All case studies
IllustrativeProfessional servicesCase study

A controlled intake system for specialist advisory work

An illustrative triage workflow that protects judgement while shortening the route from enquiry to qualified instruction.

Executive brief

This concept gives a specialist advisory firm a consistent intake process. It structures initial information, applies mandatory control gates and prepares a concise matter brief for professional review.

100%mandatory conflict gateBefore instruction
4triage classesDefined service routes
1professional approvalRequired for every matter
01

Situation

Specialist firms often receive enquiries through email, telephone and referral networks. Initial information varies in quality, and senior professionals spend time reconstructing the same facts before they can decide whether and how to proceed.

The opportunity is not to automate professional judgement. It is to improve the completeness, consistency and traceability of the information presented to it.

02

Diagnostic

The intake is decomposed into factual capture, eligibility controls, service classification and professional acceptance. Mandatory controls stop progression when required evidence is missing.

The design also identifies sensitive information that should not enter general-purpose AI services.

  • Collect only information needed for the next decision
  • Separate mandatory controls from commercial prioritisation
  • Keep professional acceptance explicit and attributable
03

Solution design

A secure intake form creates a matter candidate. Deterministic rules check completeness and routing criteria. A language model may draft a summary from approved fields, but it cannot accept the matter or change the control result.

The reviewer sees the original evidence, the structured facts, open questions and the generated brief in one workspace.

04

Risk and governance

The data model applies retention rules, access controls and source attribution. Prompt and output logs support quality review. Sensitive categories can be excluded from model processing or routed to an approved private environment.

A representative evaluation set tests omissions, ambiguous enquiries and attempts to bypass required controls.

05

Expected value

The expected benefit is faster triage with a more consistent evidence base. Pilot measures would include time to first decision, incomplete enquiries, rework and the proportion of generated summaries accepted without material correction.

Operating baseline

From current operating friction to a controlled target state

The diagnostic tests the current state using observable work, then describes the controlled operating state the release is intended to create.

Operating areaCurrent conditionEvidence to collectTarget condition
Factual completenessInitial enquiries vary by channel and referrerSample information available at first professional reviewMinimum facts collected before routing
Control executionChecks may be reconstructed across toolsTrace identity, conflict, eligibility and approval evidenceDeterministic gates with durable results
Professional attentionSenior time spent reconstructing and reformatting factsMeasure preparation and clarification effortOriginal evidence and structured brief in one review
Data boundarySensitive content may enter tools without a common ruleInventory channels, model services and accessApproved processing route by information category
Research context

External evidence sharpens the engagement hypothesis

Findings are paraphrased from the linked original publications. Their scope and populations differ, so they inform the thesis rather than prove a universal outcome.

3 outputs

ICO guidance combines audit methodology, organisational guidance and practical tools

Data protection should be evidenced through design records, tests and operating controls, not policy wording alone.

Source: ICO, Guidance on AI and data protection
Lifecycle

NIST treats generative AI risk as an issue across design, deployment, operation and review

Professional intake controls should be tested before launch and monitored as data, models and use patterns change.

Source: NIST, Generative AI Profile
Delivery work packages

The engagement is organised around four evidence-producing work packages

Each work package ends with an explicit decision and a tangible output. The sequence keeps delivery connected to operating evidence.

01

Intake and control taxonomy

Decision
Which facts and mandatory gates precede professional judgement?
Work
Define fields, matter types, identity, conflict, eligibility and stops.
Output
Approved intake and decision policy.
02

Secure matter-candidate flow

Decision
How is evidence captured and routed with least privilege?
Work
Build intake, document handling, rules and reviewer workspace.
Output
Traceable end-to-end intake release.
03

Bounded AI summary

Decision
Can approved fields reduce review preparation safely?
Work
Evaluate omission, distortion, sensitive data and source fidelity.
Output
Drafting service with source context and no acceptance authority.
04

Assurance and adoption

Decision
Does the release improve triage while preserving professional control?
Work
Pilot representative enquiries, sample decisions and train reviewers.
Output
Acceptance evidence, operating guide and risk review.
Decision architecture

Each operational decision has evidence, control and a measure

DecisionRequired evidenceControlPerformance measure
Is the enquiry complete?Structured facts and supporting documentsMandatory field validationRequests for missing information
Do controls permit review?Identity, conflict and eligibility resultsDeterministic stop gatesControl exceptions
Which service route applies?Matter type, urgency and complexityApproved triage taxonomyReclassified matters
Will the firm accept?Original evidence, controls and professional assessmentNamed professional approvalTime to first decision
Delivery and operating risk

Risks are designed into the operating model before launch

Risks become manageable when the early signal, control and accountable owner are agreed before release.

RiskEarly signalPrimary controlAccountable owner
Generated summaries omit a decisive factReviewers find material facts only in original evidenceSource-visible review and omission testsProfessional owner
Mandatory controls are treated as model judgementSimilar facts produce different stop resultsDeterministic gates outside the modelRisk owner
Sensitive data reaches an unapproved servicePrompt logs contain excluded categoriesData classification and routing policyPrivacy owner
Faster intake shifts delay to acceptancePrepared matters queue without reviewer capacityEnd-to-end service measure and workload routingService lead
Measurement system

Acceptance links the release to observable operating performance

Measures are useful only when their definition is stable and their movement changes a management decision.

OutcomeDefinitionLeading evidenceDecision supported
Complete first reviewEligible matters reach review with required facts and evidenceClarification request rateChange form, taxonomy or guidance
Control integrityMandatory results are correct, attributable and reproducibleControl exception countStop release and correct policy
Preparation effortProfessional time before substantive assessmentReformatting and search minutesImprove workspace or data capture
Summary fidelityDraft accepted without material omission or distortionCorrection categoryRestrict use or improve evaluation
Exhibit 1

Automation supports preparation, not professional acceptance

Illustrative allocation of responsibility across the intake decision.

Structured data captureSystem led
Mandatory control checksRules led
Matter summaryAI assisted
Acceptance decisionHuman led
Source: Quiet Gears illustrative control design. Bar length represents automation suitability.
Exhibit 2

Delivery follows a controlled progression from evidence to operation

01Capture

Gather structured facts and source evidence.

02Control

Apply eligibility, conflict and completeness gates.

03Prepare

Draft the brief and surface unanswered questions.

04Decide

Require accountable professional acceptance.

System blueprint

Policy gates sit outside the language model

01candidate = intake.validate(payload)02controls = policy.check(candidate)03if (!controls.pass) return hold()04brief = model.summarise(approvedFields)05decision = reviewer.accept(brief, evidence)
01Secure intake
02Policy controls
03Approved data view
04Drafting service
05Reviewer decision
Recommended next steps

Move from design to evidence in a bounded release.

  1. Map mandatory and discretionary decisions
  2. Define the approved data boundary
  3. Build a redacted evaluation set
  4. Pilot with one service line and weekly quality review

Where is friction
holding you back?

Talk it through with us