It can read anything. It cannot change anything without your say-so.
Farm software makes you learn its filing system. Which menu holds fuel logs, which tab has the withholding period, which report shows what you spread and where. The information is in there somewhere, and finding it is your problem.
Ask Alffi inverts that. You describe what you want in your own words — a question, or a job that needs doing — and it works out which records to read, what arithmetic to run, and what needs to change. Then it shows you the change and waits.
That last part is the whole design. Alffi’s language model has no ability to write to your data. None. It can only draft a proposal. When you approve one, tested and deterministic code applies it, running under your own permissions. There is no second path into your records.
Which makes the two halves deliberately lopsided. Reading is wide open — ask anything, in any words, about anything you have recorded. Writing goes through a small, fixed set of doors, every one of them typed, tested, and shown to you as a card before it does anything. The flexibility is in the asking; the discipline is in the changing.
It also remembers. Anything you tell Alffi about how your place works — with your say-so — it keeps, so it starts each conversation knowing more about your operation than it did the last one.
Standing in a paddock you can read an answer perfectly well. What you can’t comfortably do is type the thirty-word sentence that produces one — one-handed, in the sun, with a glove off.
So now you don’t have to. Tap the microphone, say what you want, and the words appear in the message box — on the field screen, on the dashboard, anywhere you can type. It works on the phone you already have, and like the rest of the AI it draws a small amount from your prepaid balance each time you speak.
What you said arrives as ordinary text you can read and correct, exactly as if you had typed it. You press send. Nothing reaches your records because you stopped talking.
While the words are still the machine’s rather than yours, the screen says so — and that marker disappears the moment you edit them.
Before it listens, Alffi hands the recogniser your own paddocks, mobs, products, properties and machinery — read straight off your records. Nothing to maintain: name a paddock and it becomes a word Alffi expects to hear.
On our own farm that took farm names from two in five to roughly nine in ten. Ordinary words were never the problem — it is the paddocks, mobs and products that trip a recogniser that has never heard of your place.
Say something it mangles — “Boonegala”, “Hepper”, “Cydictin” — and Alffi puts the name it thinks you meant beside the box as something you can tap. It never quietly swaps a word for you.
That rule is deliberate. A wrong name applied silently books work against a paddock nobody named, and you would have no way of knowing.
Your audio goes to an Australian endpoint, comes back as text, and the bytes go no further — not stored, not kept for replay. The request is also marked so the speech provider cannot use your farm’s audio to train its models.
If the paid path is unavailable it falls back to the phone’s own free recogniser, still matching against your names, and the screen tells you which one produced the words.
This is an actual conversation from Mullinroo Estate, lightly trimmed for length. Nothing has been added, and the numbers are the platform’s own.
You have 166 kg of Green Urea NV® on hand.
That’s based on 1 purchase in and 1 application out, with the last movement on 16 July 2026.
Want me to check how much you’d need for a specific paddock, or look into ordering more?
Let me look up the Crop Paddock and the Urea product details at the same time.
Here’s the picture:
Let me set up the job for Dave and the planned spread operation.
Let me handle all three things at once — reschedule Dave’s job, record the Urea delivery, and find the Urea store.
Once both are approved, you’ll have 1,166 kg on hand — well above the 748 kg needed for the Crop Paddock spread. Do you want me to record the cost of the delivery as well?
Later in the same conversation, the farmer asked for a second job to be created and assigned to himself.
I’d be happy to create that job for you — but I don’t have your name on file. What’s your name so I can assign it correctly?
In most farm software, a dashboard is furniture you have to assemble before it’s any use. You pick the charts, track down the right data, wire up the filters, get the date ranges right. It takes time and a certain amount of technical confidence — so on most farms it happens once, if at all.
The setup cost isn’t even the real problem. You have to know what you want to see before you have it. So you build for the questions you could anticipate, and the question you actually have on a Tuesday in July isn’t one of them. Changing the dashboard needs the same skill all over again, so nobody does, and it slowly drifts out of step with the business until no one quite trusts it.
Alffi does that work instead. Ask the question in your own words and it works out which records to read, what arithmetic to run, and what shape of answer fits — a trend over time, a breakdown by machine or mob, a single number. Then it draws it. No field-picking, no filter-building, no date ranges to fiddle with.
And the answers worth keeping, you keep. Save one and it becomes a report you can run again. Which turns the usual order on its head: rather than guessing up front what you’ll want to watch, your dashboard accumulates out of the questions that actually mattered.
Ask however the question comes out — there is no phrasing to learn and no list to memorise. Alffi works out which of your records to read and what to calculate. Changing things is deliberately narrower: a closed set of typed actions, each one shown to you before it runs. Below is a sample of both.
A great deal of what makes a farm run lives in somebody’s head and nowhere else — which gauge is the house one, which paddock floods after 40mm, who supplied that bull, what you actually mean by “the back block”. New software makes you re-explain all of it, every time.
Alffi has a memory. Tell it something once and it keeps it, so the next conversation starts from what it already knows rather than from scratch. The longer you use it, the less explaining you do — and the more of your operation it can reason about without being walked through it first.
Two rules keep that comfortable rather than creepy. Alffi asks before it remembers anything, so nothing is quietly inferred and filed away. And everything it has learned sits on one settings page where you can read it, correct it, or delete any of it. Your farm’s memory is yours to inspect.
The payoff is immediate as well as cumulative: once a fact is stored, a question that used to mean looking things up is simply answered.
Once approved, I’ll know that “My Fergie” refers to the Massey Ferguson MF S.125 whenever you mention it.
Let me look up the Massey Ferguson MF S.125 for you.
No engine-hour readings have been recorded against My Fergie yet — neither via fuel logs nor service events.
Hours are captured as a meter reading when you log fuel (“put X litres in My Fergie at 1,250 hours”) or record a service. Once you log one, I’ll be able to tell you the latest figure. Want to record a reading now?
I’ll resolve yesterday’s date and find the diesel fuel store at the same time.
Got everything I need. Nothing is recorded until you approve the card.
Plenty of products will tell you their AI is never wrong. We won’t, because no honest one can.
The guarantee we can make is architectural, not statistical: whatever Alffi gets wrong, it cannot quietly put into your records.
The best demonstration is your data, your paddock names, your questions.