Do not convert your CSV to Excel to help the agent

Something we see in nearly every lab we work with is that scientists love Excel. The plate reader writes a CSV, and within a few minutes that file has been opened, given a colour code, split across half a dozen tabs, with some numbers highlighted and others left alone, and then sent on by email with a note explaining what the colours mean. That preparation is exactly the wrong file to hand an AI tool to analyze plate reader data, because the colour code and the tabs are where you already wrote down what you expect to find, and the agent reads that as an instruction.

The agent sees your annotations and has to guess

Elnora reads those markings and reports them back: it will tell you that the header cells of six columns are filled green, and what colour each one is. Deciding what that green means stays with you, because a green header means "analyse these" in one lab and "already excluded" in the next, and the difference is in your head and in that covering email.

So the agent asks you which reading applies, or it proceeds using every column regardless, and either way the time spent colouring and splitting the file in Excel changed nothing about what happens next.

The preparation writes down what you already believe

We look at a plate already believing something about it: that the effect is in those six columns, that the two conditions at the end were probably a bad dispense, that the third replicate always looks odd on that instrument. Preparing the file is how that belief gets written into it, without anyone sitting down and deciding to write it in. Highlighting six columns points the agent at those columns before a single reading has been checked. Splitting the plate across tabs makes the comparisons inside one tab easy and leaves the comparisons across tabs as the ones that never get made.

When a scientist prepares a file for another scientist, both of them know the conventions and can argue about them. When the same file goes to an agent, nobody is there to explain what the colours and the tabs mean, so the agent reads the conventions as instructions from no one in particular. You end up reading back an analysis that matches the assumption you brought to it in the first place, which is a slow way to confirm something you already thought.

Give Elnora the file the machine wrote and the analysis runs on the readings, including the two conditions you were ready to write off.

Excel makes the analysis harder as well

A CSV is a table with the same delimiter and the same encoding from the first line to the last. A workbook is a document that happens to contain a table, and the difference shows up in ordinary work.

Sheet order changes between exports. Two workbooks from the same software, describing the same experiment, arrived with their sheets in opposite order, so anything reading the first sheet analysed a different dataset in each file. A full-sheet chart counts as a sheet and holds no rows. A workbook can also contain merged cells, hidden rows, a formula that recalculated when somebody opened it last Tuesday, and a column that displays as a date and stores as a number.

The glossary tab has the same problem. One workbook came with a sheet explaining the column names, written so that whatever read the file would understand them, and it was the last sheet, so it arrived after about five hundred rows of data and behind everything it was written to explain.

Saving the file back out of Excel changes it again. On a European locale you get semicolons between the fields and commas inside the numbers, and the character encoding changes too, so a plate map with a micro sign or a degree sign in a column name can come out unreadable. Elnora handles both, and both problems start the moment the file passes through Excel.

What to send

Export from the instrument as CSV and send that file. If something about it needs explaining, put the explanation in the data: a column, a line of text above the table, or the filename, where the next person to open the file will read it.

Better than that, connect the instrument to Elnora and let it pull the readings itself. A connected plate reader is already reached through Elnora's run planning, on the schedule you set, and taking the export step out removes the moment where a person opens the file and starts tidying.

Then ask your questions in the conversation. Elnora binds your plate map to the readout, so the analysis runs on one table keyed by condition, and the wells excluded from a fit come back listed with the reason they were excluded. You can still tell it that the third replicate looks wrong on that instrument. The difference is that you state it as a claim, one your colleagues or the agent can question, and the raw numbers stay in the table exactly as the instrument recorded them.

That is the same reason the agent drafting your report reads your knowledge base directly, with no summary in between, which is how we draft every document here.