A compliance output that cannot be verified is not evidence. It is an assertion. Regulators do not accept assertions.
Describes the ProfytAI evidence architecture for in-product obligations and workflows. Your institution remains responsible for legal interpretation, scope, and supervisory engagement.
The Real Question
Banks ask whether tools like Copilot, ChatGPT, Claude, or Gemini
can replace a purpose-built compliance evidence system.
This is not a question of capability. It is a question of architecture.
How it reads regulations
General-purpose AI
Searches your documents and finds fragments that seem related to your question.
ProfytAI Evidence System
Ingests the full regulatory document. Preserves every obligation word-for-word. Assigns a unique ID to each.
What it produces
General-purpose AI
Writes a new answer in its own words every time you ask. The answer is different each time.
ProfytAI Evidence System
Returns the exact regulatory text, the same way, every time. The output is a stored record, not generated text.
What happens after
General-purpose AI
The answer appears in a chat window. When the session ends, nothing is stored. No audit trail exists.
ProfytAI Evidence System
Every output is stored, versioned, and linked to its source. The bank maps controls and attaches evidence. Users are assigned tasks. Audit readiness is tracked.
Under examination
General-purpose AI
The examiner asks you to reproduce the output. You cannot. The examiner asks for the source text. It was paraphrased. The examiner asks for the audit record. It does not exist.
ProfytAI Evidence System
The examiner requests any obligation by ID. The system returns the verbatim text, the mapped controls, the attached evidence, and the complete version history.
General-Purpose AI
Copilot, ChatGPT, Claude, Gemini
ProfytAI Evidence System
Purpose-built for regulatory examination
How it reads regulations
Searches your documents and finds fragments that seem related to your question.
How it reads regulations
Ingests the full regulatory document. Preserves every obligation word-for-word. Assigns a unique ID to each.
What it produces
Writes a new answer in its own words every time you ask. The answer is different each time.
What it produces
Returns the exact regulatory text, the same way, every time. The output is a stored record, not generated text.
What happens after
The answer appears in a chat window. When the session ends, nothing is stored. No audit trail exists.
What happens after
Every output is stored, versioned, and linked to its source. The bank maps controls and attaches evidence. Users are assigned tasks. Audit readiness is tracked.
Under examination
The examiner asks you to reproduce the output. You cannot. The examiner asks for the source text. It was paraphrased. The examiner asks for the audit record. It does not exist.
Under examination
The examiner requests any obligation by ID. The system returns the verbatim text, the mapped controls, the attached evidence, and the complete version history.
These are not two versions of the same tool.
General-Purpose AI is a search and summarisation assistant.
ProfytAI is a compliance evidence system.
They solve fundamentally different problems.
The Examination Test
Standard examination questions.
Any compliance tool that cannot answer them is a liability, not an asset.
“Show me the exact regulatory text your control satisfies.”
Every examiner asks this first. The answer must be verbatim. Paraphrasing is not acceptable.
Paraphrases source text. The regulator's exact words are not preserved.
Stored word-for-word at ingestion. Retrievable on demand.
“Reproduce this compliance output exactly as it was generated.”
Reproducibility is a core examination requirement. If the output changes on re-run,
it is not evidence.
Probabilistic by design. Every run produces a different output.
Same obligation ID returns identical output every time. Guaranteed.
“Provide the audit record and evidence chain for this finding.”
Without a stored record, there is no audit trail. Without an audit trail, there is no compliance.
Session-based. When the session ends, nothing is stored. No record exists.
Every output is stored, versioned, and retrievable. Always.
General AI was designed to answer questions.
ProfytAI was designed to produce evidence that withstands examination.
These are not the same system.
The Architectural Gap
The same compliance question enters two different systems. The divergence begins at step one and is irreversible. Follow each pipeline to see where they end.
