AI for a Translation Agency: Projects, Glossaries, Deadlines

An AI agent for a translation agency works best as a project coordinator's assistant, not as a replacement for linguists. It reads incoming requests, checks them against each client's glossary and style notes, tracks deadlines across in-house and freelance linguists, and sends every quality decision to a human reviewer. The practical result is fewer missed details at the start of a job and fewer surprises at the end, with linguists still owning the words.

Key takeaways

• The strongest use of AI in a translation agency is coordination: intake, terminology lookup, deadline tracking and delivery preparation. It is not unsupervised translation.
• Glossary memory should be kept separately for each client and retrieved at the moment of work, so one client's terminology never leaks into another's project.
• Deadline tracking is a judgment problem as much as a calendar problem. The agent should flag risk early and propose options, while the project manager decides.
• Linguists and reviewers approve every term, every correction pattern and every delivery. The agent assembles evidence and never marks quality as passed.
• Start with one role, one client group and one observable success criterion, then widen only after reviewing real projects.

What does an AI agent do in a translation agency, and what does it leave alone?

Most of the work in a translation agency is not translation. It is everything around it: understanding what the client really needs, finding the right linguist, giving that person the right reference material, keeping terminology consistent across months of work, and delivering on time. That surrounding work is where projects quietly go wrong, and it is where an agent is most useful.

A well-scoped agent handles tasks like these:

• Reading a new request and building a structured project record from it.
• Asking the project manager's approved clarifying questions when something is missing.
• Retrieving the client's glossary, protected terms, tone notes and past reviewer comments.
• Preparing a briefing for the assigned linguist.
• Tracking milestones and warning the project manager when a deadline is at risk.
• Assembling a pre-delivery checklist for the reviewer.

It leaves other things alone on purpose. It does not decide whether a translation is good. It does not choose between two defensible renderings of a sensitive sentence. It does not send final files to a client without a person approving them. If your agency also uses machine translation with post-editing, that is a separate production decision with its own rules. The coordination agent can respect the outcome of that decision, but it should not be confused with it.

How should project intake work?

Intake is where small gaps become expensive. A request arrives by email, a client portal or a chat message, often with attachments and a vague deadline. The project manager then spends time working out what is missing before anyone can start.

An agent can do the first pass. It reads the request and the attachments, fills a standard project record, and lists what it could not determine. A useful intake record covers:

• Client and contact, matched to the existing client record.
• Source language, target language and regional variant (for example, Brazilian or European Portuguese).
• Subject area and intended audience.
• File formats and whether the files are editable.
• Deadline with a time zone, and whether it refers to first draft or final delivery.
• Reference material supplied, such as previous translations or style guides.
• Special requirements: certification, notarization, desktop publishing, subtitling timing.
• Confidentiality terms that apply to this client.

Where something is missing, the agent drafts a clarifying question and places it in front of the project manager. The manager edits and sends it, or approves it as written. Early on, no message should go to a client without that approval. After the team has reviewed enough drafts and trusts a narrow category (for instance, asking for the editable source file), you can decide to let that category go out automatically.

One security point matters here. Client documents are untrusted input. A file may contain text that looks like an instruction, whether by accident or by design. The agent should treat document content as material to read, never as commands to follow, and it should have no ability to take actions that a bad instruction could exploit. The [OWASP prompt injection prevention guidance](https://cheatsheetseries.owasp.org/cheatsheets/LLM_Prompt_Injection_Prevention_Cheat_Sheet.html) is a good reference for the design habits involved.

Agencies that handle confidential contracts will find the same intake discipline discussed in AI agents for law firms, where matter intake and deadlines carry similar weight.

How does glossary memory per client work?

Terminology consistency is the part of translation quality that clients notice most and that is hardest to maintain across people and time. A glossary file in a shared folder helps, but only if someone remembers to open it. The agent's job is to bring the right entries to the linguist at the right moment.

The memory itself should be a curated store, not a model trained on client text. Retrieval keeps the glossary editable: when a linguist corrects a term on Monday, the next project on Tuesday uses the correction. It also keeps the data deletable: when a contract ends, the client's memory can be removed without touching anything else. The trade-offs between retrieval and fine-tuning are laid out in the [AWS comparison of retrieval and fine-tuning](https://docs.aws.amazon.com/prescriptive-guidance/latest/retrieval-augmented-generation-options/rag-vs-fine-tuning.html), and for terminology that changes often, retrieval is the more practical fit.

A good glossary entry carries more than a pair of words:

• Source term and approved target term, per language pair.
• Status: approved, proposed or retired.
• Context note, such as the product line or document type where the term applies.
• Who approved it and when.
• Protected terms that must stay untranslated, such as product names.
• Style rules: formality, punctuation conventions, treatment of units and dates.

Separation between clients is the rule that matters most. Each client gets its own memory space, and the agent retrieves only from the space tied to the project at hand. This is the same principle explored in AI for engineering firms, where client walls decide what information may be seen together.

The agent can also notice things. If a delivered translation uses a term that is not in the glossary, or two linguists rendered the same source term differently, the agent proposes a glossary addition or flags the conflict. It does not resolve it. A linguist or terminologist approves, rejects or edits the proposal, and only then does the entry become active. This keeps the memory trustworthy, because every approved entry has a human name next to it.

How can an agent track deadlines across linguists and vendors?

Deadline tracking in a translation agency is rarely about a single date. A project has acceptance by the linguist, a first draft, a review pass, possible client review, and delivery. Each step depends on a different person, often in a different time zone, with a different workload.

