O trabalho do setor público é executado com base em registros escritos, e a maioria das ferramentas de IA deseja processar esses registros nos servidores de terceiros. Este guia separa a camada de captura, que deve ficar no próprio dispositivo do policial, das ferramentas de raciocínio em nuvem, que são uma categoria diferente. Ele percorre um fluxo de trabalho off-line prático para anotações de visitas, atualizações de casos e correspondência de residentes, e é direto sobre o que a IA no dispositivo não resolve.
Local government runs on written records. A home visit note, a case update, a letter to a resident, an enforcement observation, a committee paper: none of it is glamorous, and all of it is evidence of what a public body did and why. The writing is also the part of the job that gets pushed to the end of the day, then to the end of the week, then reconstructed from memory.
So the appetite for AI in the public sector makes sense. The problem is that most of the tools built to relieve that burden work by sending your words to a server. When the words are a research query, that is a reasonable trade. When the words name a vulnerable resident, describe a safeguarding concern, or record what you saw inside somebody's home, the calculation changes completely.
The constraint in public sector work is rarely the licence cost. It is custody. Who holds a copy, where it sits, how long they keep it, and what you would have to say if somebody asked. On-device AI changes that answer at the root: if the audio and the note never leave the device, there is no third-party copy to disclose, retain, or lose.
This guide is not legal, procurement, or compliance advice, and Yaps holds no government accreditation of any kind. Treat it as a practical look at where your words actually go, and follow your own authority's IT, data protection, and records guidance before you put any tool near resident information.
The phrase covers two categories that barely overlap, and mixing them up is where authorities get into trouble.
The first is a reasoning engine. It summarises a long policy document, drafts a first pass at a report, answers a resident question in a chat window, or searches across legislation and guidance. To do any of that well it needs a large model and often a large corpus, which is why it lives in the cloud and sends your text out to work. That category has real value, and Yaps does not try to be it.
The second is the capture layer. This is the constant, unglamorous part of the job: getting your own observations down in words. The note after a home visit. The three lines about a phone call before the next call starts. The description of a defect at a site. The paragraph of a report you would rather speak than type. This step does not need a server to work, and it is exactly where the most sensitive material appears first, in your own voice, with real names and addresses attached.
Authorities trip up by reaching for one cloud tool to do both jobs. The capture, which is the most confidential step and the one that needs the least computing power, quietly inherits a cloud round-trip it never needed. Your half-formed note about a resident is not a research query. It does not need to leave the building to become clean text.
Cloud-first capture
You dictate a visit note into a cloud dictation or notes service. The audio and the transcript travel to a vendor to be processed, then come back. For that trip, and for however long the vendor retains it, personal data about a named resident sits on infrastructure your authority does not control, under a retention policy nobody in your team wrote, possibly in a jurisdiction nobody in your team chose.
On-device capture with Yaps
You dictate the same note with the Yaps hotkey. Recognition and cleanup happen on the phone or laptop in your hand, and the finished text saves into a local vault. The audio is never uploaded to become text, so the confidential part of the work does not depend on a server being trustworthy, online, or in the right country.
Strip away the vocabulary and it is simple. On-device AI is artificial intelligence that runs on the computer or phone you are holding, instead of on a company's servers somewhere else.
Four things follow from that, and all four matter in public sector work. It is private, because the material never travels, so there is no third-party copy of a resident's information to secure, disclose, or breach. It keeps working with no connection, which matters in a basement flat, a rural property, a plant room, or anywhere the mobile signal dies. It does not silently change when a vendor updates a service, so the tool you validated in March behaves the same in November. And it costs the same to use heavily as it does to use lightly, because nobody is metering server time.
There is a fifth benefit that is specific to your world. A tool that never sends resident data anywhere is a much shorter conversation with your information governance colleagues than a tool that does. You are not arguing about a data processing agreement for the capture step, because there is no processor in it. Our plain-English explainer on what a local AI app is covers the general idea, and is voice typing private looks at what mainstream dictation tools actually do with your speech.
Yaps is a voice assistant for capture, not a policy engine or a chatbot for residents. It does a small number of concrete things, and it does all of them on your own device.
It turns speech into clean notes. Press the Yaps hotkey (on desktop the key is Fn, on Android it is the dictation button on the Yaps keyboard) and talk the way you actually think, with restarts and asides included. Recognition runs on the device and handles about 25 spoken languages, auto-detected, so a note about a conversation held in Polish or Urdu does not require you to change a setting first. Because it works across every app on the system, you can dictate straight into a form, an email, or a blank document. The dictation feature page covers the mechanics.
