See it on your own records.
Talk to usBatch record review that reads the paperwork with you.
BatchPilot reads a finished batch record, drafts the review, and points out the missing signatures, out-of-spec results, and copy errors a reviewer has to catch. You spend your time deciding, not hunting.
Set up your company in under a minute. A sample batch record is built in, so you can run a full review without uploading anything of your own.
A real flag from the sample record that ships with BatchPilot.
98.6%
caught by deterministic checks alone
100
batch records in the eval corpus
7
defect types checked on every record
The daily grind
Engineers and QA lose hours to slow paperwork that is easy to get wrong: batch records, deviations, change controls. Every entry has to be checked line by line. Every result has to be matched against its spec. One missed signature can hold up a release.
What BatchPilot does about it
It does the first read for you. You get a draft summary and a ranked list of issues. Each one points to the exact spot in the record and says why it matters. You accept, edit, or reject each one. Nothing is approved unless a person signs off.
- Missing signature
- Missing data
- Out of specification
- Date / time gap
- Transcription mismatch
- Calculation error
- Other
What every review looks for
Goodbye, line-by-line hunting.
The rules and the AI look for these 7 kinds of defect on every record, and point at the exact line.
01A worked example
Three real flags, on one real record.
This is Metformin HCl 500 mg Tablets, batch MET-24-0087, a sample that ships with the product. Pick a defect to see what the record says, what BatchPilot raises, which part of the review raised it, and what a reviewer does about it.
What the record says
Verified by: ______________ Time: 09:05
Step 5: BLENDING (Verified by)
The "Verified by" entry is blank.
A significant step requires the documented identity of the person who performed it and the person who checked it.
FDA reference: 21 CFR 211.188(b)(11) (Second-person check). Suggested reference.
What you decide
Accept it and send the record back for the second-person signature, or reject it if you can show the clearance was signed elsewhere.
02The workflow
How a review goes, and who does what.
Load the record
Pick one of the built-in samples or paste your own. BatchPilot reads the whole thing, then keeps the document beside the review so every flag points at the line that caused it.
| Product | Batch | Status | Flags |
|---|---|---|---|
| Metformin HCl 500 mg Tablets | MET-24-0087 | Completed | 5(2 high) |
| Ibuprofen 200 mg Film-Coated Tablets | IBU-24-0153 | In review | 2(2 high) |
| Amoxicillin 250 mg Capsules | AMX-25-0042 | Not reviewed | 4(3 high) |
| Cefalexin 500 mg Capsules | CEF-25-0118 | Not reviewed | 3(3 high) |
Rules check, then the AI reviews
Fifteen rule checks run first and say what they could not check. Then the AI reads the record, looks into it, and checks its own findings. A finding both raise is shown once, with an FDA reference.
You decide
Every flag is only a suggestion. Accept it, reject it, or edit it with a comment. Nothing moves without your call.
Sign and export
Finish with an electronic signature. A compliance report, clause by clause, is saved and locked with it. Then export a record with a history that cannot be changed without it showing.
The 15 rule checks in step 2
- R1Blank signature or verification
- R2Blank required field
- R3Cross-reference with no identifier
- R4Result above a not-more-than limit
- R5Result below a not-less-than limit
- R6Result outside a range
- R7Yield inconsistent with its inputs
- R8Step limit looser than the approved specification
- R9Timestamps out of sequence
- R10Hold time over the validated limit
- R11Carried-forward quantity does not match
- R12Disposition disagrees with the laboratory result
- R13Unit weight contradicts the blend
- R14Yield taken against the wrong basis
- R15Deviation note with no reference
03Measured performance
How well it does, including where it falls short.
100 batch records carrying 345 planted defects, each drawn from a failure mode published in FDA inspection findings and produced together with its own answer key. Two engines, scored against the same defects with the same scorer.
| Engine | Defects caught | Flags on a clean record |
|---|---|---|
Deterministic checks No model. Runs in milliseconds, for nothing. | 98.6% | 0.0 |
AI review claude-opus-4-8, run over the same records. Measured 5 September 2026 on the earlier single-step AI review. | 67.1% | 9.4 |
Both together Rules and AI together, as the app now runs them. Measured 5 September 2026 on the earlier single-step AI review. | 99.7% | 9.5 |
What that table says
Most of a batch-record review is mechanical. A blank signature is a blank signature; comparing two timestamps is arithmetic. Deterministic checks do that part perfectly, instantly, and identically every time.
Adding the AI review on top of those checks caught four more defects out of 331 and raised almost seven hundred additional flags to do it. On a record with nothing wrong, it had something to say nine and a half times.
We would rather publish that than imply the model is doing work it is not.
Those AI figures are from 5 September 2026. Since 8 October the AI reviews in three steps (read, look into it, check its own findings), and a finding the rules already raised is shown once. That version has not been scored on the full set yet. When it has, these numbers will change.
| Defect type | Caught | Rate |
|---|---|---|
| Calculation error | 51/56 | 91% |
| Missing signature | 35/35 | 100% |
| Date / time gap | 52/52 | 100% |
| Missing data | 47/47 | 100% |
| Transcription mismatch | 43/43 | 100% |
| Out of specification | 76/76 | 100% |
| Other | 36/36 | 100% |
Deterministic checks over 100 records, 23 October 2026.