The agent keeps a milestone record for every project and watches for signals that a date is slipping: a linguist who has not accepted, a draft that is late relative to the plan, a reviewer who is already carrying other work, a source file that changed after assignment. It also accounts for local holidays and working hours of the people involved, which is a common reason for a deadline that looked fine on paper.

When risk appears, the agent should propose, not act. A simple decision guide helps the team agree on this in advance:

• No response from a linguist after the agreed window: send the standard reminder and note it on the project.
• Still no response after a second window: alert the project manager with a shortlist of qualified alternates, drawn from the agency's own vendor records.
• Source files changed after assignment: alert the project manager and the linguist, and show what changed.
• Draft late but the review buffer is intact: tell the manager, and suggest no action yet.
• Review buffer at risk: ask the manager to choose between extending the client deadline, splitting the work or reassigning.

The agent never reassigns work or tells a client about a delay by itself. Those moves involve relationships and commitments that belong to a person. The same follow-up logic, applied to suppliers instead of linguists, appears in AI agents for procurement teams.

An illustrative scenario: a client sends a user manual for review on a Thursday with a Monday morning deadline. The agent builds the project record, retrieves the client's glossary and notes that the previous manual used a protected product name. It sees that the preferred linguist is on a public holiday on Friday in her own country and flags that to the project manager before assignment. The manager chooses a different linguist for the first half and keeps the preferred one for review. Nobody discovers the holiday on Friday afternoon.

How do linguists stay in charge of quality?

The boundary is simple to state. The agent prepares and reminds. People judge and approve. Putting that into practice requires a few firm rules:

• The agent does not edit translation text and does not mark any quality check as passed.
• The agent does not send deliverables to a client. A project manager approves each delivery.
• Glossary changes require approval from a named linguist or terminologist.
• Every flag the agent raises is a prompt for a human to look, not a verdict.

Before delivery, the agent can assemble a checklist for the reviewer: which approved glossary terms appear in the target text, which deviate and where, whether protected terms stayed intact, whether all files in the request are present, and whether the client's special requirements were met. The reviewer sees the evidence and decides. A flagged deviation may be a mistake, or it may be a deliberate choice for that sentence. Only the reviewer can tell.

Reviewer corrections are the most valuable data the agency produces. When a reviewer changes a term the same way several times, the agent can propose a glossary entry or a style note. That turns scattered corrections into shared memory, and it gives linguists a visible say in how the system improves.

To keep the agent honest, review it regularly. Take a sample of recent projects and compare the agent's intake record, glossary retrieval and risk flags with what the project manager and reviewer actually did. Where they differ, decide whether the agent or the process needs to change. Frameworks such as the [NIST AI Risk Management Framework](https://www.nist.gov/itl/ai-risk-management-framework) offer vocabulary for this kind of ongoing review, and the [Anthropic guidance on building effective agents](https://www.anthropic.com/engineering/building-effective-agents) is a useful reminder to prefer simple, observable workflows over clever autonomy.

What does a sensible first engagement look like?

Do not try to cover every client and every language pair at once. A first engagement defines one useful role, the memory and tools it needs, the approval boundaries and one success criterion you can observe. For a translation agency, a practical sequence looks like this:

• Step one: choose one client group with recurring work and a reasonably stable team of linguists.
• Step two: collect what already exists for those clients: glossaries, style guides, past briefs and reviewer comments. Clean it and mark what is approved.
• Step three: define the intake record and the questions the agent may draft for the project manager.
• Step four: set the boundaries in writing: what the agent may read, what it may draft, what needs approval, and what it may never do.
• Step five: connect only the tools the role needs, with read access first and narrow permissions after that. The [OWASP authorization guidance](https://cheatsheetseries.owasp.org/cheatsheets/Authorization_Cheat_Sheet.html) is a good model for least privilege.
• Step six: run the agent alongside the team, with managers comparing its records to their own.
• Step seven: pick a success criterion such as fewer intake questions asked late in a project, and review it with the team on a regular schedule.
• Step eight: widen to the next client group only after the first one behaves predictably.

Be Human designs, implements and operates this kind of setup remotely, in English and Portuguese. The managed side covers monitoring, recurring evaluations, integration maintenance, permission review and incident response, and improvements are introduced in a controlled way, so changes to the agent are reviewed rather than silently applied. The agency keeps control of the glossaries, the approval rules and the permissions.

If your agency also manages creative or localized marketing work, the approval patterns in AI for creative studios show how briefs and sign-offs can be tracked in a similar way.

Frequently asked questions

Does an AI agent replace project managers or translators? No. The agent takes over repetitive preparation and tracking, such as filling intake records, retrieving terminology and watching milestones. Project managers still make commercial and relationship decisions, and linguists still produce and approve the translation.

Can the agent also translate the documents? It can be connected to translation tools if the agency chooses, but that is a separate decision with its own quality rules, client consent and review requirements. The coordination role described here works the same way whether a project is translated fully by people or with post-editing, as long as a linguist remains responsible for the result.

How do we keep client documents confidential? Give each client its own memory space, grant the agent only the access its role needs, and keep a record of what it read and did. Treat document contents as data, never as instructions, and agree with clients on retention and deletion. When a contract ends, the client's memory can be removed independently of everyone else's.

What should we measure to know it is working? Pick something your team can observe directly, such as how often a project manager has to ask the client for basic information after assignment, or how often a reviewer finds a glossary term that the agent should have surfaced. Review a sample of projects on a regular schedule and adjust the agent's instructions, memory or permissions based on what you find.

Be Human editorial guidance. Examples are illustrative unless explicitly attributed. Sources provide background; implementation choices depend on your workflow.