It cleans up the ramble. On by default and running locally, cleanup strips filler words and self-corrections, fixes punctuation and capitalisation, and turns a spoken list into an actual list. A muttered stream of "so, um, attended at eleven, sorry, quarter past eleven, no answer at the front, spoke to the neighbour who said" comes out as tidy sentences you can read back. You get the speed of talking with the finish of typing, and none of it needed a connection.
It keeps everything in a private searchable vault. Captured notes live in a local vault on your device that you can search without a cloud index of your caseload. Notes are not only walls of text: they support plain text, checklists, and kanban boards, so a spoken brain-dump after a visit becomes a working task list. When the text needs to go elsewhere, notes export to Markdown and plain text. The voice notes feature goes deeper on the structure.
It syncs between your phone and your laptop. Vault sync is a premium feature, and it pairs your own two devices with a QR code over your local network or an encrypted peer-to-peer link, rather than routing through a central server. The note you dictate in the car outside a property is waiting on the laptop when you get back to the office, without emailing yourself anything.
There is one more piece worth spelling out precisely, because it is the thing people most often assume wrongly. If you have an audio file you recorded yourself under your authority's own rules, the Yaps Studio editor imports that file and transcribes it offline into text or an SRT subtitle file. On Windows and macOS desktop, Studio can also identify different speakers in that recording: it is optional and off by default, you install a small speaker-identification pack once from the Features screen and switch the toggle on before transcribing, and it runs on the device like everything else. That feature is not available on Android or Linux.
What that is not is a bot in your meetings. Yaps does not join a call, does not sit in a committee meeting, and does not transcribe live. Studio works on a recording that already exists on your machine. If your authority wants automatic minutes from a live meeting, that is a different product category, and our post on an AI notetaker without a meeting bot explains the distinction and the workarounds.
Here is the loop that replaces writing everything up from memory at six in the evening.
Time, place, who was present, what you saw, what you did, what is outstanding. In the car or the corridor, with no signal needed.
Filler and false starts out, punctuation and lists in. Nothing is uploaded to make that happen.
Drop the follow-up onto a checklist or a board card so the action is recorded while you still remember it.
QR pairing over your local network or an encrypted direct link between your own two devices. No mailbox in the middle.
Dictate the narrative around your captured facts, then paste the clean text into your authority's system.
That is the question that decides whether a tool belongs anywhere near resident data. Not how capable the model is, but where your words rest and who else ends up holding them. Put on-device capture next to a cloud AI notes service on exactly that axis and they stop looking like rivals.
| Where your words go | Yaps (on-device capture) | Cloud AI notes service |
|---|---|---|
| Your dictated audio | Stays on your device | Uploaded to a server |
| Where the note text is stored | Local vault on your device | Vendor cloud account |
| Works with no connection | Yes, for core capture | No |
| Account needed to capture a note | No, only for billing | Usually yes |
| A third party holds resident data | No, nothing is uploaded | Yes, the vendor processes it |
| Transcribe a recording you made | Yes, in Studio, offline | Yes, by cloud upload |
| Joins a live meeting as a bot | No | Often yes |
| Summarise a long policy document | No, not its job | Yes |
| Connects to your case management system | No, paste-ready text only | Sometimes, by integration |
| Best role in your stack | Private capture and notes | Meetings and document work |
Read that as a fork rather than a scoreboard. The rows where Yaps says no are not apologies. A tool that never uploads your audio cannot leak it, and a tool with no live meeting bot cannot record a conversation somebody did not expect to be recorded. The missing features are mostly the features that would put a resident's information on somebody else's server.
Documentation in local government is not one task. It is a handful of recurring formats, and voice capture suits some of them far better than others. These are the ones where the loop above earns its place.
Visit and contact notes. Home visits, welfare checks, tenancy visits, and doorstep conversations. Short, frequent, and almost always written up late, which is exactly the profile that voice capture fixes.
Case updates between appointments. The three lines that keep a case file honest: who you spoke to, what was agreed, what happens next. Dictated in the gap before the next appointment rather than saved up for a Friday afternoon.
Inspection and enforcement observations. A repeating list of checks with an exception noted against each one. Build the round as a checklist once, then walk it and dictate the exceptions, so the negative evidence is recorded as well as the findings.
Resident correspondence. The reply you have already composed in your head while reading the complaint. Dictate the response, let cleanup punctuate it, then paste it into your mail or case system and edit for tone. See the voice notes feature for how these drafts stay organised.
