A useful digital twin is an operating model, not a personality clone. It should reproduce approved knowledge and decision boundaries while exposing uncertainty and handing sensitive decisions back to a person.
01 / CHOOSE THE ANSWERSDefine the small set of questions worth standardising
Begin with recurring, low-ambiguity questions whose answers already exist: onboarding, document location, standard process, approved language, and escalation points.
- Exclude private client detail by default.
- Exclude legal conclusions that require a fresh factual assessment.
- Exclude commitments, payments, publication, or external sending without approval.
02 / MODEL THE RECORDTurn every answer into a governed record
A paragraph alone cannot carry provenance or responsibility. Use stable fields.
Question: Approved answer: Source document or URL: Source excerpt / page: Owner: Effective from: Review date: Audience / permission: Confidence: Escalate when:
03 / SEPARATE LAYERSKeep facts, judgment, examples, and tone apart
Facts change on one timetable, policy on another, and writing style on a third. Separate them so a style update cannot silently change a rule.
- Fact layer: names, dates, product states, rules, and definitions.
- Judgment layer: approved criteria and exceptions.
- Case layer: examples with personal data removed.
- Voice layer: tone and formatting guidance.
04 / HUMAN GATERequire approval where consequences begin
The assistant may retrieve and draft. A named owner approves external advice, a new exception, a legal conclusion, or anything that changes money, rights, privacy, or reputation.
Automation may shorten the route to a decision; it must not erase who owns the decision.
05 / TEST WITH REAL QUESTIONSEvaluate answers, citations, refusals, and escalation
Create a small test set containing ordinary questions, outdated questions, ambiguous cases, permission conflicts, and questions with no answer.
- Does the answer cite the correct record?
- Does it refuse to invent missing facts?
- Does it prefer the latest effective version?
- Does it escalate the right cases to the right owner?
06 / MAINTAINTreat deletion and expiry as product features
Review records on a schedule, mark superseded material, record every edit, and give people a way to correct or remove personal information. A knowledge base that only grows will eventually become confidently wrong.
Primary references
- OpenAI — Retrieval and knowledge guidance — Primary documentation for retrieval-backed applications.
- NIST AI Risk Management Framework — A governance reference for validity, accountability, and risk.
Begin with twenty approved answers, not every file the team owns.
A small governed source is more useful than a vast unreviewed archive.
Back to the field guide Next: traceable consultant skill

READER COMMENTS
Leave the thought this article gave you.
No comments yet. You can leave the first one.