Deterministic checks, per defect type, worst first.
Why you should discount these numbers
The corpus is generated. We wrote a generator that plants defects, then wrote rules that look for defects of that shape, then scored both engines against it. A test made of mechanical patterns is going to be won by a mechanical checker, and the hardest tier of planted defects scoring 100% is evidence of that rather than evidence of quality.
Real records bring formatting variation, ambiguity, and judgment that rules cannot express and a model handles better. Our records do not, so this comparison is unfair to the model and flattering to the rules. It is a regression control, not a validation study, and it is not evidence of performance on your products or your formats.
Real executed batch records are confidential and none are published, so nobody can honestly claim a validated figure from public data. If we get real records from a partner site, this number will move, and it will probably move down.
The AI review is also not deterministic: the same four demo records scored 13, 13 and 12 of 14 across three runs. A single run is a sample, not a measurement. Rules run 23 October 2026 over all 100 records; the model run covers the 96 records that returned, on prompt version brr-2026-08-b.
04Compliance
Built for a regulated environment. Checked by rules, decided by people.
The details are on the security page, written for the people who have to answer for it.
- #001record.uploadedhumanA. Rao
- #002review.ai_draftedairules + claude-sonnet-5
- #003flag.acceptedhumanS. Mehta
- #004deviation.openedhumanS. Mehta
- #005review.completedhumanS. Mehta
Every action is hash-linked. Change an earlier entry and the chain breaks.
A person makes every call
The AI never approves or rejects anything on its own. A trained reviewer makes every decision, and the record shows who did.
A history you can trust
Every action is saved in a linked log. If someone changes an earlier entry, the link breaks and you can see it happened.
Every draft is traceable
Each draft records which AI model and version wrote it, so you can repeat the review later and defend it.
A compliance report, locked at signing
Signing saves a report with one line per FDA clause: a finding, a pass, or what could not be checked and why. Once saved, it cannot be changed.
Only the rules can say pass
A clause is marked as passing only when a rule checked it and found nothing. The AI can raise a finding, but it can never clear one.
05Comparison
A manual review vs. BatchPilot. Same people, less hunting.
A manual review
- Every line is read and checked by eye.
- Results are matched to their specs by hand.
- Two reviewers can reach different answers on the same record.
- It is hard to show afterwards which checks were done.
- The regulation behind a finding is looked up separately.
With BatchPilot
- 15 rule checks and a three-step AI review read every line first.
- The rule checks give the same answer every time.
- Each finding points at its line and suggests the FDA clause it relates to.
- It says plainly what it could not check, and why.
- A clause-by-clause report is saved and locked with the signature.
The same in both: a qualified person decides every finding and signs the review.
06Questions
Questions we get asked.
If yours is not here, ask us. We would rather answer it straight than have you find out later.
Does the AI approve or reject anything on its own?
No. It writes a draft review and lists suggested flags. A qualified person accepts, edits, or rejects each one. Nothing is released without a person signing it, and the record shows who signed.
Is BatchPilot validated for 21 CFR Part 11?
BatchPilot is built to support 21 CFR Part 11 and EU GMP Annex 11 expectations: unique named accounts, optional two-step sign in with an emailed code, an electronic signature that requires the reviewer to re-enter their password at the moment of signing, and a tamper-evident audit trail. Formal computer system validation is done per deployment with your quality unit. We do not claim a certification we have not earned.
Read the security pageWhich regulations does it check against?
Today, US FDA 21 CFR Part 211. Each finding names the clause it relates to, and every signed review saves a compliance report with one line per clause. The clause wording is copied from the eCFR as it stood on 21 October 2026, and once a day BatchPilot checks the eCFR for any later amendment to those clauses and raises it for review. It never changes the wording on its own. Until a qualified person has approved how our checks map to each clause, every citation is labelled a suggested reference, not a determination.
What happens to our batch record text? Does it train a model?
Record text is sent to our AI provider over an encrypted connection only to generate a review. The provider does not use API content to train its models. The API key stays on our server and is never sent to the browser.
Read the security pageWhere is our data stored?
Your records, users, and audit trail are stored in a managed Postgres database, encrypted in transit and at rest. Ask us for the current hosting region and we will confirm it in writing before you send us a real record.
Ask us about hostingCan we see who did what, and prove nothing was changed?
Yes. Every meaningful action is written to an append-only audit trail. Each entry is linked to the one before it with a SHA-256 hash, kept separately for each company. If an earlier entry is altered, the chain breaks and you can see it. Editing a record needs a reason and bumps the revision number.
Read the security pageCan another company see our records?
No. Users, records, deviations, and the audit trail are scoped to your company, and that scope is checked on the server on every request.
Read the security pageDoes it connect to our MES or eQMS?
Not yet. Today you paste or upload the executed record text. Tell us which system you use and we will tell you honestly whether a connector is on the roadmap.
What does it cost, and can we run a pilot?
There is no public price list yet. We run short paid pilots on a small number of your real records, so you can judge it on your own paperwork. Tell us your monthly review volume and we will come back with a number.
Ask about a pilot07Get in touch
Talk to us on your own records.
Want to see BatchPilot on your own records, or set up a demo for your team? Send us a note and we will get back to you.
We read every message. You will hear back from a person, usually within one working day.