Report and committee paper drafting. Not the finished paper, but the first honest pass at the argument. Most officers speak a rationale far more clearly than they type it, and a spoken first draft is easier to edit than a blank page.
Handover notes. The most undervalued document in any authority and the one that depends most on detail living in one person's head. Talking a handover out loud is faster and more complete than typing it, because you narrate it the way you would explain it to the colleague taking over.
Your own working notes. The half-formed thoughts, the questions to raise with a manager, the thing that felt off about a property. Keeping those in a local searchable vault rather than a cloud notes app is the single easiest privacy improvement most officers can make this week.
Send the research question to the cloud on purpose. Do not let a note about a resident go there by default.
Yaps
Public sector procurement runs on precise claims, so here are the limits without hedging.
Yaps holds no security certification or government accreditation. There is no FedRAMP authorisation, no Cyber Essentials certification, no ISO certification, and no government cloud approval. Nobody should present it to an assurance panel as accredited, because it is not. What it offers is an architectural argument rather than a badge: the sensitive step happens on hardware your authority already controls.
Yaps has no records management, retention, or disclosure features. It does not classify a record, apply a retention schedule, redact anything, or help you respond to an information request. A note in the vault is a document on a device, and your authority's own policies decide what happens to it. If your requirement is records lifecycle management, that is a different class of system entirely.
Yaps does not integrate with any case management, CRM, or line-of-business government system. There is no connector, no sync, and no plugin. It produces clean, structured, paste-ready text that you copy into whichever system your authority uses. That is a genuine limitation, and it is also why there is no second vendor holding your data.
Yaps does not join or transcribe live meetings. A committee meeting, a case conference, or a Teams call is not captured automatically, and there is no bot. For a recording that already exists and was made under your authority's own rules, Studio transcribes the file offline. Recording people is a governance decision, not a software feature, and it belongs with your organisation's policy.
There is no iPhone or iPad app yet. Yaps runs on Android, Windows, macOS, and Linux today, and iOS is coming soon rather than shipping now. If your field staff are issued iPhones, that is the blocker to weigh, and it is worth checking the current status before making plans. The read-aloud voices are also English speakers in practice, so if you need documents read aloud across many languages, a multilingual cloud reader is the more honest fit for that specific job.
Finally, nothing here is legal, procurement, or compliance advice. Our wider guide to voice privacy in regulated industries covers the same reasoning across health, legal, and financial settings, and the offline dictation use case explains what genuinely keeps working with no connection.
If you want to introduce on-device capture properly rather than quietly, it helps to arrive with the right questions already answered. These are the ones that come up in nearly every review, and the honest answers are short.
Who processes the personal data? For the capture step, nobody outside your authority. The audio is turned into text on the device, and the note is saved on the device, so there is no external processor in that part of the workflow. That is usually the fastest part of the conversation, because there is no data processing agreement to negotiate over something that does not leave the building.
Where is the note stored, and who backs it up? In a local vault on the device, which means your existing device policy is what protects it. Disk encryption, screen lock, mobile device management, and your backup arrangements all apply exactly as they do to any other document on that laptop or phone. On-device does not mean unmanaged, and a lost unencrypted phone is still a lost unencrypted phone.
What travels over the network, and when? Subscription status, model downloads, and optional vault sync between your own two devices. Sync pairs the devices directly over your local network or an encrypted peer-to-peer link, so it does not park a copy of your notes on a central server. If your policy is that nothing syncs at all, leave the feature switched off and the workflow still works on each device.
What must never be dictated? That is a policy decision, not a product one, and it is worth writing down before rollout rather than after. Agree with your team where dictation is appropriate, what gets recorded and what does not, when a resident should be told, and which system the finished text belongs in. The tool is deliberately dumb about that, and your organisation should not be.
Start with the capture layer, and start with Yaps. That is where resident information first appears in your working day, it is the step that gains nothing from a cloud round-trip, and it is the easiest thing to change without a procurement cycle. Install it on a work laptop or an Android device with your IT team's agreement, run the trial on your own non-sensitive notes for a week, and see whether the write-up burden actually drops before you take it anywhere near a live case.
Then, deliberately, keep using cloud tools for the jobs that genuinely need them. Summarising a two hundred page consultation response, searching guidance, or drafting from a large corpus is real work that a large cloud model does well. Choosing to send a specific, reviewed piece of text to that model is a completely different act from letting every raw thought about a resident default there.
The division of labour is the whole point. Private capture on the device, deliberate reasoning in the cloud, and a clear answer whenever somebody asks where the note about Mrs Patel on Tuesday afternoon is stored.
