DEEN
Consultation
AI for confidential data

Privacy AI Bridge

Use the most capable AI in the world – without giving away a single client name. A sealed-off AI first redacts everything confidential. Only then does the text go to the large cloud AI, which never learns who it is about.

The process step by step Get advice, no obligation
Names leave your house only as placeholders Figures, quantities and deadlines are preserved Data processing agreement under Art. 28 GDPR
Who it is for

Built for people who work with other companies' data

Interim managers, consultants, buyers, lawyers, auditors. Your work consists of contracts, calculations and supplier data that is not yours – and which you have signed to keep confidential.

That is exactly why AI goes unused in these professions, even though it would save the most time there. Privacy AI resolves the conflict technically rather than by abstaining: the analysis happens – the names stay with you.

The process

What happens to your document

Six steps, played through on a real example: a framework agreement you want assessed.

Privacy AI – process

Step 1 of 6
privacy-ai.dss-success.com
New analysis Connection encrypted Signed in as m.schulz@interim-projects.com
1 · Documents
📄
Drag files here PDF, Word, Excel, text · up to 50 MB
📄Framework_agreement_2026.pdf14 pagesready
📊Price_list_Q3.xlsx312 rowsready
2 · Your question
Assess the termination and price adjustment clauses against the market. Where is there room to negotiate? Are there risks around delivery times?
Contract analysisRisksMarket comparison＋ Template
Start analysis
The files are used solely to process this request and are deleted afterwards. No training takes place.

Just like any AI service – only with one step in between. You upload and ask your question. What happens next is what sets Privacy AI apart from ChatGPT and Copilot.

privacy-ai.dss-success.com/detection
The sealed-off AI reads along runs locally – without an internet connection 47 findings across 14 pages
Framework_agreement_2026.pdf · page 3
§ 4 Term and termination

The agreement between Bosch GmbH (“client”) and Meyer Logistics KG (“contractor”) begins on 01.01.2026 and runs for 36 months.

The notice period is 6 months to the end of the term. The negotiator was Dr Sabine Krüger (s.krueger@bosch-example.com).

The annual volume amounts to €2.4 m with a price adjustment clause capped at 3.5 % p.a. Internal project number: BSH-2026-0417.
Found & mapped
FirmaBosch GmbH[SUPPLIER_A]
FirmaMeyer Logistics KG[CLIENT_B]
PersonDr Sabine Krüger[PERSON_1]
Emails.krueger@…[MAIL_1]
IdentifierBSH-2026-0417[PROJECT_X]
Deliberately left in place:
Amount€2.4 m
Deadline6 months · 36 months
Percentage3.5 % p.a.
The decisive trick: Only names and identifiers are replaced. Amounts, quantities, deadlines and percentages stay untouched – which is why the cloud AI can still calculate, compare and assess. Analysis quality is preserved almost entirely.

An AI of our own that gives nothing away. It recognises company names, people, addresses, email addresses and internal identifiers – including spelling variants and inflected forms. A placeholder is created for every finding. The mapping table stays within your area of responsibility.

privacy-ai.dss-success.com/approval
You see exactly what goes out Approval required Review all (47)✓ Approve and send
Exactly this text goes to the cloud AI – nothing else
§ 4 Term and termination

The agreement between [SUPPLIER_A] (“client”) and [CLIENT_B] (“contractor”) begins on 01.01.2026 and runs for 36 months.

The notice period is 6 months to the end of the term. The negotiator was [PERSON_1] ([MAIL_1]).

The annual volume amounts to €2.4 m with a price adjustment clause capped at 3.5 % p.a. Internal project number: [PROJECT_X].
Missed something?Select a passage in the text and add it manually. The rule remembers it for future documents.
Redacted too much?Individual placeholders can be undone – for instance when a term is a product rather than a company name.
Your own term listProject names, code designations or abbreviations that should always be redacted are stored once.

No flying blind. Before sending you see word for word what the cloud AI gets to read. You can tighten it up or take things back. Only your approval triggers the next step – automatically on request, once you trust the procedure.

privacy-ai.dss-success.com/analysis
The large AI works – without knowing who forrunning · approx. 40 seconds
With you / with us in the protected area
✓ Original document
✓ Mapping table [SUPPLIER_A] → Bosch GmbH
✓ Real names and identifiers
✓ File attachments in the original
only
placeholders
At OpenAI / Google / Anthropic
✕ No company name
✕ No person
✕ No mapping table
✓ Subject matter with figures
Progress
Termination clauses against the marketcomplete
Assess price adjustmentcomplete
Delivery times and risksrunning …
Formulate a recommendationwaiting

Full capability, half the knowledge. The model receives the complete subject matter including all figures – but no indication of which companies are involved. Even if the request were logged at a cloud provider, no conclusion about your clients could be drawn from it.

privacy-ai.dss-success.com/result
Result – with real names, assembled locallydone↧ As PDFStore in Sales
Assessment of framework agreement · Meyer Logistics KG
1. Notice period
The period of 6 months to the end of the term is at the upper end for the logistics sector. Three months from the second contract year is market standard. Recommendation: negotiate a staggered reduction.

2. Price adjustment
The cap of 3.5 % p.a. is, over a term of 36 months for Meyer Logistics KG advantageous provided an index link is added. Without an index there is a risk of it being fully applied every year.

