I build Antoti, software that turns customer emails into Jira tickets. It runs on a computer in the company, and so does its AI. No message text goes to a cloud service. This article shows what the local model does there, what it does not do, and which checks it needs around it. It is useful for anyone building a local model into a real workflow.
Disclosure: Antoti is my own product. It costs €49 per month per company.
Why local
Many companies may not send customer emails to ChatGPT or other AI services. Data protection officers, works councils or their own clients forbid it. A model on the company’s own computer removes the problem at the root: the emails already sit with the email provider and the tickets in Jira, and no further service gets to read them in plain text.
The price: the model is small. Antoti uses Ministral 3 8B by Mistral AI, Apache-2.0, quantised as Q4_K_M. The file is about 5.2 GB and runs through llama.cpp on Mac, Windows and Linux. A model like that is not GPT-5, so the workflow has to be built so it does not need to be.
What happens before the model
The most important rule: whatever a rule can decide reliably, the model never sees.
- Check the sender. Antoti checks the DKIM signature of every email. Without a valid signature, no ticket is ever created automatically, whatever the model thinks.
- Filter the noise. Newsletters and automated emails carry fixed markers such as
List-Unsubscribe,Precedence: bulkorAuto-Submitted. Antoti sets them aside without the model, unless the sender is a registered customer. - Match the customer. First by saved addresses, then by the customer’s domain. Only if neither matches may the model pick from the list of Jira projects, and the ticket gets a note.
- Catch instructions. If sender, subject or body contain an instruction aimed at an AI, Antoti holds the email before the model reads it. It also checks readings a model can see but a person cannot: text in invisible tag characters, full-width letters, words split by invisible separators.
Only then does Antoti ask the model. The question is narrow: is this a request, and if so, what is it about?
What the model may never do
The model cannot execute anything. It has no tools, no access to Jira and no mailbox. It returns a classification as JSON, and the code decides what follows. A prompt injection that slips past the filter can at most lead to a wrong classification, which ends up in the console where a person sees it.
If the model is unsure or fails, Antoti does not guess. The email waits under “Zu entscheiden” (to decide) until a person decides. Emails set aside stay visible with the reason, and one click turns them into a ticket.
Memory: load only when needed
A computer that also does office work should not reserve 5 GB for a model all day. Antoti starts llama.cpp only when there are emails to check and stops it again after a while without requests. That costs a few seconds on the first call but frees the memory for the rest of the day.
Antoti only touches processes it started itself. If another model already runs on the computer, for example in LM Studio or Ollama, it stays untouched.
16 GB of RAM is recommended. On an Apple silicon Mac, llama.cpp computes on the GPU. How much memory which model needs is covered in How much RAM does a local LLM need on a Mac?.
When the model fails
A local model fails: too little memory, an update, a crashed process. Antoti does not lose an email when that happens. The raw message stays stored, Antoti retries after a pause, and until then the email waits. There is deliberately no fallback to a cloud AI, because that would break the promise “local” exactly when it matters.
What I learned
- The model is the weakest link, so it gets the smallest job. Everything rules can do reliably, rules do.
- Safety comes from missing permissions, not from the prompt. A prompt-injection filter helps, but it stays a heuristic. The real boundary is that the model cannot do anything.
- Uncertainty is a valid result. A model that may say “don’t know” is worth more in production than one that always has an answer.
Antoti is available for Windows, macOS (Apple silicon) and Linux, free for 30 days: See Antoti. Why Jira Cloud does not accept mail from customers without a licence at all, and the options you have, is covered in Customer emails as Jira tickets.
Frequently asked questions
Is an 8B model enough to sort customer emails?
For deciding whether an email is a request, yes, as long as rules filter out the obvious cases first and a person decides unclear ones. Antoti uses Ministral 3 8B in Q4_K_M, a file of about 5.2 GB.
How much memory does it need?
Antoti recommends 16 GB of RAM and about 6 GB of free disk space for the model. The model loads only when there are emails to check and unloads again afterwards.
How do you protect a local model from prompt injection in emails?
Antoti looks for instructions aimed at an AI before the model runs, including in hidden characters, and holds such emails. More important: the model cannot execute anything. It only returns a classification that the code checks.
Transparency
Sources and references
This technical assessment draws on the primary and reference sources below. Vendor claims and external benchmarks are labeled as such in the article.