Longhand AI Atlas

Do this

1. 1

   Use the equation: RICE score = (Reach × Impact × Confidence) ÷ Effort.

2. 2

   Reach: count affected people, events, or workflow runs in one shared time period. Do not compare monthly reach with annual reach.

3. 3

   Impact: estimate effect on the shared result. Intercom's example scale is 3 massive, 2 high, 1 medium, 0.5 low, and 0.25 minimal.

4. 4

   Confidence: discount the estimate based on evidence. Intercom's example scale uses 100% high, 80% medium, and 50% low; enter the equation as 1, 0.8, or 0.5.

5. 5

   Effort: estimate total person-time across implementation, data preparation, review, testing, training, governance, and maintenance using one shared unit.

Create a RICE scoring guide for these AI workflow candidates. First choose one reach period, one impact scale tied to the same result, a confidence scale based on evidence quality, and one effort unit that includes implementation and human review. Do not score the candidates until the units are explicit.

Candidates: [paste]

Do this

1. 1

   Attach a source and observation date to reach: workflow logs, transaction counts, support volume, or another traceable record.

2. 2

   Tie impact to the TDAR result and state whether the estimate comes from experiments, comparable cases, user evidence, or judgment.

3. 3

   Set confidence after reviewing the evidence, not after seeing the desired score. Lower confidence when reach, impact, or effort is weakly supported.

4. 4

   Estimate effort across all contributors and recurring operating cost, including human review and exception handling.

5. 5

   Label unknowns and record the cheapest evidence that could narrow them before a build.

Audit this RICE table before calculating. For each factor, identify the source, date, unit, assumption, and uncertainty. Recommend a confidence level only after reviewing the evidence. Include full implementation and operating effort. Do not invent missing data.

RICE table: [paste]

Do this

1. 1

   Identify the least certain or most decision-sensitive factor for the top candidates.

2. 2

   Recalculate a plausible low and high case. If Candidate B effort rises from 2 to 4 person-weeks, its score falls from 75 to 37.5 and Candidate A becomes first.

3. 3

   State whether the ranking is stable, fragile, or tied and what evidence would reduce the important uncertainty.

4. 4

   Apply constraints outside RICE: strategic fit, safety, legal or policy restrictions, dependencies, capacity, equity, and the cost of failure.

5. 5

   Do not use RICE to launder a prohibited or unsafe option into consideration; remove ineligible candidates before ranking or explicitly gate them.

For the top two RICE candidates, identify the most uncertain decision-relevant inputs. Calculate plausible low and high cases, state whether the ranking is stable or fragile, and recommend either decide now or gather specific evidence first. Apply safety, policy, dependency, capacity, and strategic constraints separately from the score.

Table and evidence: [paste]

Do this

1. 1

   Choose at least three recurring workflow candidates tied to one result and reach period.

2. 2

   Define the scales and units before scoring. Attach a source or assumption to every input.

3. 3

   Calculate the scores and show the equation for each candidate. Check that confidence is entered as a decimal.

4. 4

   Run a plausible sensitivity case for the top two candidates and apply eligibility constraints.

5. 5

   Select the next candidate to test, defer, reject, or investigate. Explain what evidence would reverse the decision and save the complete brief.

# RICE AI Workflow Decision Brief

## Shared result and scoring rules
[Result, reach period, impact scale, confidence rule, effort unit]

## Candidate table
| Candidate | Reach + source | Impact + evidence | Confidence + rationale | Full effort + basis | RICE score |
|---|---:|---:|---:|---:|---:|

## Calculations
[Show each substitution into (R × I × C) ÷ E]

## Sensitivity analysis
[Low/base/high case for the top two; stable, fragile, or tied]

## Eligibility and constraints
[Safety, policy, authority, dependencies, capacity, strategic fit]

## Decision
[Test, defer, reject, investigate, or choose an alternative; explain why]

## Evidence that would reverse the decision
[Specific evidence and how it would change an input or constraint]

Do this

1. 1

   Result: name the business or service outcome as a number, rate, time, quality measure, or observable state—not as 'use AI.'

2. 2

   Baseline: identify the current value, source, period, and conditions so the later comparison is meaningful.

