which tells you what changes when a run finishes, and what is waiting for a person to look at. All three run inside the projects you already have.
AI Translation
Writes. Produces the first translation of your content, reading your memory and termbase for every sentence and following instructions you write in plain language. Installed as a machine translation system, so it runs wherever machine translation already runs.
Read moreAI Post-Editing
Writes. Brings an existing translation up to publishable quality. Built for updating fuzzy matches and applying your approved terminology, and for cleaning up machine translation output before a reviewer ever sees it.
Read moreAI Quality Assurance
Annotates, and changes nothing else. Findings land in the Editor quality panel beside your built-in rule results, each with a description and, where one exists, a suggested correction. The findings are the output.
Read moreThe Scope You Set
Agents Touch What You Allow, and Nothing Else
Each workflow classifies content by where it came from, so nobody sorts files first, and work your people edited is protected before you change anything. Below is what each does out of the box. Which of the nine content origins it may touch is yours to set.
AI Translation
Out of the box
- Works on
- Segments with no usable translation
- Never touches
- Anything pre-translated or human-edited
Customizable
- Which origins
- Any of the nine
- Your instructions
- Tone, brand, terms to keep
Runs as a machine translation system. Fills the gaps, and only the gaps.
AI Post-Editing
Out of the box
- Works on
- Fuzzy matches and MT output
- Never touches
- Human edits, 100% and 110% matches, and previous versions
Customizable
- Which origins
- Any of the nine
- Your instructions
- What to raise, what to leave
Runs as a supplier job on a workflow step. The two states that need a pass, nothing else.
AI Quality Assurance
Out of the box
- Works on
- Every segment, whatever produced it
- Never touches
- Nothing. A finding is free to ignore
Customizable
- Which origins
- Any of the nine
- Your instructions
- Refinements, not new checks
Runs as a QA profile rule or a supplier job. Annotates only, changes nothing.
How a Run Works
Agentic Workflows Run in Batches, so Each One Builds on the Last
An agentic workflow moves through your document in batches, and builds each one from your project as it stands at that moment. Because results are saved before the next batch is prepared, a decision made at the start of a long document is reference material by the middle of it, and the last page reads like the first.
- The next batch is prepared with reference material retrieved fresh
Step 1 · Batch
Take a Batch
The workflow takes a handful of segments rather than sending the whole document off in one block.
At this step
- Runs on text content in any format Wordbee can process
- Skips segments you told it to leave alone
- Locked segments are held back by default
Multi-Agent Workflows
Not Everything Needs Multiple Agents, and You Decide What Does
The choice is how many passes the agents make. One agent works in a single pass over the batch. Three agents add two more: the second reads every segment against the source and your references and flags what does not hold up, and the third rewrites only what was flagged. Those are the repetitive passes, and you set how far the agents go on each workflow you configure.
- 1 Agent: one pass over the batch, and nothing after it
- 3 Agents: a second reads every segment and flags what does not hold up, a third rewrites only what was flagged
- Translation and post-editing run the same three passes. Only the job of the first agent differs
- LQA is the exception, because LQA agents flag and never correct. A second agent adds meaning checks rather than a fix
- Configure the same workflow twice with different instructions, one for marketing and one for legal
Agents as Workflow Suppliers
Staff a Step, Not a New Workflow
An AI agent sits in a supplier pricelist next to your linguists, and gets assigned to a workflow step the same way. Assign the job, and the agent works the batch, completes the job, and triggers whatever comes next. No new workflow, no new structure, and no change to which provider your AI runs on.
- An agent is defined once as a supplier service, then assigned to any job
- Agents start when the job moves to In Progress, complete the work, and trigger the next step
- Any step can be fully automated, fully human, or a mix, decided step by step
- An agent appears only where you assign one, so a project you keep AI out of never gets an AI supplier or an AI translation profile
- Your own provider account and your own key, any provider, with different models on different steps of the same workflow
- Live progress while an agent works: the sparkle icon in job lists and job details opens the AI work status popup with a percentage
- Failures carry a flow ID and a request ID, so support traces the exact run rather than asking you to reproduce the failure
- Cancelled and Failed are distinct states, cancellation genuinely stops the run, and Completed means the work completed
Common Questions
What Buyers Ask About AI Workflows
See It Run on Your Own Content
Which workflows fit depends on what you publish, how much of it there is, and how much review it needs. We work that out with you first, then run a pilot on your material rather than on a demo. Already on Wordbee? Your account manager can add AI workflows to your current plan.