A language
intelligence stack.
Akkordia will offer a set of AI inference models trained on expert-verified datasets that read the language of an interaction and return evidence-based indicators of coercion and manipulation in real time. Each component is at a different stage of development.
Five components, one evidence trail.
Akkordia is building the models, the data assets and the inference layer as one system.
Language Intelligence Models
Trained to identify the linguistic patterns associated with coercion and manipulation.
- Pattern-level detection across a whole interaction
- Every signal reported by name
Fraud Language Library
Records the signals detected, the language that produced them and the policy that applies, so any output can be reviewed and audited.
- Evidence phrase retained for every signal
- Built for human review and audit
Risk Inference Engine
Combines model output into evidence-based indicators for an interaction, each expressed as a named signal with the language behind it.
- Indicators, with no automated decision
- Judgment stays with the operator
AI Agent Support
A capability to give customer-facing staff and AI agents a safer next step during a high-risk interaction, drawn from the policy they already operate.
- Prompts drawn from the customer's own policy
- Planned capability for integration, not yet built
Coercion has a linguistic
shape, and it repeats across
scripts, scams and languages.
The pressure builds across an interaction as a pattern of restricted choice. A single keyword never carries it, which is why the models must read the whole conversation.
Time pressure
Deadlines that exist only inside the conversation.
Isolation and secrecy
Requests not to consult a colleague, a partner or a second channel.
Borrowed authority
An executive or a regulator invoked to override a control.
Unusual language pattern
The account may be right, but the language does not fit the person.
Every scam warning
must show its evidence.
Every prompt must name the signals it detected, quote the language behind them and record the action that followed.
- Interaction
- CALL-2026-08-24-04127 · inbound · EN-GB
- Signals detected
- Time pressure, isolation, secrecy request, borrowed authority
- Evidence phrase
- “please do not call the office line”
- Policy referenced
- Payment Authorisation Procedure, clause 3.2, callback on registered number
- Prompt shown
- Do not proceed. Verify on registered number.
- Agent action
- Callback initiated. Instruction not released.
- Outcome
- Attempted fraud prevented. Case referred to the fraud team.
What if we could help people
detect a scam instantly?
Many communications contain subtle indicators of coercion and manipulation. Akkordia is being built to demonstrate AI inference models can spot those signals instantly, alerting people to scams before harm occurs.