Scaling QIS Analytics without scaling headcount

Every QIS (Quantitative Investment Strategies) platform needs holdings data. Thousands of funds. Hundreds of formats. An analyst joined the team and spent his first three weeks doing the same thing. Not the same kind of thing but the exact same task, in a slightly different shape, with a different fund family each time. Find the file. Identify the holdings column. Separate value from quantity. Flag the missing attributes. Then start over with the next one. When the team studied where onboarding time was actually going, 60% of it was here in decoding, before a single number had been checked.

For most QIS and risk-tech platforms, getting holdings data is close to a solved problem. Collection has been made routine by years of work on scrapers, pipelines, and standardization. So, when our onboarding time was studied, the finding was a little uncomfortable. The constraint had moved downstream. The question was no longer how the data was obtained. It was how thousands of funds could be covered without an analyst being added for every few hundred.

The bottleneck was decoding, not checking

The analyst time was spent understanding a new disclosure format, not checking whether the data was right. Before a single number was validated, the source had to be found, the holdings file pulled, the value column told apart from the quantity column, and the missing attributes identified. Necessary work. Also, the same work in a slightly different shape, every single time.

So, the team built for that shape. Files are collected from each fund site by a Python scraper. They are standardized into one clean client format by an AI agent. The funds are then checked against our rules by a QC macro, which surfaces the outliers and returns cleaner data. A cycle that took about 75 minutes now takes 30.

It did not work cleanly the first time. A fund family changed their disclosure structure mid-project and the scraper returned nothing for three days. The fix was straightforward. The lesson was automation that cannot detect its own failures is not automation. It is a quiet liability.

The check is the point, not the overhead

The team was careful about one thing. A mapping model is perfectly capable of being confidently wrong. A settlement value can be mapped to market value and handed over as if nothing is off, and the error stays hidden until it surfaces in a risk number of weeks later. So, the analyst’s job was not removed. It was moved. Instead of decoding layout, they now decide whether the draft is right.

Own the edge, rent the rest

That is what makes build versus outsource mostly the wrong question. The useful one is narrower. The scraping logic, data models, validation rules, and analytics methods are the actual edge, and they are kept in-house. Document understanding, orchestration, and first-pass checks can now be bought and adapted rather than built from scratch.

The teams that scale coverage well over the next few years will not be the ones with the biggest operations headcount. They will be the ones honest about which of their work is real expertise and which is repetitive decoding that a machine can draft and a person can check. The first is worth protecting. The second was never a good use of an analyst’s day.