Thinks in-house.
The language layer runs on the Apple chip – no cloud, no subscription. What you ask and what the assistant answers stays in-house.
Full ERP AI without the cloud. Including your Windows workstations.
The language layer computes in-house, on Apple Silicon. A Mac mini as the language node brings the assistant to every Windows workstation: the server only relays, the thinking happens at your place. AI stays OFF by default; every answer carries its source.
New: the assistant
John Paul AI. Ask your books – on your devices, never in the cloud.
The ewws assistant carries a Swiss first name: Urs in German-speaking Switzerland, Jean in Romandy, Giuseppe in Ticino, Reto in Grisons, John Paul in English. It understands questions like "Which invoices are overdue?" or "How are we doing with Muster AG?" and answers with numbers from your system – every answer carries its source, nothing comes from the model.
It only reads. You confirm.
"Send reminders to the overdue debtors" prepares the dunning run – nothing is sent until you confirm in the familiar dialog. The same goes for the period close. The assistant never books, pays or sends anything itself.
Explains the payroll.
"Why is Anna Müller's pay lower this month?" – the assistant compares the two most recent statements component by component and names the biggest movement. Seconds instead of half an hour of recalculating.
Knows the receipt inbox.
"What is in the receipt inbox?" – how many receipts are waiting, how many are ready for conversion, how many need your review. Conversion happens in the OCR inbox, after your confirmation.
Briefs trustees.
"Which mandate needs attention today?" – the morning triage across all mandates, most urgent first, with the francs at risk.
Four languages, one price.
The assistant speaks German, French, Italian and English – and it is part of ewws: the edition prices remain unchanged.
The language node
One Mac mini brings the AI to the whole company – Windows included.
Mixed environments are the norm in Swiss SMEs: accounting on the Mac, sales and warehouse on Windows. The language node solves this without the cloud: a single Mac with Apple Silicon does the thinking for every workstation. Full ERP AI without cloud dependency in mixed environments – no other Swiss SME ERP we know of offers this in this form.
One device thinks for everyone.
The Apple Foundation Models run on the Neural Engine of one Mac on the company network – a Mac mini is enough. A switch in the ewws settings turns it into the language node; no additional licences, no AI subscriptions.
Windows asks, the node thinks.
The Windows workstation asks in the assistant. The ewws server relays the question to your tenant's language node and carries the answer back: question and answer stay within your installation, the thinking happens on the node.
Visible, not just claimed.
The cockpit shows the language node as a presence signal: green when it is ready, amber when it has not reported for a while. And every answer carries its source – which tool, which figures.
Prerequisite: a Mac with Apple Silicon on the company network. The node is bound to your tenant; without an available node, the Windows workstation honestly says the voice feature is not ready.
Data sovereignty first
Your data belongs to you. Even when AI helps.
In ewws, AI is a choice – per tenant, reversible at any time, and OFF by default. Four paths are open: leave it off, on device with Apple Foundation Models (no document leaves the Mac), Anthropic Claude or OpenAI with your own API key (BYOK, audit trail per tenant), or your own AI under your own responsibility. As delivered, ewws routes not a single document text to an external provider.
OFF by default. Three included options plus your own AI integration.
Off by default: no suggestions, no cloud requests. Whoever needs AI enables it per tenant and picks the mode. The deterministic modules (dunning, forecast, payroll anomalies) run locally with your database anyway – no language model required.
Integrate your own AI (Bring Your Own AI)
Customers who want to integrate their own AI (a local model such as Ollama, an own cloud LLM, a company-specific provider or an on-premises inference server) can do so. ewws provides the provider integration pattern. The responsibility for the choice of model, the contract with the provider, the data flow, compliance and ongoing operation lies entirely with the customer. We do not review or certify external models and give no data protection guarantee for providers that are not included.
Professional confidentiality
On-device AI for professions with a duty of secrecy.
By default, AI document classification in ewws is switched off. No document leaves the Mac. Whoever needs AI activates it per tenant; the standard recommendation is Apple Foundation Models on device, where documents do not leave the Mac. Whoever explicitly wants a cloud LLM brings their own API key for Anthropic Claude or OpenAI (BYOK, audit trail per tenant). At Bexio, Abacus, Sage and the third-party AI providers for Topal (Finmatics, Kontera, m2 machine), cloud is the only and unchangeable option; Abacus runs a Swiss cloud, which from a confidentiality point of view is not the same as on device. As of May 2026, checked against the vendor websites.
Fiduciary offices
Client documents are classified on the Mac, not in a cloud console. The duty of secrecy toward the client does not end at the AI model.
Law firms
Client-related invoices, fee notes and expenses are processed locally. No prompt leaves the firm, no provider sees the client name.
Medical practices
Document classification and posting suggestions run on the Mac machine. Patient-related notes on documents stay within the practice.
Associations and foundations
Member data, donation receipts and board decisions stay in the association office. Even with small structures, without an AI subscription with third parties.
Self-hosted means: the ewws server stands on your hardware or in your public cloud in your own account. On device additionally means: the AI model runs on the Neural Engine of the Apple Silicon Mac, not via a server request to an LLM provider. BYOK (Bring Your Own Key) means: whoever activates Anthropic Claude or OpenAI signs a contract directly with the provider and stores their own API key in ewws. We do not route through a collective account and do not see the contract content.
