The method

A gated, self-validating audit, end to end.

Siana applies the discipline of an audit firm to AI deployment: a standardised programme, validation gates, and a remote expert layer for the risky residue. We don't re-read engagements — we design each step to prove itself.

Nine phases, each with its validation gate.

A phase only closes when its acceptance criterion is green. Otherwise the tool blocks and shows what to fix.

  1. Engagement acceptance

    Qualify the buyer, budget and risk; rule out prohibited sectors.

  2. Planning & materiality

    Price the cost of an error — this threshold drives the level of human oversight.

  3. Evidence gathering

    Interviews, data inventory, process mapping, personal-information scan.

  4. Analysis & prioritisation

    Score each use case and pick the first, best-justified initiative.

  5. Design & feasibility

    Choose the architecture and hit the target on real data. Hard gate: otherwise kill or pivot.

    Gate

  6. Costing & proposal

    Estimate cost and ROI, generate the proposal — with no forbidden promise.

  7. Compliance

    Law 25, risk classification, model card. Hard gate on decisions affecting people.

    Gate

  8. Governed implementation

    Shadow deployment, continuous evaluation, client sign-off before cutover.

  9. Monitoring & learning

    Drift tracking and recalibration of estimates for the next engagements.

Seven validators — five automatic.

Each acceptance criterion relies on one or more of these validators. Humans only touch the risky residue.

RULE

Deterministic rule

Artifact presence, format and coherence.

CALC

Metric vs threshold

Does a metric clear the required threshold.

EVAL

Eval on real data

Does the solution hit the target on real cases — the supreme judge.

LLM-J

Second model as judge

Critiques reasoning, detects forbidden promises.

ANOM

Anomaly detection

Is the estimate an outlier versus calibrated history.

CLIENT

Client sign-off gate

Has the client explicitly accepted.

SAMPLE

Quality sampling

Human review of a risk-weighted subset.

Maker — checker — partner.

Like an audit firm: the field auditor produces, a remote senior reviews, a partner signs off material decisions. Expert load grows as log(N), not N — which is what makes 100 engagements manageable without sacrificing quality.