3. 3

   AI contribution: state the bounded part of the workflow the system will perform or assist.

4. 4

   Human contribution: state the judgment, approval, relationship, accountability, or exception handling that remains human.

5. 5

   Remeasurement: set the date, sample, owner, and comparison rule before the experiment begins.

Help me draft a TDAR Results Brief for this recurring workflow. Ask for one missing fact at a time. Do not invent a baseline. Use these fields: measurable result, current baseline and source, bounded AI contribution, accountable human contribution, and remeasurement date/sample. End with a section called 'Evidence still missing.'

Workflow: [describe it]
Current evidence: [paste or say unknown]

Do this

1. 1

   Choose a result measure already connected to the work: cycle time, first-pass accuracy, corrections, throughput, cost, satisfaction, or another operational measure.

2. 2

   Name the source, period, sample size, exclusions, and unit. Preserve the raw record or a reproducible query when permitted.

3. 3

   Record companion measures that could reveal displacement, such as faster drafting paired with more review corrections.

4. 4

   If evidence is unavailable, say 'baseline unknown' and specify a collection window. Do not turn an assumption into a historical fact.

Audit this proposed AI baseline. Separate observed facts from estimates, identify comparability problems, and list the minimum data needed before a result claim is credible. Do not fill missing values.

Proposed baseline: [paste]
Planned result: [paste]

Do this

1. 1

   Walk the workflow step by step and identify which inputs the AI may access and which actions it may propose or perform.

2. 2

   Mark human judgment at the step where significance, truth, risk, relationship, or professional authority is decided—not only at the end.

3. 3

   State what deterministic checks or source verification must pass before the human review.

4. 4

   List external changes the system may not make without explicit approval, including sending, filing, publishing, purchasing, deleting, or changing a system of record.

Map this workflow into three columns: bounded AI contribution, accountable human contribution, and deterministic validation. For each external or high-impact action, add the required approval and a stop condition. Do not assume the AI has authority.

Workflow: [paste steps]

Do this

1. 1

   Choose a recurring workflow with an authorized practice context. If you cannot use workplace information, use the provided incident-report case with different targets and justify the changes.

2. 2

   Complete all five TDAR fields and attach or describe the baseline source. Label every estimate and unknown.

3. 3

   Add one quality, safety, or customer guardrail that prevents apparent improvement from hiding a worse outcome.

4. 4

   Write continue, revise, and stop decision rules for the remeasurement date.

5. 5

   Self-check the brief for outcome value, evidence quality, reliability, safety, and reflection. Save the artifact for later assessment preparation.

# TDAR Results Brief

## Workflow and decision
[What recurring work is being considered, and what decision will this experiment inform?]

## Measurable result
[Measure, target, unit, population, and time period]

## Baseline
[Current value or 'unknown'; source; period; sample; exclusions; guardrail]

## Bounded AI contribution
[Authorized inputs, work performed, outputs, and prohibited actions]

## Accountable human contribution
[Judgments, reviews, approvals, and owner of consequences]

## Remeasurement
[Date, sample, comparison method, continue/revise/stop rules]

## Evidence still missing
[Facts that must be collected or verified]

## Reflection
[Which assumption is most likely to change the decision?]

Do this

1. 1

   Big What: write the consequential decision that must be made now—build, test, defer, reject, or choose another path.

2. 2

   Why: state the result, evidence, urgency, and strategic reason that make the decision worth attention.

3. 3

   What: define the workflow, users, inputs, outputs, boundaries, and finish line precisely enough to evaluate.

4. 4

   How: describe the smallest credible operating model, including people, AI surface, data, deterministic checks, effort, and permissions.

5. 5

   What If: expose uncertainty, failure modes, alternatives, stop conditions, and what evidence would change the decision.

Help me open a DMP decision record for this AI opportunity. Ask one question at a time in this order: Big What, Why, What, How, What If. Separate observed evidence from assumptions. Always include the current/manual process as a possible alternative. Do not recommend a build until every section has enough evidence to review.

Opportunity: [describe]
TDAR result: [paste]

Do this

1. 1

   Write the decision as one sentence with a verb, object, boundary, and timing: for example, 'Run a four-report pilot of source-grounded incident drafting; do not automate approval or publishing.'