For the capture side of the job, Yaps is a strong choice because dictation, cleanup, note storage, and search all run on your own device, so notes about residents never travel to a server to become text. It runs on Android, Windows, macOS, and Linux. For summarising long policy documents or searching guidance, a cloud assistant is the better tool. The practical answer is to pair them: keep capture local, and send only deliberate, reviewed text to the cloud.
It depends entirely on where the processing happens, which is why the architecture matters more than the marketing. A cloud tool means personal data about a resident is processed by a third party, which needs a lawful basis, a contract, and your information governance team's agreement. An on-device tool like Yaps avoids that at the capture step because the audio and text never leave the device. Your organisation's own data protection and records rules still apply to the note either way.
On-device AI means the artificial intelligence runs on the phone or laptop you are holding rather than on a company's servers. In practice that means no third-party copy of the note exists, the tool keeps working with no connection, and the behaviour does not change when a vendor updates a hosted service. It does not mean the note stops being a record of your authority. It is still your organisation's document, subject to your own retention and disclosure rules.
No. Yaps holds no FedRAMP authorisation, no Cyber Essentials certification, no ISO certification, and no government cloud approval, and it should never be presented as accredited. What it offers instead is that the sensitive processing happens on hardware your authority already owns and manages. Whether that satisfies your assurance process is a decision for your IT, security, and information governance teams, not for a vendor page.
Not live, and not automatically. Yaps does not join meetings, does not act as a bot in a call, and does not produce minutes while a meeting is happening. If your authority has already recorded the meeting under its own rules and you have the audio file, the Yaps Studio editor transcribes that file offline into text or an SRT subtitle file, and on Windows and macOS desktop it can optionally identify different speakers after you install a small speaker-identification pack.
No. Yaps has no integration, connector, or sync with any case management system, CRM, or line-of-business government platform. It produces clean, structured, paste-ready text that you copy into whatever system your authority uses, and notes also export to Markdown and plain text. That is a real limitation for teams who want automation, and it is also the reason no second vendor ends up holding your case data.
Yes. Speech recognition, cleanup, note storage, and local search all run on the device and work with no internet connection, so a note dictated in a basement flat, a plant room, or a rural property is captured normally. The network is used only for a short list of things such as subscription status, model downloads, and optional vault sync between your own devices. Capturing and saving a note never requires you to be online.
Yaps offers a 7-day free trial, after which it is a paid subscription. The plans are Pro at $15 a month and Max at $25 a month, and an account is required for billing only, not to capture a note. There is no permanently free tier. Because the processing runs on your own device, heavy use does not cost more than light use.
Yaps dictation recognises about 25 spoken languages and detects them automatically from your speech, so a note that switches language does not require you to change a setting. That matters in authorities serving diverse communities, where a caseworker may need to capture detail from a conversation held in another language. Recognition runs on the device in every supported language, so the privacy benefit does not disappear when the language changes.
Yes, as a premium feature. Vault syncing pairs your own two devices with a QR code and moves notes over your local network or an encrypted peer-to-peer link rather than through a central server. A note dictated outside a property is waiting on the laptop when you get back to the office. Because the sync runs directly between your devices, no vendor holds a copy of the note in transit.
No. Yaps has no records management, retention scheduling, classification, redaction, or information request handling features of any kind. A note in your vault is simply a document on your device, and your authority's own policies govern how long it is kept and how it is disclosed. Keeping the note local reduces the number of third parties involved, but it does not change your legal obligations over the record itself.
Yaps runs on Android as a full customisable keyboard, and on Windows, macOS, and Linux as a desktop app, plus a Chrome extension for saving web pages into your vault. On desktop you trigger dictation with the Yaps hotkey, and on Android you tap the dictation button on the Yaps keyboard. There is no iPhone or iPad app yet: iOS is coming soon and is not publicly available, so teams issued with iPhones should check the current status first.
Yes, within limits worth knowing. Yaps includes text-to-speech with 18 voices on desktop, eight of which work offline, and two voices on mobile, triggered with Option and Fn on desktop or by choosing to read aloud from a selection on Android. The shipped voices are English speakers in practice. If your requirement is reading material aloud in many community languages, a multilingual cloud reader is the more honest fit for that job.
Yes, for the jobs it is built for. Cloud assistants are genuinely good at summarising long documents, searching guidance, and drafting from a large corpus, and those tasks need computing power that does not fit on a laptop. Use Yaps for the capture layer, where notes about residents belong on your own device. The two are complementary, and the discipline is simply to send text to the cloud on purpose rather than by default.