General-Purpose AI Pipeline
You ask a question
Natural language input
Finds document pieces
Fragments that seem related
AI writes a new answer
Different every time
Text appears in chat
In the AI's own words
Session ends
Nothing stored or versioned
You ask a question
Natural language input
Finds document pieces
Fragments that seem related
AI writes a new answer
Different every time
Text appears in chat
In the AI's own words
Session ends
Nothing stored or versioned
ProfytAI Evidence System Pipeline
Regulation ingested
Full document, verbatim
Obligations extracted
By modality and each with a unique ID
Controls mapped
Added by the bank's team
Evidence attached
Tasks assigned to users
Artifact stored
Versioned, reproducible record
Regulation ingested
Full document, verbatim
Obligations extracted
By modality and each with a unique ID
Controls mapped
Added by the bank's team
Evidence attached
Tasks assigned to users
Artifact stored
Versioned, reproducible record
General-Purpose AI
Session ends with generated text. Nothing is stored. The output cannot be reproduced, traced, or submitted as evidence.
ProfytAI Evidence System
Every run ends with a stored compliance artifact: verbatim source text, obligation ID, mapped controls, attached evidence, and a complete version history.
The divergence begins at step one. A language model ends with generated text.
No schema, no citation enforcement, no stored record.
ProfytAI ends with a persisted, verifiable artifact.
Why General AI Fails Under Examination
These are properties of how language models work.
No model, no prompt, and no configuration removes them.
Failure 01
Non-Determinism
Same query. Different output. Every time.
Run 1
Institutions are expected to maintain a classification framework.
Run 2
Institutions should implement a data categorization process.
Outputs Diverged: No Canonical Record
Under Examination
When the examiner asks you to reproduce your evidence, the output will differ. A record that cannot be reproduced on request is not a record.
Can this be fixed by configuring the AI differently?
No. Language models sample from probability distributions. Even at temperature zero, outputs vary across sessions and model versions. There is no mechanism that guarantees identical output.
Failure 02
Paraphrasing
The regulator's words, rephrased by a model.
MAS TRM § 7.7.4 · Verbatim Governing Text
“The FI should configure system events or alerts to provide an early indication of issues that may affect its IT systems’ performance and security.”
AI Output
“Financial institutions are expected to configure appropriate monitoring mechanisms to ensure early detection of system issues.
Obligation Reworded: Legal Precision Lost
Under Examination
The examiner asks for the exact words of the obligation your control satisfies. The AI output does not contain them. A paraphrase of a statute is not that statute.
Can this be fixed by configuring the AI differently?
No. Language models generate text. Generation means producing new words. The original regulatory text is lost in the process, even with retrieval-augmented generation.
Failure 03
Inferred Citation
A reference with nowhere to go.
“Refer to MAS TRM, Part IV (data management section) for the applicable requirements.”
Citation
Extracted
Source Text
Retrieved
Verbatim
Confirmed
Record
Stored
Evidence Chain: Broken at Retrieval
Under Examination
You cite the obligation. The examiner asks for the source text your system retrieved. You cannot produce it. The citation was generated, not retrieved.
Can this be fixed by configuring the AI differently?
No. The AI infers what the citation should be based on patterns in its training data. It does not retrieve a verified source record. The reference is a prediction, not a retrieval.
Changing models does not resolve these failures. Every large language model is non-deterministic, produces paraphrase, and cannot retrieve a verified citation because these are properties of the architecture, not limitations of any specific model. No amount of prompting, fine-tuning, or configuration changes this. An evidence system requires a different architecture entirely.
The ProfytAI Standard
Verbatim
Source text preserved
Every obligation stored exactly as written in the regulation. No paraphrasing. No summarization. The regulator's exact words,
retrievable on demand.
Reproducible
Same output every time
Any compliance artifact can be reproduced identically at any point in time. The output is the same output that was produced at ingestion.
Traceable
Every citation verified
Every obligation links directly to its source clause. Every control maps to the obligation it satisfies. Every evidence item is attached and retrievable.
The Four Guarantees
Every obligation is traceable to its exact source clause
Every output is reproducible on demand
Every artifact is stored, versioned, and retrievable
Every citation is independently verifiable
These properties apply to content and workflows delivered through the ProfytAI platform for covered regulatory material. They do not replace your policies, sign-offs, or external legal advice.
“A language model produces answers. An evidence system produces proof. The difference is not quality. It is architecture.”
ProfytAI is not a compliance assistant. It is a compliance evidence system - the layer between your institution and your regulator that makes examination a process of retrieval, not reconstruction.
Request a Demonstration
We will walk through a live compliance artifact:
obligation ID, verbatim source text, evidence chain, and audit record.