You already have Copilot. Why would you need anything else for translation and review?
If your firm has rolled out Copilot, you have probably already tried it on a research report. Paste in a Japanese draft and ask for English, and you get English. Paste in a draft and ask what a compliance reviewer might object to, and you get a list. For a few minutes it looks as though translation and review are solved, and the obvious question is why you would buy a specialist tool for either.
It is a fair question, and this post is our answer. The short version: generating a plausible translation or a plausible list of concerns was never the hard part. The hard part is doing it the same way for every report, every analyst and every week, against reference data someone keeps current, with a result the firm can stand behind in front of a regulator. A general assistant is built for the first problem. Our products are built for the second.
What a general assistant gives you, and what it hands back
A general assistant is designed to be good at anything you type into it. That is its strength, and it is also why it is a poor fit for a regulated process. Everything that makes the output reliable has to come from the person typing.
In practice, three things end up back on your desk:
- The prompt. Someone has to work out how to ask for a translation that keeps the house style, or a review that covers the right rules. That knowledge lives in one person’s head, or in a shared document that drifts. Two analysts asking for the same thing get different results, and nobody can say which is right.
- The data. The assistant knows what is in the chat and what it can find in your tenant. It does not know your glossary, your past translations, how a particular issuer’s figures were handled last quarter, or what the issuer actually disclosed yesterday, unless someone keeps feeding it. Keeping that current is a job, and it tends to land on whoever set the thing up.
- The judgement about the output. A paragraph of fluent prose or a bulleted list of concerns is an opinion. It is not tied to specific sentences in the draft, it carries no reason, and nothing records what was checked and what was not. When a reviewer asks why a figure was let through, the answer is a chat transcript.
None of this is a criticism of Copilot. It is what a general tool looks like when you ask it to be a process. The firm ends up running a small, undocumented software project inside the research department, maintained by whoever is most enthusiastic, and it stops when that person is busy or leaves.
A workflow that runs the same way every time
Our translation editor and Second by Sentience are workflows, not prompts. The same steps run in the same order on every report, whoever submits it and whichever day it is. A translation goes through the same terminology, style and consistency passes. A pre-check runs the same set of review checks against the same rules. If two analysts submit the same draft, they get the same findings.
That consistency is not something your team has to build or look after. We design the workflows, test them against real research, and maintain them. When a rule changes, a disclosure format changes or a model improves, we update the workflow and every customer gets the change. Nobody in your research department writes prompts, trains colleagues on how to ask, or keeps a document of what works. The department gets a tool, not a project.
It also means the process is the same whether the person using it is the most experienced analyst on the floor or someone who joined last month. The quality of the result does not depend on who is typing.
The data underneath
A workflow is only as good as what it checks against, so the other half of the product is the reference data behind it. For translation, that is your firm’s terminology and style, built up from your own past reports and maintained as a controlled resource: changes are deliberate, versioned and applied to every job from that point on, not scattered across individual chats. For review, it is a maintained set of checks mapped to the rules a reviewer actually applies, together with the issuer’s own filings, so a figure in a draft is compared with what the company disclosed rather than with what a model remembers.
The point is who owns the upkeep. With a general assistant, the reference material is whatever someone last pasted in or saved to a shared folder. With our products, it is a database we maintain on your behalf, kept current as filings land and as your house style evolves, and under controls that say who can change it and when. Your team benefits from it without having to curate it.
The same applies to where your drafts go. Unpublished research is some of the most sensitive material a broker holds. Our products are built so that drafts are processed in a controlled environment set up for that purpose, with a clear record of what was sent where, rather than in a general assistant whose data handling is configured for the whole company.
One model with a prompt, or a harness of specialists
Under the surface, a general assistant is one model answering one request. It reads the whole draft, holds everything in mind at once and produces one answer. For a short email that is fine. For a report with earnings tables, forward-looking statements, named companies and regulatory language, it is asking one reader to be an expert in everything simultaneously, and the result is that some things get checked thoroughly and others get a glance.
Our products work differently. The workflow sits inside a harness: a controlled frame that splits the job into tasks, hands each task to a specialist agent with one responsibility and the reference data it needs, runs them in parallel, and then merges and grades what comes back. One agent checks figures against the source. Another checks terminology. Another looks for statements a compliance reviewer would flag. The harness decides what runs, in what order, with what data, and it records what each agent did.
For the person using it, the difference shows up in the output. Instead of a paragraph of opinion, you get findings attached to specific places in the draft, each with the check that produced it and the reason. A reviewer can see what was examined, what was passed and why. That is the difference between a second opinion and a review record.
We built the harness so that it can be reconfigured. The same frame runs a translation workflow and a review workflow, and it can take on new checks as rules or customer needs change, without anyone rebuilding the product. That is how the workflows stay replicable and how we can maintain them centrally for every customer at once.
Where a general assistant still belongs
Copilot is a good tool for the things it was made for. Summarising a long thread, drafting a client email, getting a first rough read of a document you have not seen before, or asking a quick question about a term. Most of our customers use it for exactly that, and we would not suggest they do otherwise.
The line is between individual productivity and departmental process. If the task is one person getting something done faster, a general assistant is often the right answer. If the task is something the whole department has to do the same way every time, with an audit trail, against maintained data, and with no one in-house on the hook to keep it working, you want a workflow built for the job. Translation of research and pre-review of research are both firmly in the second category.
Who is responsible next month?
The question to ask is not whether Copilot can translate a report or spot a problem in a draft. It can, some of the time, for the person who knows how to ask. The question is who is responsible for it doing so reliably next month, for everyone, when the rules have moved on and the person who set it up has moved desks.
With our products, that responsibility is ours. The workflows are designed and maintained by us, the data behind them is curated and controlled by us, and the harness that runs them records what it did. Your analysts get the result and your reviewers get the record.
If you are weighing a general assistant against a dedicated workflow for translation or supervisory review, we are happy to run both on one of your own reports and compare. Request a demo.
Gareth Edwards
CEO, Sentience Inc.