Operator overview
What the AI wants today, you see immediately.
The overview card shows open recommendations, uncertain cases and the current OCR adoption rate. After sign-in, the operator knows in 5 seconds where attention is needed and where the AI works independently.
Important today at a glance
- Top AI recommendations with severity (critical, review, info), sorted by impact.
- An uncertain queue with the cases from all modules that have low confidence.
- The OCR adoption rate of the last 30 days, with top suppliers and correction share.
Cross-module recommendations
Four recommendation sources, one place. Every recommendation with evidence.
The AI connects OCR, dunning, banking and reordering into a shared stream of suggestions. Not "the AI thinks", but a concrete statement with a sample size and a before-after comparison.
What the four recommendation sources deliver
- The OCR correction trend per supplier, when the operator increasingly corrects posting suggestions.
- Dunning clusters across several customers with the same risk profile.
- A reorder correction trend per article as well as a banking review trend per IBAN, each with a delta over 30 days.
AI question
Natural language over your own books, entirely without an LLM.
The operator asks in German, French, Italian, Romansh or English about open receivables, solvency risk, critical dunning, liquidity, supplier overrides or counterfactual explanations. The answer comes deterministically from your own tables, with a source and follow-up suggestions.
Eighteen intents in five languages
- Cash and risk: open receivables/payables, critical dunning, cash-flow horizon, solvency risk, largest overdue customer.
- OCR and AP: supplier with the most overrides, documents posted automatically today, confidence trend.
- Operational and XAI: stock warnings, top customer revenue, payroll run, VAT quarter, counterfactual gap. Patterns in de-CH, fr-CH, it-CH, rm-CH, en-US; the answer rendered locally.
Payroll anomaly check
Detect anomalies before the ESTV submission, not after.
Five plausibility detectors check every payroll run before submission: gross jump, deduction rate, employer-cost rate, net plausibility and new employees. Every anomaly comes with a comparison value from the last periods.
What the detectors detect
- A gross jump per employee with a comparison to the median of the last periods.
- Conspicuous deductions or an employer cost rate in individual payslips.
- New employees without a historical reference and net plausibility against gross.
From the application
AI that justifies its recommendation
No black box: every recommendation shows its evidence - critical items, scores, open amounts and overdue status. And the dunning score rates every open item individually, with a justification and a per-line breakdown.
Comparison
Where ewws AI stands today, compared with the known Swiss providers.
Fact-based, not bashing. If a provider has a capability, it says so. If not, likewise. As of 2026, checked against the vendor websites.
| Capability | Bexio | Topal | Abacus | Sage | Banana | ewws |
|---|---|---|---|---|---|---|
| OCR document capture with top 3 accounts | Cloud, simple suggestion | via a cloud third party | Cloud, simple suggestion | simple | - | Off by default; opt in on device (Apple) or BYOK (Anthropic/OpenAI) |
| Dunning risk scoring | - | - | rudimentary | - | - | with a factor breakdown |
| Liquidity forecast with sensitivity | - | - | classic | - | - | with assumptions and sensitivity ±10 pp |
| Payroll anomaly before ESTV | - | - | - | - | - | 5 detectors with a historical comparison |
| Cross-domain recommendations | - | - | - | - | - | 4 recommendation sources with evidence |
| Natural language query | - | - | - | - | - | self-hosted, deterministic |
| Active learning queue | not publicly documented | - | not publicly documented | - | - | transparent, with operator control |
| AI data leaves the company | yes (cloud) | via a cloud third party | yes (cloud, hosted in Switzerland) | yes (cloud) | n/a | Off by default; choice on device (Apple) or BYOK with your own contract |
As of May 2026, checked against the vendor websites. With the other providers, cloud is the only option; with ewws, AI is OFF as delivered, on device the standard activation path, and BYOK (Anthropic Claude or OpenAI with your own API key) the explicit opt-in choice per tenant. Whoever wants to integrate their own AI can do so on their own responsibility. The Swiss cloud at Abacus is good in terms of data protection, but from a confidentiality point of view not identical to on device. Documents are still sent to an external provider. With cloud AI ornamental features (auto mail, generative texts, chatbot), ewws is deliberately restrained.
What lies ahead
Honest about what the AI cannot do yet and when it is expected.
No AI hype. No promise that shines in the next demo and disappoints in everyday use. These two areas are next on the roadmap, with their own iterations. (The vector recognition for unknown suppliers has shipped since 17.05.2026, see Phase 6 above.)
Industry-specific anomaly profiles
The payroll detectors are generic. Hospitality and construction need different thresholds, for example because of 13th-month-salary advances, vacation payouts, hourly-wage models. Industry profiles for hospitality and construction follow with their specific thresholds and wage models.
An optional LLM layer for summaries
For customers who explicitly want LLM summaries, an optional layer is offered. Voluntary activation, with a clear record of which data goes where and to which provider. By default it is and stays off.
Taking stock
AI in ewws is a property of the product, not a marketing layer on top.
Sixteen AI functions are shipped, all self-hosted, all with source references. Built so trustees, lawyers and SMEs get their books under control faster – and the person who signs stays responsible. The demo is the fastest way to see it: ask the assistant yourself.