2. 2

   Connect the Why to the TDAR result and baseline. Identify which evidence is observed, estimated, or missing.

3. 3

   State why the decision is needed now and what happens if no change is made.

4. 4

   Define the smallest reversible commitment that can answer the important uncertainty before a larger build.

Review this Big What and Why. Identify vague verbs, hidden scope, unsupported urgency, missing baseline evidence, and irreversible commitments. Rewrite it as the smallest reversible decision that can test the most important uncertainty.

Big What: [paste]
Why: [paste]
TDAR: [paste]

Do this

1. 1

   Define the user, trigger, authorized inputs, produced artifact, downstream consumer, and explicit non-goals.

2. 2

   Choose the simplest AI surface that can access the required information and perform only the authorized actions.

3. 3

   Add deterministic checks for schemas, totals, required fields, citations, or invariants where exactness is testable.

4. 4

   Place human judgment and approvals at the relevant steps and include their real time in the effort estimate.

5. 5

   Describe failure handling, evidence capture, and how the workflow returns to a known state.

Turn this proposed scope into a bounded operating model. Use sections for user and trigger, authorized inputs, AI work, deterministic validation, human judgment, external actions and approvals, failure handling, evidence retained, non-goals, and full effort. Flag any permission or ownership assumption.

Proposed What/How: [paste]

Do this

1. 1

   List the assumptions most capable of changing the decision: demand, data quality, permissions, reliability, review capacity, harm, cost, or adoption.

2. 2

   For each, define an observable signal and the earliest safe point at which it can be detected.

3. 3

   Specify continue, revise, pause, and stop responses with an accountable person for the decision.

4. 4

   Compare at least one non-AI or lower-complexity alternative and the option to keep the current process.

Convert these What If concerns into a table with assumption, evidence, early signal, consequence, continue/revise/pause/stop rule, accountable human, and alternative. Add the current/manual process as a comparator. Avoid generic risks without detection or response.

Concerns: [paste]

Do this

1. 1

   Choose a candidate workflow and define the immediate decision. Use a synthetic or de-identified case when authorization is limited.

2. 2

   Complete Big What, Why, What, How, and What If with facts, assumptions, alternatives, and source references separated.

3. 3

   Select build, bounded test, defer, reject, or non-AI alternative. Explain why the rejected options are weaker under the current evidence.

4. 4

   Add the evidence that would reverse the decision and the date or event that triggers reconsideration.

5. 5

   Save the DMP record beside the TDAR brief for the planned Opportunity Diagnosis assessment.

# DMP Decision Record

## Big What
[Exact decision, boundary, and timing]

## Why
[TDAR result, baseline evidence, urgency, observed facts, assumptions]

## What
[User, trigger, workflow, inputs, outputs, finish line, non-goals]

## How
[People, AI surface, data, deterministic checks, permissions, review, effort, failure handling]

## What If
[Decision-changing uncertainties, signals, continue/revise/pause/stop rules, owners]

## Alternatives compared
[Current process, non-AI option, lower-complexity AI option, proposed option]

## Decision and rationale
[Build, bounded test, defer, reject, or alternative]

## Evidence that would reverse this decision
[Specific evidence]

## Reconsideration trigger
[Date or event]

Do this

1. 1

   Name the authorized purpose and the result being pursued. List secondary uses that are not authorized without a new review.

2. 2

   Identify the accountable human for the workflow, data decision, professional judgment, external action, and incident response. A model cannot hold these accountabilities.

3. 3

   List applicable organizational policies, contractual limits, jurisdictional questions, and professional standards. Mark unknown requirements for qualified review.

4. 4

   Define documentation, change control, review cadence, and who can pause or terminate the workflow.

5. 5

   Prohibit impersonation, fabricated authority, undisclosed external action, and representations that AI output received a review it did not receive.

Create the Govern section of an AI risk-control record for this workflow. Include authorized purpose, prohibited secondary uses, accountable humans by decision, applicable policies/contracts/jurisdiction questions, documentation, change approval, review cadence, and pause/termination authority. Mark unknown requirements for qualified review; do not give legal conclusions.

