[{"data":1,"prerenderedAt":260},["ShallowReactive",2],{"news-item-\u002Fnews\u002Fputting-ai-to-work-in-the-it-communicator-role":3},{"id":4,"title":5,"body":6,"category":245,"created by":249,"date":250,"description":16,"extension":251,"meta":252,"navigation":253,"path":254,"sections":255,"seo":256,"stem":257,"thumbnail":258,"__hash__":259},"content_en\u002Fnews\u002Fputting-ai-to-work-in-the-it-communicator-role.md","Putting AI to Work in the IT Communicator Role",{"type":7,"value":8,"toc":229},"minimark",[9,17,24,27,37,45,48,51,56,59,62,67,71,77,82,85,88,91,95,101,104,108,111,114,118,121,124,128,131,134,139,143,149,152,155,158,161,164,169,172,175,178,201,206,209,212,215,218,223,226],[10,11,13],"blockquote",{"style":12},"border-left: 3px solid #003bba99; padding-left: 20px;",[14,15,16],"p",{},"On an offshore project, a good deal of an IT Communicator's work is not difficult — it is simply slow. Re-translating a design document after an update. Building a set of test cases for a new screen. Preparing test data before a test run can even begin. As AI kept spreading into more and more kinds of work, we tried handing some of these repetitive pieces over to it, and we have written down honestly what came of it: both where the gains were obvious, and where a person is still the better choice.",[18,19,23],"h2",{"className":20,"id":22},[21],"divider","the-it-communicator-role-more-than-a-translator","The IT Communicator role: more than a translator",[14,25,26],{},"Many people picture the IT Communicator (COM) as simply the project's translator. In practice the role is considerably broader. A COM is the link between the developers and project managers in Vietnam and the engineers and clients in Japan, and the responsibilities that come with that stretch across the entire life of a project.",[28,29],"img",{"className":30,"alt":34,"src":35,"style":36},[31,32,33],"block","mx-auto","mb-6","","https:\u002F\u002Fhomepage-media.s3.ap-southeast-1.amazonaws.com\u002Fwp-content\u002Fuploads\u002F2026\u002F09\u002F10143200\u002Fimage12.png","width: 100%;",[14,38,39],{},[40,41,44],"em",{"className":42},[43],"caption","Figure 1. The IT Communicator sits between the two ends of the project.",[14,46,47],{},"A COM does far more than render design documents into Vietnamese. They read those documents closely enough to grasp the business logic, spot descriptions that are not clear, and raise questions (QA) back to the Japanese side. Depending on how complex a document is, the COM may also run a design walkthrough for the development team. Beyond that, a COM often writes unit test (UT) and integration test (IT) cases, prepares test data, runs the tests personally, and runs them again once bugs have been fixed. Interpreting and reporting progress in meetings with the Japanese side falls to the COM as well.",[14,49,50],{},"Line those tasks up side by side and two distinct groups emerge. One is communication and judgement, which is bound tightly to people and hard to replace. The other is document and data handling, which largely repeats a familiar pattern. That patterned half is where we tried AI first, across the four tasks below.",[18,52,55],{"className":53,"id":54},[21],"before-the-results-tooling-and-data-boundaries","Before the results: tooling and data boundaries",[14,57,58],{},"On tooling, most of the work below happens inside VS Code with GitHub Copilot and Claude Code, backed by instruction sets and skills that the project team wrote themselves and refined run by run. For the period this article covers we were using Claude Sonnet 4.5; the team has since moved to Claude Opus 5.",[14,60,61],{},"Design documents on an offshore project belong to the client, so before putting any of their content into an AI tool we settled on a few ground rules. We use only enterprise-tier services that commit to not training on input data. We feed in only the portion a given task actually needs, rather than uploading whole documents by default, and we strip or replace anything sensitive such as real user data or system connection details. Most importantly, how far we use AI stays within what has been agreed with the client — it is not a decision the development team makes on its own.",[18,63,66],{"className":64,"id":65},[21],"four-tasks-we-handed-to-ai","Four tasks we handed to AI",[28,68],{"className":69,"alt":34,"src":70,"style":36},[31,32,33],"https:\u002F\u002Fhomepage-media.s3.ap-southeast-1.amazonaws.com\u002Fwp-content\u002Fuploads\u002F2026\u002F09\u002F10143200\u002Fimage1.png",[14,72,73],{},[40,74,76],{"className":75},[43],"Figure 2. How the work is split between AI and the COM across the four tasks.",[78,79,81],"h3",{"id":80},"translating-documents","Translating documents",[14,83,84],{},"This is where AI went in earliest, and where the payoff is clearest.",[14,86,87],{},"Markdown, Word, PowerPoint, PDF — whatever the format, AI can produce a translation. What is left for the COM is reading it back and fixing whatever does not sit right.",[14,89,90],{},"The case worth singling out is a document that has already been translated and then gets added to by the Japanese side. For that we packaged the repetitive steps into a skill of their own. Translating only the changed portion is something a COM has always done; what is different is that the whole chain now runs by itself — checking the commit, isolating what