To prepare company data for sale, check your rights first, then work through the files: profile them, remove duplicates, scrub personal data, hash join keys, test re-identification risk, scan for secrets, convert to validated JSONL, write a data card and build a checksum manifest. Free browser tools cover each step without uploading anything.
Before you touch the files
Preparation starts with rights, not with formats. A dataset that is perfectly cleaned but restricted by a customer contract cannot be sold. Run the pre-check at /tools/data-rights-checklist before doing anything else. It asks about ownership, customer agreements, privacy notices and earlier licenses, and shows which answers need a closer look from counsel.
Next, decide the scope. Pick one source system and one record type for a first pass, such as resolved support tickets from the helpdesk. Export a working copy. Never work on the system of record, and keep the original export untouched so every later step can be traced back to it.
All of the tools named below are free and run in the browser. Files are read on your device and are not uploaded. They are built for working copies and samples of moderate size. For very large exports, use them on a sample to settle the method, then apply the same method in your own pipeline.
Step 1: Profile the dataset
Open the export in /tools/dataset-profiler. A profile tells you what you actually have: the number of rows, the fields, the type of each field, how many values are empty and how many are distinct. It is the fastest way to find out that a field you planned to sell is mostly blank.
Record the results. Row count, field list, fill rates and date range all go into the data card later, and a buyer will run the same profile to compare. Look for fields that should not be in the export at all, such as internal notes, agent names or billing details. Deciding to drop a field now is cheaper than scrubbing it later.
If the export is too large to open, profile a random sample of rows drawn from across the full date range. The first rows of a file usually come from a single period and give a misleading picture.
Step 2: Remove duplicates
Run the file through /tools/duplicate-remover. Exports often contain repeated rows from retries, system merges or overlapping date ranges. Auto-replies and templated messages can repeat many times over.
Buyers measure duplicates and discount for them, so remove them before you count volume. Decide what counts as a duplicate for your data: an exact match on the whole row, or a match on a key such as ticket ID. Note the rule and the number of rows removed. Then profile the file again so your record count reflects what you will deliver.
Near duplicates deserve a look as well. Two tickets that differ only by a reference number or a timestamp are the same content to a model, and a buyer's duplicate check may normalize those differences away.
Step 3: Scrub personal data and hash keys
Use /tools/pii-scrubber on the text fields. It finds and replaces structured identifiers such as email addresses, phone numbers and card-like numbers. Pattern matching does not catch everything. Names, addresses and identifying details inside free text need a second pass, so read a random sample by hand after scrubbing.
Some identifiers carry structure you want to keep. A customer ID links tickets from the same account, and that linkage is valuable to a buyer. Use /tools/column-hasher to replace such columns with a salted hash, or to drop columns outright. The hash keeps records joinable without exposing the original value. Keep the salt secret and separate from the dataset. Under EU law, hashed identifiers are generally treated as pseudonymized, not anonymous, so hashing reduces risk without ending privacy obligations.
Step 4: Check re-identification risk and scan for secrets
Removing direct identifiers is not the end of the privacy work. Combinations of ordinary fields, such as job title, office location and start date, can single out one person. Run /tools/reidentification-risk-checker on the scrubbed file and choose the columns that describe people. It reports how many records are unique or nearly unique on that combination. Where the groups are too small, generalize the values, for example by replacing a full date with a year, or drop a column.
Then run /tools/secrets-scanner. Logs, code, configuration files and chat threads often contain API keys, passwords and access tokens that someone pasted in. A leaked credential in a delivered dataset is a security incident for you and a liability for the buyer. Remove every hit, and rotate any credential that was real.
Step 5: Convert, validate and count
Most AI buyers ask for JSONL: one JSON object per line. Use /tools/csv-json-converter to turn a CSV or JSON export into JSONL. Keep field names plain and consistent across files.
Check the result with /tools/jsonl-validator. It parses every line and reports the line number of any that fail, along with records whose keys differ from the rest. Fix those before a buyer finds them.
Finally, run /tools/token-counter. Buyers think in tokens, not rows, because tokens measure how much text a model will read. The counter gives an estimate you can put in the data card next to the record count. State that it is an estimate, since each model family counts tokens a little differently.
Step 6: Write the data card and build the manifest
A data card is the short description that travels with the dataset. Fill in /tools/data-card-generator with what you recorded along the way: source system, record type, time span, volume, fields, the de-identification method, the duplicate rule, known gaps and the license scope you are prepared to offer. Be exact about gaps. A buyer who finds an undisclosed gap trusts the rest of the card less.
Last, build a checksum manifest with /tools/delivery-manifest. It lists every file with its size and SHA-256 hash. The buyer recomputes the hashes on arrival to confirm that nothing was lost or changed in transit. The manifest also fixes exactly which files the deal covers.
| Step | Tool | What you keep |
|---|---|---|
| Rights pre-check | /tools/data-rights-checklist | List of open rights questions for counsel |
| Profile | /tools/dataset-profiler | Row count, fields, fill rates, date range |
| De-duplicate | /tools/duplicate-remover | Duplicate rule and rows removed |
| Scrub personal data | /tools/pii-scrubber | Identifier classes replaced; sample review notes |
| Hash or drop columns | /tools/column-hasher | Columns hashed or removed; salt stored separately |
| Re-identification check | /tools/reidentification-risk-checker | Group sizes before and after generalizing |
| Secrets scan | /tools/secrets-scanner | Hits removed; credentials rotated |
| Convert | /tools/csv-json-converter | JSONL files with consistent keys |
| Validate | /tools/jsonl-validator | Zero parse errors |
| Count tokens | /tools/token-counter | Estimated token total |
| Describe | /tools/data-card-generator | Completed data card |
| Seal | /tools/delivery-manifest | SHA-256 manifest of the delivery |
What the tools do not do
The tools prepare files. They do not decide whether you may sell them, and they do not certify the result. Keep these limits in view.
When the package is ready, a buyer will run its own checks against it. Outrich Data, as a direct buyer, runs a data title search on the rights and handles de-identification of the full dataset itself before any AI lab sees it, so a seller's own preparation is a head start and not a requirement.
- They do not review contracts. The rights pre-check raises questions; counsel answers them.
- Pattern-based scrubbing misses names and context in free text. Sample and read.
- A low re-identification score on chosen columns is not a legal finding of anonymity.
- Browser tools suit working copies and samples, not very large production exports.
- A data card you write is a statement you will be asked to stand behind, so keep it accurate.