3. Room to negotiate
At a volume of €2.4 m and with contact Dr Sabine Krüger we estimate the realistic room at €40–60k per year, primarily via the price adjustment clause.
Reassembly
[CLIENT_B] → Meyer Logistics KG
[PERSON_1] → Dr Sabine Krüger
[SUPPLIER_A] → Bosch GmbH
Insertion happens by plain string replacement – without a second AI model. So nothing can be invented along the way.
✓ Intermediate data deleted
✓ Mapping table discarded
✓ No training with your data
✓ Record generated for your files

No AI is involved in the reassembly, by design. The placeholders are simply replaced with the remembered names – no second language model that could add anything. This rules out content being altered while it is put back together.

privacy-ai.dss-success.com/record
Evidence for your files↧ Download record
TimeStepProcessed whereWhat was transmittedStatus
14:02:11Documents receivedProtected area–encrypted
14:02:14Detection of confidential passagesSealed-off AI, no network–47 findings
14:02:49Approved by userBrowser–confirmed
14:02:50Analysis commissionedCloud modelredacted text onlyTLS 1.3
14:03:31Result receivedCloud modelResponse with placeholdersok
14:03:31Names reassembledProtected area–without AI
14:03:33Intermediate data deletedProtected area–irreversible
For your clientsThe record documents which content left the protected area – and which did not. Suitable for the project file.
For your complianceDocumented processing path, named processor, defined deletion periods.
For yourselfTraceable even months later on which data basis a recommendation was made.

The decisive difference to quickly asking ChatGPT. At the end there is a record you can present to your client. It documents what left the protected area – and what did not.

Three levels

Depending on how far the data may travel

From an online service you can use immediately to a solution that runs entirely in your own house.

Level 1

Interim managers & solo

A pure online service. Nothing to install, ready straight away. We run the sealed-off AI that redacts and provide the cloud AI.

  • ✓ Ready to use in a day
  • ✓ No IT coordination needed
  • ✓ For individuals and small teams
  • ✓ Cancellable monthly
Enquire
Level 2

Your private AI server

We run a protected AI server for you that only your team accesses. No public cloud AI in the processing path.

  • ✓ Your own instance, not shared
  • ✓ Access via your user directory
  • ✓ Tiered by team size
  • ✓ Your own term and redaction rules
Enquire
Level 3

Everything in your own house

Entirely on your infrastructure, including the AI itself. Can be run without any internet connection if you wish.

  • ✓ Operation without network connectivity possible
  • ✓ For corporations, public authorities, critical areas
  • ✓ Set-up, training and support by us
  • ✓ Full data sovereignty on your side
Enquire
Use cases

Typical areas of use

Contract analysis

Assess clauses, find risks, estimate room to negotiate – without the contracting party becoming identifiable.

Procurement & spending

Check price lists and invoices for anomalies, identify savings, compare quotes.

Supplier assessment

Evaluate performance data, complaints and delivery reliability without revealing supplier names.

HR & people

Analyse applications, appraisals and compensation data without personal data leaving the environment.

Legal documents

Summarise and classify extensive pleadings while preserving client confidentiality.

Finance

Have analyses, forecasts and reports reviewed without figures being attributable to a company.

Safeguards

Which layers of protection apply today

With confidential data what counts is verifiability, not promises. That is why we set out openly which measures already apply contractually and technically – and which level is currently being built.

Active
  • ✓ Encrypted transmission via TLS 1.3
  • ✓ Redaction before every handover to cloud models
  • ✓ The mapping table never leaves the protected area
  • ✓ Data processing agreement under Art. 28 GDPR
  • ✓ Deletion of intermediate data after processing
  • ✓ No training of the models with your data
  • ✓ A complete processing record for every request
Being built
  • › Processing inside an encrypted enclave (confidential computing), so that even we as the operator cannot technically read the plain text
  • › External audit of the procedure

For level 3 the question does not arise anyway: there everything runs on your own infrastructure.

Frequently asked questions

Does the redaction reduce the quality of the analysis?
Barely. What gets replaced are names and identifiers – precisely the details that are not needed for a professional assessment. Amounts, quantities, deadlines, percentages and the entire subject matter stay untouched. So the AI can still calculate, compare and classify.
What happens if the redaction misses something?
That is what the approval step is for: before sending you see word for word the text that is due to leave the environment. You mark up anything missed, and the rule remembers it for future documents. In addition you can store your own term lists that are always redacted.
Why not simply use a local AI for everything?
Because locally operable models are noticeably weaker than the large cloud models on demanding analyses. Privacy AI combines both: the local AI for detection, where it is strong enough – and the large model for the actual thinking. Anyone who still wants to work purely locally chooses level 3.
Does Digital Shared Services see my data?
At levels 1 and 2 we process your documents as a data processor – so access would technically be possible at present, while contractually it is excluded and logged. Closing exactly this gap is the purpose of the planned processing inside an encrypted enclave. At level 3 everything runs at your site, where we see nothing at all.
Which models are used?
For detection we rely on established methods for recognising personal data in German. For the analysis you can choose between the leading providers – or specify a locally operated model. The choice is noted in the record.
What does it cost?
That depends on the level, the number of users and the data volume. After a short conversation you receive a concrete quote.

Bring along a real document

In the initial consultation we run through the process on one of your own cases – you see immediately what gets redacted and how good the analysis still is.

Arrange a consultation Back to the overview