Workflow: [paste]

Do this

1. 1

   Inventory every input, generated output, metadata source, connection, retention location, recipient, and downstream system. Classify sensitivity and authorization for each.

2. 2

   Minimize data to what the purpose requires. Identify consent, notice, confidentiality, licensing, copyright, provenance, and deletion questions.

3. 3

   Map people who receive benefit, bear error, are represented in data, are evaluated, or may have limited ability to contest the output.

4. 4

   Classify every action as read-only, reversible, approval-gated, external, or prohibited. Verify actual permissions rather than inferring them from tool capability.

5. 5

   Describe foreseeable misuse, out-of-scope use, dependency, and downstream effects—including accessibility barriers and disparate error consequences.

Create the Map section of an AI risk-control record. Build tables for data/source/authorization/sensitivity/retention, affected people and possible impacts, and action/reversibility/approval. Include consent, confidentiality, provenance, copyright or licensing, accessibility, contestability, and foreseeable misuse. Mark unknowns as blockers; do not assume consent or permission.

Workflow: [paste]

Do this

1. 1

   Define quality measures tied to the task: claim support, completeness, classification accuracy, correction rate, safe refusal, accessibility, or another observable result.

2. 2

   Create representative normal, boundary, adversarial, missing-data, and prohibited-action cases. Include groups or contexts that could experience different error rates or consequences.

3. 3

   Specify acceptance thresholds before running the test and retain inputs, outputs, versions, reviewer decisions, and failures needed for inspection.

4. 4

   Test human review: give the reviewer the information, criteria, time, and authority needed to catch consequential errors; measure review workload and misses.

5. 5

   Record uncertainty and residual risk. Passing a finite test set does not prove universal safety or correctness.

Create the Measure section for this AI workflow. Define task-specific quality and safety measures, normal/boundary/adversarial/missing-data/prohibited-action cases, affected-group checks where relevant, predeclared thresholds, human-review tests, retained evidence, and residual uncertainty. Do not claim universal safety from a finite test set.

Workflow and context map: [paste]

Do this

1. 1

   Choose layered controls: data minimization, least privilege, read-only access, allowlists, deterministic validation, approval gates, rate or scope limits, isolation, monitoring, and fallback procedures.

2. 2

   Assign each residual risk to accept, mitigate, transfer, avoid, or escalate—with an accountable human and rationale.

3. 3

   Define monitoring signals, incident severity, containment, notification, evidence preservation, recovery, and post-incident review.

4. 4

   Set change triggers for new model versions, sources, tools, permissions, users, jurisdictions, or consequences. Require regression testing before material changes enter use.

5. 5

   Define retirement: revoke connections, stop schedules, archive required evidence, delete data under policy, and inform affected users when appropriate.

Create the Manage section of an AI risk-control record. For each material risk, list prevention, detection, human review, response, evidence, owner, residual risk, and accept/mitigate/transfer/avoid/escalate decision. Add monitoring, incident, change-regression, and retirement procedures. Reject controls that are only disclaimers.

Mapped risks and test results: [paste]

Do this

1. 1

   Classify the consequence of error and identify whether the work affects legal rights, finances, health, safety, employment, education, security, privacy, benefits, or another high-stakes interest.

2. 2

   Use AI for bounded preparation—organizing provided sources, extracting terms, identifying questions, or drafting for review—only when data and use are authorized.

3. 3

   Do not claim to be a lawyer, clinician, financial professional, security authority, or other licensed expert. Do not label output privileged, approved, compliant, or professionally reviewed unless a qualified person actually made that determination.

4. 4

   Require authoritative source verification, jurisdiction and date checks, uncertainty disclosure, and qualified review before reliance or external action.

5. 5

   Stop when authority, consent, facts, source currency, policy, jurisdiction, reviewer availability, or the ability to contest a decision is unclear. Record the escalation and do not continue by assumption.

Review this proposed AI workflow for high-stakes boundaries. Identify affected rights or consequences, what bounded preparation AI may perform, required authoritative sources, qualified professional review, prohibited representations or actions, stop conditions, and the escalation record. Do not provide the professional conclusion itself.

Workflow: [paste]