Neurosymbolic AI

AI understanding,
separated from execution.

DataKite is built on a neurosymbolic architecture. The model interprets what the customer means; deterministic platform logic decides what runs, what reaches the customer, and what gets logged.

The architecture problem

Why standard AI architecture struggles in banking.

Guardrails are probabilistic

LLM + guardrails reduces risk but does not create deterministic control over what executes.

Errors compound

Errors can compound across prompts, orchestration, tools, and output before anyone sees them.

Hard to audit

Banks need execution that is auditable, replayable, and policy-controlled — not generated.

Neurosymbolic architecture

Two layers, distinct responsibilities.

A neural model handles language; a deterministic layer handles execution. Each layer does the work it is good at, and neither is asked to do the other’s job.

01
Neural layer
Understands language. Extracts intent, entities, and values. Produces structured interpretation, not free-form action.
02
Symbolic / deterministic layer
Validates policy, workflow state, API permissions, and output templates. Decides what actually runs.
Execution flow

The same governed path, every request.

Every request through the platform follows the same path. The model interprets; deterministic logic decides what runs; execution goes through governed skills and bank APIs; output is templated; the full path is logged.

  1. Step 1

    Input

    A customer request, app event, or scheduled trigger enters the platform.

  2. Step 2

    Interpretation

    The banking-tuned model resolves intent, entities, and values as structured output.

  3. Step 3

    Validation

    Deterministic logic evaluates policy, workflow state, and skill eligibility.

  4. Step 4

    Skill / API execution

    Execution runs only through governed skills and the bank’s own APIs.

  5. Step 5

    Templated response

    Customer-facing output is rendered from templates populated with validated data.

  6. Step 6

    Audit log

    Input, interpretation, decisions, actions, and output are joined in one reviewable record.

What this architecture enforces

Controls, not guardrails.

Free-form output disabled

Customer-facing surfaces never emit raw model text. The model is upstream of execution and rendering.

Skills and bank APIs only

The platform executes through governed skills and the bank’s own APIs. No tool or endpoint is reachable unless it has been registered as a skill.

Templated, grounded responses

Responses are rendered from templates populated with validated data from the execution result — not generated free-form.

Audit and replay by design

Every request joins input, interpretation, decisions, executed skills, and rendered output in one record that can be reviewed and replayed.

Comparison

LLM + guardrails vs Rawi.

LLM + guardrails

  • Model interprets and may influence output and actions
  • Guardrails are probabilistic
  • Harder to audit
  • Free-form output risk
  • Tool misuse risk

Rawi

  • Model only interprets
  • Deterministic platform acts
  • Governed skills and bank APIs
  • Templated customer output
  • Audit and replay by design
Data posture

Bank data stays in the bank.

The model is not the system of record. Rawi reads from the bank’s approved data sources at execution time, organized as a Knowledge Graph the deterministic layer can reason over.

01
Bank data stays in the bank
Rawi reads structured context from approved bank sources at execution time. There is no second persistent copy of bank data outside the bank environment.
02
Knowledge Graph posture
The data layer organizes the bank’s context — products, customers, entitlements, policy — as a graph the deterministic layer can reason over, rather than handing the model raw documents to summarize.
03
Grounded interpretation
The model interprets against this structured context, so intent and entity extraction stay tied to what actually exists in the bank.
Deployment

Bank-controlled, by design.

Rawi runs inside the bank’s environment, through the bank’s own APIs, with policy, templates, and audit logs in the bank-controlled deployment. Private cloud, on-prem, and hybrid are supported depending on the bank’s requirements.

01
Private cloud
Rawi runs inside the bank’s tenancy on its chosen cloud.
02
On-prem
For banks with on-prem requirements, Rawi runs inside the bank’s own data centers.
03
Hybrid
Sensitive workloads stay inside the bank; supporting services run in approved regions when the bank chooses.
Selected references

External authority on neurosymbolic AI.

Neurosymbolic AI is an emerging architecture direction for more governed, explainable AI systems. DataKite applies this architecture to banking.

  • DARPA
    Assured Neuro Symbolic Learning and Reasoning (ANSR)

    Category legitimacy and external validation that neuro-symbolic approaches are being used for assured, robust AI systems.

    Visit source →
  • EY-Parthenon
    Neurosymbolic AI

    Business and financial-services relevance — auditability, explainability, and enterprise use.

    Visit source →
  • Springer
    Academic review of neuro-symbolic AI

    Technical grounding and support for neuro-symbolic AI in high-stakes domains.

    Visit source →

See the architecture in a bank deployment.

We’ll walk security, risk, and architecture teams through the neural and symbolic layers, the execution flow, the data posture, and the deployment model.

Explore Rawi