Joe Drafts, You Confirm: How Our AI Works
Ask Farmer Joe reads across your whole farm, the medicine book and the books-books, and drafts the answer or the entry. It never takes the next step on its own. That line is fixed, not adjustable.
Ask Farmer Joe reads across your whole farm, the medicine book and the books-books, and drafts the answer or the entry. It never takes the next step on its own. That line is fixed, not adjustable.
Ask Farmer Joe drafts, it never submits. Whether it's a cattle passport application, a movement report, or a VAT-coded expense line, Joe prepares the entry from what you've told it and stops there. Nothing is ever sent to a government system or posted to your books without you actively tapping to confirm it, and this rule has no setting to switch it off.
Joe sits across both halves of a farm's data at once, the statutory compliance side and the financial side, which is the specific thing most farm software cannot do because it only ever built one half. That means Joe can answer a question that needs both datasets joined together, what did that TB test actually cost per head, by pulling the vet invoice and matching it against the medicine-book entries for that herd, rather than making you do the arithmetic between two separate systems yourself. It drafts a passport application from a calving you described out loud. It suggests a VAT code for a receipt photo. In every case, the output is a draft sitting in front of you, not an action already taken.
It would be easy to soften this into something vaguer, "with your permission," "you're always in control," language that sounds reassuring but leaves room for an AI to act first and ask forgiveness through a notification you might not read in time. We didn't want that ambiguity anywhere near a system that touches government registrations and financial records. So the rule is specific and mechanical rather than a promise: a draft is not an action. A cattle passport application does not exist as far as BCMS or LIS is concerned until you tap confirm. A book entry is not posted until you tap confirm. There is no draft state that quietly becomes real if you ignore it for long enough.
Joe runs server-side, on frontier AI models with structured outputs, behind isolation that keeps one farm's data completely separate from another's. The model that answers a question about your herd can only ever see your herd, never a neighbouring farm's records, never an aggregate view across FarmHQ's whole user base. This matters more here than it would for a generic productivity assistant, because the data in question is a farm's actual compliance history and financial position, the kind of information that should never leak sideways even accidentally, let alone be usable to infer anything about a farm that has not chosen to share it.
When you're weighing up any AI-powered farm tool, ask it one direct question before you trust it with real records: what happens if you're wrong? If the honest answer involves the AI having already submitted or posted something before a human looked at it, that's a materially different, and riskier, product than one where being wrong just means correcting a draft.
An AI that gets a shopping list wrong is a minor annoyance. An AI that gets a cattle passport application wrong, and submits it, can leave an animal permanently barred from the food chain, the exact outcome the 27-day passport rule exists to prevent through human diligence in the first place. An AI that misreads a receipt and posts the wrong VAT code without review can distort a return HMRC treats as your legal responsibility, not the software's. The stakes on this side of farm software are not hypothetical, which is exactly why the confirm step is not negotiable.
Speaking a calving out loud into the phone, using streaming speech-to-text tuned for farm vocabulary and accents, is a genuinely faster way to log a record with muddy gloves on than typing. But voice does not skip the confirm step, it just changes how the draft gets created. Whatever reaches Joe, typed, spoken, or photographed, goes through the exact same "drafts, you confirm" gate before anything is saved or sent. The input method changes. The rule does not.
In practice, a draft is shown back in plain language before anything happens to it, a passport application with the calf's details laid out as Joe understood them, a VAT-coded expense line with the amount and category it picked, not a legal document or a raw API payload you're expected to parse yourself. If Joe misheard a tag number or picked the wrong VAT code, the fix happens right there, before confirming, not after the fact as a correction to something already filed. That single design choice, showing the draft in the same plain language a farmer would use to describe it themselves, is what makes the confirm step meaningful rather than a formality nobody actually reads.
Every feature we add to Ask Farmer Joe has to fit inside this rule, not the other way round. If a capability can't be designed to stop at a draft, it doesn't ship, however useful it would otherwise be.
See how FarmHQ worksNo. Joe can draft a passport application, a movement report, or a VAT entry from what you've told it, but nothing is ever submitted to BCMS, LIS, HMRC or posted to your books without you actively tapping to confirm it first.
Joe sits across both a farm's compliance data and its financial data and answers questions that need both, such as joining a vet invoice to the medicine book to work out what a TB test cost per head. It drafts entries and answers from your own farm's records, and stops there.
No. Joe runs server-side behind per-farm data isolation, so the model only ever sees the farm it is talking to. There is no shared context between one farm's data and another's.
No. Speech is transcribed and turned into a draft the same way typed input is, and the same confirm step applies before anything is saved or submitted, regardless of how the information reached Joe.