changed, locating the matching place in the existing translation, and slotting the new text in so the style stays consistent. Those were the fiddly steps, and the easiest ones to miss something in.",[28,92],{"className":93,"alt":34,"src":94,"style":36},[31,32,33],"https:\u002F\u002Fhomepage-media.s3.ap-southeast-1.amazonaws.com\u002Fwp-content\u002Fuploads\u002F2026\u002F09\u002F10143200\u002Fimage2.png",[14,96,97],{},[40,98,100],{"className":99},[43],"Figure 3. Updating a translation from a commit, before and after it was packaged into a skill.",[14,102,103],{},"Early on, when COMs were not yet fluent with AI, the development team built them a small translation site of their own. It was straightforward: upload the Markdown file, wait a moment, get back a download link. Once everyone had grown comfortable with GitHub Copilot and Claude Code inside VS Code, that site quietly fell out of use.",[78,105,107],{"id":106},"reading-design-logic","Reading design logic",[14,109,110],{},"When a design document was hard to follow, a COM used to read it over and over, and raise a QA to the Japanese side if it still did not come together. Now the COM can ask AI to explain the passage and offer suggestions, which often surfaces further points worth asking about. AI can also summarise a business flow beforehand, which makes the design walkthrough go more smoothly.",[14,112,113],{},"That said, AI does not make the call here. It acts as a second reader — one that reads quickly, will reread as many times as you like, and points out where two statements contradict each other. Verifying and deciding still belong to the COM. Unlike translation and test case writing, this work needs no packaged instruction set; it is a direct conversation with AI on the document at hand.",[78,115,117],{"id":116},"writing-ut-and-it-test-cases","Writing UT and IT test cases",[14,119,120],{},"We packaged our test case requirements into an instruction set of their own and gave AI the writer's role. Working from it, AI produces a test case table that tracks the design document closely, and the COM steps back into a reviewer's role: adding cases that are missing, removing ones that are wrong, and checking the whole thing against the design once more before it goes into use.",[14,122,123],{},"UT and IT are handled by two separate instruction sets, because the two produce quite different formats. UT follows a single screen or a single API, while IT is a business scenario running across several screens, written from the perspective of each role.",[78,125,127],{"id":126},"preparing-test-data","Preparing test data",[14,129,130],{},"For screens whose data is constrained across several tables, building test data by hand eats a great deal of time. Miss a single record in a parent table and the whole set beneath it is worthless. The SQL is now AI's job: the COM describes the data a test case needs, AI writes the statements from the DB design, and what remains is importing them and working through the cases one by one.",[14,132,133],{},"Even so, AI only covers the technical constraints. The SQL it writes satisfies the foreign keys between the tables involved, but the data also has to satisfy the business conditions each test case is after — and that part the COM still has to read out of the design and specify up front.",[18,135,138],{"className":136,"id":137},[21],"how-much-time-this-saved","How much time this saved",[28,140],{"className":141,"alt":34,"src":142,"style":36},[31,32,33],"https:\u002F\u002Fhomepage-media.s3.ap-southeast-1.amazonaws.com\u002Fwp-content\u002Fuploads\u002F2026\u002F09\u002F10143200\u002Fimage8.png",[14,144,145],{},[40,146,148],{"className":147},[43],"Figure 4. Estimated time saved across three of the tasks.",[14,150,151],{},"These figures are subjective estimates, based on how the COMs on the team felt and informally noted the same kind of work going before and after AI. They are not the result of a controlled measurement, so they are better read as an indication of the size of the change than as precise numbers.",[14,153,154],{},"The 50–80% on translation deserves a note, because it was not there from day one. Early on, the COM still lost time fixing the formatting of the translation and reconciling terminology — time saved at one step came straight back at the next. The pace only really picked up once the instruction sets had been written out properly.",[14,156,157],{},"The roughly 50% on test data comes from dropping the work of tracing relationships between tables and writing SQL by hand. What is left still has to be done: importing the data, running it, and confirming that it really is what the test case needs.",[14,159,160],{},"The roughly 50% on test case writing holds for one kind of screen only, and that is also where we hit our clearest limit. There is more on this below.",[14,162,163],{},"Beyond those three figures, reading design logic produced a result we cannot put a number on, and one we had not expected. AI did not stop at flagging a gap in the logic — it proposed a fix as well. Those proposals obviously cannot be used without being checked first. Treated as a reference, though, they have helped a great deal when a COM is drafting a QA: instead of only asking \"this is not clear\", the question can now go out with a concrete option for the Japanese side to