An extract from a task evaluation record, prepared with synthetic evidence.
AI earns its place by serving the business.
Seichō’s AI consultancy starts with the commercial objective. Automation and well-made software do most of the work; AI is chosen where interpretation improves the operating task.
- Bases
- Dublin, Belfast and London
- Clients
- Scaling and established businesses across Ireland and the UK
- Decision
- Whether a model improves the work enough to justify its cost and oversight
AI consultancy connects the model to a bounded task.
A bounded model task sits inside an accountable workflow.
Key
The model has a defined task and restricted authority.Ordinary automation carries approved records onward.
The system in words
This illustrative extraction workflow receives permitted enquiry records. A model proposes a classification and structured fields. The application validates the fields and holds the consequential action for a named reviewer, who can correct or reject the proposal. Only an approved instruction passes into the operating record through ordinary software and automation. Invalid output and supplier failure enter an owned exception queue.
Triage, drafting, extraction and classification are suitable candidates where source material varies. The illustrative map follows an enquiry into a proposed record, with human approval before a consequential action.
Defined rules belong in ordinary software.
Calculations, scheduled transfers, required fields and known approval rules run through software and automation. A model adds review effort and usage cost; the investment case includes both before any build is approved.
The business retains the authority to decide.
A model receives restricted sources and tools. The application validates output and holds consequential actions for a named reviewer. Rejected output, supplier failure and repeated errors enter an exception route with an accountable owner.
Evaluation addresses the work the model is meant to do.
SeichōAI consultancy / Synthetic extract
Task evaluation record
The model’s output is assessed against the operating task.
- Proposed task
- Extract the quantity, format and delivery date from an enquiry into a draft operating record.
- Representative case
- A message contains two quantities and no confirmed delivery date. The source remains visible to the reviewer.
- Expected behaviour
- The draft preserves both quantities and marks the date as missing. The model cannot invent a date or release the order.
- Acceptance evidence
- The evaluation records extraction accuracy, reviewer correction time and rejected outputs against the agreed test set. The process owner accepts the operating threshold.
- Failure and authority
- An invalid response or unavailable model enters the exception queue. A named person owns the queue, and approval remains required before the consequential action.
- Operating decision
- The investment case includes model usage, review effort and correction cost. A rule-based alternative is assessed against the same objective.
Representative cases include incomplete inputs and conflicting evidence. Acceptance records the required accuracy, correction effort and operating cost. Governance defines authority, change review, the action record and the tested way back.
450+messages handled semi-autonomously each month, saving more than 30 hours
Message volume from the system's logs, averaged over recent months. The time saving is estimated from handling time per message.
The concierge combines model output with human approval.
The messaging concierge connects assessment responses to suggested replies and a review queue. An outbox carries approved work into the sales system.
Read the messaging concierge study.
The B2B print study provides a different application: inbox triage extracts print specifications into a numbered lead. Versioned quotes, production records and invoice connections carry the work onward.
The operating record determines the right tool.
- Where a model earns its place
- Variable enquiries, unstructured documents and drafting tasks can justify model interpretation. The evaluation tests whether the proposed output improves the agreed task once review and correction are included.
- Where automation fits
- A known rule with a dependable input belongs in process automation. Arithmetic and permissions remain explicit software behaviour, with a defined exception when the input fails validation.
- Where judgement remains human
- The business retains decisions requiring professional judgement or commercial authority. A model can prepare material for review within the agreed boundary; the authorised person determines the outcome.
- Participation and evidence
- The client supplies permitted sources, representative cases and process expertise. A sponsor owns the commercial objective, and named reviewers establish acceptable output and approval responsibilities.
Process automation carries approved records through the rest of the workflow. The model’s role stays narrow enough to evaluate and govern.
The Diagnostic establishes the investment decision.
The Operating Model Diagnostic examines the workflow, records and controls before recommending an investment. It establishes whether to build, buy, defer or stop. Accepted equivalent evidence can support an Operating Plan.
The firm connects advice to delivery and continuing operation for manufacturing, professional services, healthcare, energy and distribution.
- Operating Model DiagnosticA data map, ranked opportunities and a signed recommendation.2 to 3 weeks from agreed access and interviews.
- Operating PlanAn investment plan, target architecture and first-release specification.4 to 6 weeks after a Diagnostic or accepted equivalent evidence.
- BuildWorking software, acceptance evidence and an operating runbook.First-release planning range: 8 to 12 weeks.
- StewardshipA named service owner, monitoring and operating reviews.6-month initial term, then quarterly renewal.
Each commission has a defined scope and approval. Fees are fixed and set out in a written proposal. The briefing establishes scope.