The screen that replaces typing
Klok’s signature component: the document on the left, the drafted record on the right, confidence flags on anything uncertain, and a confirm button. Every module that takes in a document uses it.
AI validation
Any Document You Already Have
Aadhaar, PAN, a joining form, a bank passbook, a bill, a resume. Photographed on a phone is fine — no template, no scanner.
Fields Drafted, Not Guessed
Klok extracts the values and maps them to the right fields, marking what it is unsure about instead of filling it silently.
Document Beside The Draft
You check against the source without switching screens — which is what makes reviewing faster than typing.
Nothing Writes Without You
The rule is enforced in code and covered by tests: no record is created or changed until a person confirms it.
Reused Everywhere
Onboarding, expenses, payroll inputs and recruitment all run through the same validator, so the behaviour is identical wherever you meet it.
A Visible Trail
What the AI drafted, what the human changed, and who confirmed it — all recorded.
Why drafting beats extracting
Most systems that read documents present you with a result. You are told the PAN is ABCDE1234F, and your options are to accept it or to go hunting for the original. That asks you to trust a machine on a field you cannot see.
Klok inverts it. The document stays on screen beside the drafted record, so verifying a field is a glance rather than an investigation. Where the model is unsure it says so, and the uncertain field is flagged rather than quietly filled. Confidence is shown, not hidden behind a confident-looking value.
Nothing is written until a person confirms
This is the rule the whole platform is built on, and it is not a setting. An AI draft is a draft: it occupies no record, appears in no register, and reaches no payroll run until somebody with the authority to do so presses confirm.
The practical effect is that a wrong draft costs nothing. It is corrected in the field and confirmed, and the correction takes seconds because the source document is right there. Compare that with a wrong value written automatically into an employee master, discovered four months later in a PF claim rejection.
The same component everywhere
Onboarding documents, expense receipts, résumés, invoices — all of them run through the same validator. Learning it once means learning it for the platform, and it means a correction pattern that is identical whichever module you are in.
Capture, draft, validate, commit
Capture
A photograph, a scan, a PDF, a forwarded email attachment. Whatever the document already exists as.
Draft
The model reads it and proposes field values, marking anything it is not confident about rather than guessing silently.
Validate
The document sits beside the draft. You check what matters, correct what is wrong, and leave the rest.
Commit
Only on confirm does anything reach the record — and the trail keeps who confirmed it, and when.
Questions about AI validation
What happens when the AI reads something wrong?
You see the mistake, because the draft sits on screen beside the document it was read from, with anything the AI was unsure about flagged rather than smoothed over. You correct the field and confirm — a misread costs a ten-second correction, not a wrong record in payroll, because nothing is saved unsupervised. Wrong readings are expected; the design's whole job is making them impossible to miss.
Does it work with handwriting or poor photos?
Printed documents read reliably. Handwriting, poor lighting and creased photocopies are where confidence drops — and dropped confidence is displayed, not hidden, so you know exactly which values to check rather than re-verifying everything. A confident wrong value is the dangerous one; an honest "I am not sure about this field" is the behaviour you actually want from a machine touching payroll.
Can we turn the AI off entirely?
Yes. Every screen in Klok can be filled by hand, and the AI is an accelerator you can ignore indefinitely — there is no workflow that requires it. In practice most teams try the validator on one joining form, watch the draft appear beside the document, and never type one again. But that is a choice the product earns, not one it forces.
Which AI models process our data, and on what terms?
Anthropic's Claude and Google's Gemini, under business terms that do not permit using your data to train their models — stated in our privacy policy, not buried. Only the content needed for the reading you asked for is sent, and the output comes back as a draft a person must confirm. If data residency is a hard requirement, raise it before you buy: model processing happens with those providers, and we will not pretend otherwise.
Who inside our company sees what the AI reads?
The same people your roles allow and nobody else — the AI does not create a side-channel around your permission structure. A document dropped for an employee is confirmed by that employee and verified by HR; the draft, the confirmation and every later view are written to the audit trail. Every company's data is isolated from every other's by a boundary that is enforced in code and covered by automated tests.
Can we switch this off our bill if we do not use it?
Modules are switched on per company, and what is off is not billed — there is no bundle you must buy through to reach the part you want. AI validation ships with the document workflows it powers, so in practice you are choosing which modules run, and paying for exactly those. Billing follows the switches, and the switches are yours.
See it on your own data
Fourteen days, no card, no consultant. Import your Excel and watch the first record draft itself.