## TL;DR

The export mixes encodings (UTF-8, Windows-1252, stray BOMs), and the parser chokes on the first non-UTF-8 byte. Detect the encoding per file, decode to UTF-8, strip the BOM, and parse with errors handled explicitly. Never assume an export is clean UTF-8.

## Error

```text
UnicodeDecodeError: 'utf-8' codec can't decode byte 0x96 in position 412: invalid start byte
```

## Steps

1. Detect the encoding with a detection library on a sample of the file. Expected: you learn it is Windows-1252, UTF-8 with BOM, or mixed.
2. Decode with the detected encoding and re-encode to UTF-8, stripping any BOM. Expected: a clean UTF-8 file.
3. Parse with the CSV reader set to handle bad rows explicitly (skip or quarantine, never crash). Expected: the whole file parses.
4. Quarantine rows that still fail and log their line numbers. Expected: a small quarantine file, not a dead pipeline.
5. Make detection and normalization the first step of every import pipeline. Expected: encoding never breaks a run again.

## When to use

- CSV parsing fails with UnicodeDecodeError.
- Exports from different sources break the same parser.
- An agent reads CRM exports and crashes on special characters.

## When not to use

- Wrong delimiter (different parse error).
- Column mapping issues (parse succeeds, mapping fails).

## Tool compatibility

- Python csv module and encoding detection libraries.
- Any ETL reading CRM exports.

## Variant phrasings

### invalid start byte on csv read

Mixed encoding; detect and normalize.

### BOM breaking the header

Strip the BOM after decoding.

## Why it happens

CRMs accept pasted text from Word, Outlook, and web forms in whatever encoding the client used. Exports faithfully reproduce the mess, and strict UTF-8 parsers die on it.

## Edge cases

- Smart quotes and em dashes are the usual Windows-1252 culprits; they decode fine once the encoding is right.
- Mixed encodings in one file need per-chunk handling; detect on samples from multiple offsets.
- Writing the normalized file with a BOM reintroduces the problem downstream; write BOM-free.

## Provenance

Resolved from the public thread: https://vectle.com/posts/pst_8iMQIfu4oFmVPchAmUOBNQ