What kind of PDF do you have?
A Tamil PDF produced from a document carries a text layer and converts by extraction. A scanned Tamil PDF is images and must be recognised first. Both are accepted here. The distinction matters for expectations: extraction from a digital PDF is close to lossless, recognition is not.
Legacy encodings are common in Tamil documents
Tamil has a long history of pre-Unicode encodings — TSCII, TAB, TAM and various font-specific schemes. Documents in those encodings store bytes that only mean Tamil in the presence of one particular font, which is why they turn to nonsense when copied elsewhere. That is an encoding problem rather than a recognition problem, and converting the encoding preserves the text exactly, where OCR would only approximate it.
What is preserved
Text, paragraph structure and reading order. Fonts and precise page layout are approximated. The output is an editable document rather than a visual reproduction.
Checking the result before you rely on it
Two quick checks catch most problems. First, search the Word document for a distinctive word you know appears on the page — if search finds nothing, the text is not real Unicode and something upstream went wrong. Second, look specifically at the Grantha letters ஜ, ஷ, ஸ and ஹ and at words carrying the visarga-like ஃ, since those are the characters most often substituted when a scan is marginal. Spot-checking a paragraph against the original takes a minute and tells you far more than any overall accuracy figure, which is an average across a page and not a promise about your particular document.
Multi-column pages and reading order
Tamil magazines and academic papers are frequently set in two or three columns. Column detection decides the order text comes out in, and a page with a narrow gutter, a full-width heading spanning columns, or a floating image can be read across the columns instead of down them. If a converted document reads as interleaved half-sentences, that is what happened — converting the pages individually, or cropping to one column, usually resolves it.