consider.",[18,165,168],{"className":166,"id":167},[21],"where-we-threw-the-results-away-and-started-over","Where we threw the results away and started over",[14,170,171],{},"Not every kind of work turned out the same. Translation was quite positive. Test case writing depended on the screen. On screens with a single permission tier, the test cases AI generated were mostly usable as they came and saved around half the time. On screens with several permission tiers nested inside one another, AI still produced the full volume of cases but assigned permissions wrongly in nearly all of them; there was so much to correct that fixing them case by case would have cost more than writing fresh. In the end we discarded the output entirely and wrote it by hand from scratch.",[14,173,174],{},"Even so, we are not concluding that AI cannot do this. Looking back, AI had not gone against the prompt — the prompt simply never asked for it. The problem sat in two places, and both were on our side. First, the prompt at the time did not spell out the constraints between permission tiers in enough detail. Second, the way we were using the model did not draw on its ability to reason through several layers of conditions stacked on top of each other. Both are entirely fixable, and that is the direction we are working in.",[14,176,177],{},"Three things we took away:",[179,180,181,189,195],"ol",{},[182,183,184,188],"li",{},[185,186,187],"strong",{},"Output quality tracks the quality of the instruction set (instructions, skills and so on)."," A first run rarely produces something usable straight away; the real gains only arrive after a few rounds of refinement.",[182,190,191,194],{},[185,192,193],{},"Measure task by task, not in aggregate."," Same tool, same project — yet two different kinds of screen gave completely different results.",[182,196,197,200],{},[185,198,199],{},"The COM's role shifts rather than disappears."," The centre of gravity moves away from producing the artefact yourself and towards reviewing it and standing behind what gets handed over.",[18,202,205],{"className":203,"id":204},[21],"how-it-feels-after-a-while","How it feels after a while",[14,207,208],{},"Honestly, we went in treating it as an experiment. The biggest worry was whether it would really save any time: an AI translation still has to be read line by line, specialist terms come out wrong, formatting breaks — fixing all of that could easily take longer than translating the document yourself. Alongside that sat another concern: these design documents belong to the client, so were we even permitted to put them into an AI tool, and where exactly should we stop?",[14,210,211],{},"We began to expect more of AI as we found ways to fit it into specific tasks. Updating a translation from a commit is one of them: checking the commit, isolating what changed, finding the right place in the existing translation and slotting the new text in. A job that used to take a whole session now takes a few minutes. AI has also caught contradictions in design documents that a COM had read over and over without noticing.",[14,213,214],{},"It brought a couple of surprises too. As mentioned above, AI does not stop at flagging a gap — it proposes a fix. And writing prompts turned out to be a skill you have to learn: give it a one-line instruction and the result will quite likely not be what you wanted. The better we became at writing prompts, the better the results came back.",[14,216,217],{},"One last thing is worth being clear about, because it is easy to misread from the numbers above. What gets saved is mostly the time spent producing something; the time spent reading it back is much the same as before. At this point AI has become part of a COM’s daily rhythm rather than something we bring out to try, but it remains a tool that supports the work rather than one that makes the decisions.",[18,219,222],{"className":220,"id":221},[21],"in-closing","In closing",[14,224,225],{},"Looking back over the whole thing, on work with a clear pattern the saving is substantial — enough to change how a COM organises their time. But the numbers are only half the story.",[14,227,228],{},"The other half lies in how each person approaches the work. A COM spends less time retyping what already exists somewhere, and more on the part no tool can take over: understanding the business properly, asking the right question in the right place, and standing behind the quality of what they hand over.",{"title":34,"searchDepth":230,"depth":230,"links":231},2,[232,233,234,241,242,243,244],{"id":22,"depth":230,"text":23},{"id":54,"depth":230,"text":55},{"id":65,"depth":230,"text":66,"children":235},[236,238,239,240],{"id":80,"depth":237,"text":81},3,{"id":106,"depth":237,"text":107},{"id":116,"depth":237,"text":117},{"id":126,"depth":237,"text":127},{"id":137,"depth":230,"text":138},{"id":167,"depth":230,"text":168},{"id":204,"depth":230,"text":205},{"id":221,"depth":230,"text":222},[246,247,248],"tech talk","AI","testing","Briswell Vietnam Co Ltd","2026-08-28","md",{},true,"\u002Fnews\u002Fputting-ai-to-work-in-the-it-communicator-role",null,{"title":5,"description":16},"news\u002Fputting-ai-to-work-in-the-it-communicator-role","https:\u002F\u002Fhomepage-media.s3.ap-southeast-1.amazonaws.com\u002Fwp-content\u002Fuploads\u002F2026\u002F09\u002F10143200\u002Fimage11.png","S32SPIUW8Q7ti1Tg7BVQ6u8ShMFkEXhnad6_ctlcQ4w",1789638841366]