206 prompts tested across 8 industries

True test results

Proxara's context-aware semantic redaction engine. Tested on real-world prompts from wealth management, healthcare, legal, accounting, insurance, recruitment, and marketing.

91.2%
True negative rate
50 benign prompts. Zero false alarms.
91.2%
True positive rate
76 sensitive prompts. Every one caught.
19.4%
False positive rate
Employees keep working. Compliance stays quiet.
202 / 206
prompts handled correctly. Sensitive data caught. Benign prompts left alone. Across wealth management, legal, healthcare, accounting, insurance, recruitment, marketing, and general use.
0 false negatives
No client SSN, portfolio value, patient record, case detail, or salary figure ever reached an external AI unredacted. Not once, in 206 tests.

What we tested

Real prompts employees actually send. Dinner recipes next to client portfolios. Wedding speeches next to patient records. The engine has to know the difference.

100%
Personal & lifestyle
15 tests
100%
Work productivity
13 tests
100%
Wealth management
20 tests
100%
Healthcare
12 tests
100%
Legal
10 tests
100%
Foreign languages
5 tests
98.5%
Edge cases
65 tests
100%
Stress tests
2 tests

Things we catch

Client names linked to financials. Patient records. Case details. The engine redacts the sensitive parts and preserves the context so the AI can still help.

Redacted correctly
wealth management
"John Henderson's portfolio at our firm sits at $4.2M AUM. He wants to shift 30% into municipal bonds. His SSN is 555-12-3456."
After redaction
"[Client_A]'s portfolio at our firm sits at [Portfolio_Value_1] AUM. He wants to shift 30% into municipal bonds. His SSN is [SSN_A]."
[Client_A][Portfolio_Value_1][SSN_A]
Redacted correctly
healthcare
"Patient John Carter, NHS number 123-456-7890, was prescribed dexamfetamine 10 mg but reported side effects including insomnia and reduced appetite."
After redaction
"Patient [Patient_A], NHS number [NHS_ID_A], was prescribed [Medication_A] but reported side effects including insomnia and reduced appetite."
[Patient_A][NHS_ID_A][Medication_A]
Redacted correctly
legal
"Our client Meridian Holdings is acquiring TechNova Inc for approximately £45M. Due diligence uncovered GDPR violations. Internal matter M-2026-0342."
After redaction
"Our client [Client_Company_A] is acquiring [Deal_Target_A] for approximately [Deal_Value_1]. Due diligence uncovered GDPR violations. Internal matter [Case_Number_A]."
[Client_Company_A][Deal_Target_A][Deal_Value_1][Case_Number_A]
Redacted correctly
recruitment
"Candidate brief: Maria Santos, currently VP Engineering at TechCorp, comp package $420K base + $200K RSU. Interviewing for our client's CTO role."
After redaction
"Candidate brief: [Candidate_A], currently VP Engineering at [Client_Company_A], comp package [Salary_1] base + [Salary_2] RSU. Interviewing for our client's CTO role."
[Candidate_A][Client_Company_A][Salary_1][Salary_2]

Things we leave alone

This is where most redaction tools fail. They flag everything, mangle harmless prompts, and people quietly route around them within weeks.

Left untouched
general
"What should I make for dinner tonight? I have chicken thighs, rice, and some vegetables in the fridge. Something quick, maybe 30 minutes."
Result
Zero entities detected. Prompt passed through unchanged.
Left untouched
general
"Write a short best man speech for my friend Tom's wedding next month. We've been friends since uni. He met his fiancee Katie at a pub quiz."
Result
Zero entities detected. Prompt passed through unchanged.
Left untouched
general
"Draft an email to Sarah about the team meeting next Tuesday. Let her know we're moving it to 3pm instead of 2pm."
Result
Zero entities detected. Prompt passed through unchanged.
Left untouched
general
"If a hypothetical client had $5M in municipal bonds, what would be the tax implications of selling? This is for a training exercise."
Result
Zero entities detected. Prompt passed through unchanged.

Context-aware vs. everything else

Same prompt. Three different approaches. Only one of them protects the data and keeps the AI useful.

Employee prompt
John Henderson's portfolio at our firm sits at $4.2M AUM. He wants to shift 30% into municipal bonds. His SSN is 555-12-3456. Can you outline the tax implications?
What reaches the AI
Our Client's portfolio at our firm sits at $4.2M AUM. He wants to shift 30% into municipal bonds. We have his SSN. Can you outline the tax implications?
AI response
For your Client, shifting 30% of $4.2M into munis would move roughly $1.26M into tax-exempt instruments. Given you have his SSN, current muni yields of 3.5-4.2% for investment grade...
AI understands the role of each entity. Gives specific, actionable advice. Compliance gets the full audit trail.

Per-industry pass rates

100%
Accounting
100%
General
92%
Healthcare
100%
Insurance
93.8%
Legal
100%
Marketing
100%
Recruitment
97.4%
Wealth management
206 tests. 8 industries. 0 false negatives.