Speed
Respond sooner
Prepare a useful next step without starting every inquiry from zero.
Human-AI Interaction · Agentic Workflow Design
R.A.I.N.E.
Responsive AI for Notifications and Engagement
Designing an AI-assisted intake workflow that balances interpretation, deterministic system logic, and human control.
R.A.I.N.E. — Responsive AI for Notifications and Engagement — is a working intake prototype I designed for ForwardVantage.ai. It interprets a prospect’s inquiry, identifies what has already been explained, determines what context is still missing, and prepares a relevant follow-up for human review.
The central design decision was to avoid treating the model as the whole system. Predictable workflow behaviour stays deterministic, AI is used for semantic interpretation and preparation, and a person retains authority over AI-authored customer communication.
Human in the LoopAgent BehaviourWorkflow OrchestrationConversational AI
Responsibility-led workflow
Interpretation, execution, and authority stay distinct.
- System
- AI
- Human
- System
Inquiry
A prospect selects a service and describes the need.
- System
Normalize + validate
Known data rules determine whether the input can continue.
- AI
Interpret
The inquiry is compared with the selected information model.
- AI
Prepare
A structured interpretation and draft are proposed.
- Human
Review
A person can approve, revise, override, or cancel.
- System
Authorized action
The workflow executes the reviewed outcome and records it.
Target recovery behaviour: unusable input should exit before AI interpretation; revision returns to human review; override and cancellation remain human-controlled paths.
- Role
- Human-AI interaction and workflow designer
- Focus
- Autonomy, behaviour, orchestration, and review
- Project
- Independent operational prototype
- Evaluation
- Retrospective testing with synthetic scenarios
The design challenge
Faster intake without making it less human
How could I create a more responsive intake experience without sacrificing relevance or human oversight?
Prospective clients arrive with very different levels of clarity. One person may describe a specific workflow, the systems involved, and the outcome they need. Another may only know that part of their work feels repetitive or inefficient.
A fixed acknowledgement can respond quickly, but it does little to move the conversation forward. A fully autonomous response can be more adaptive, but gives probabilistic output too much control over an important customer interaction.
The design challenge was to balance three needs:
- Speed: reduce the delay before a useful follow-up is ready.
- Relevance: respond to what the prospect actually said and avoid redundant questions.
- Human control: keep consequential communication reviewable and correctable.
The goal was not to remove the human from intake. It was to reduce the work required before human judgment became necessary.
Relevance
Respond to the inquiry
Recognize supplied context and ask only for useful gaps.
Human control
Keep action reviewable
Let AI prepare while a person remains accountable.
System design
Giving each responsibility the right kind of intelligence
I assigned each part of the workflow according to the kind of decision it makes.
System — predictable execution
The workflow handles known operations such as normalizing input, validating a submission, routing from the service selected by the user, managing review states, and recording outcomes. If the system already knows the answer, asking AI to infer it adds uncertainty rather than value.
AI — interpretation and preparation
AI handles the parts that benefit from semantic reasoning: understanding free-form language, recognizing information already supplied, identifying gaps, choosing relevant follow-up questions, preparing a draft, and responding to reviewer direction.
Human — judgment and authority
A reviewer can inspect the source inquiry, the model’s interpretation, and the proposed response. They can approve it, request a revision, replace it manually, or cancel the action.
Human authorization is required for AI-authored prospect communication. Any predefined recovery message belongs to a separate, deterministic path with pre-approved copy.
AI for ambiguity. System logic for certainty. Human judgment for consequence.
SystemHandles what is known
Predictable execution
- Normalize input
- Validate
- Route
- Manage state
- Record outcome
AIHandles ambiguity
Interpretation + preparation
- Understand inquiry
- Find known context
- Identify gaps
- Select questions
- Prepare draft
HumanHandles consequence
Judgment + authority
- Inspect
- Direct
- Approve
- Take over
- Cancel
Agent behaviour and conversation design
An information model, not a fixed questionnaire
R.A.I.N.E. uses the service a prospect selected as an explicit routing signal. It does not ask a model to guess which service the person needs.
Each path applies a different information model:
- AI Agents: task, systems, users, outcome, and project stage.
- Automations and Workflows: process, trigger, tools, and the current bottleneck.
- AI Opportunity Audit: business area, inefficient workflows, desired outcome, and type of support.
- Exploratory: repetitive work, delays, handoffs, and the improvement the person hopes to see.
These are information goals, not fixed questionnaires. The model compares them with what the prospect already shared, marks topics as covered or missing, and asks only for useful gaps.
When someone is uncertain about the solution, the conversation focuses on their experience of the problem instead of requiring them to define the technology.
Profile 01
AI Agents
- User signal
- A defined agent opportunity
- Information goals
- Task · systems · users · outcome · stage
- Conversation
- Focused and solution-oriented
Profile 02
Automations
- User signal
- A process to improve
- Information goals
- Process · trigger · tools · bottleneck
- Conversation
- Operational and process-focused
Profile 03
Opportunity Audit
- User signal
- A search for where AI can help
- Information goals
- Business area · friction · outcome · support
- Conversation
- Strategic and diagnostic
Profile 04
Exploratory
- User signal
- A problem without a defined solution
- Information goals
- Repetition · delays · handoffs · improvement
- Conversation
- Approachable and lower-friction
Fixed questionnaire
Question 1 → 2 → 3 → 4
Everyone receives every question.
R.A.I.N.E. information model
Known ✓ · Known ✓ · Missing ? · Missing ?
Ask only about what is still unknown.
Human-AI control
AI prepares. A human authorizes the action.
I treated autonomy as a spectrum rather than an on-or-off feature. Internal, reversible preparation can have more autonomy; an external action that represents the business requires stronger human control.
The review model supports four distinct choices:
- Approve: accept the proposed response.
- Request revision: give the AI direction and review another version.
- Manual override: stop using the AI draft and replace it directly.
- Cancel: stop the action entirely.
Meaningful review requires more than a polished email and a yes-or-no prompt. The reviewer needs to see the prospect’s original message, what the model believes is covered or missing, what questions it selected, and what it proposes sending.
Correction is therefore part of the interaction, not evidence that the workflow has failed. Human control includes the ability to guide the AI, take over from it, or stop collaborating with it altogether.
Human review checkpoint
Review the source, interpretation, and recommendation before authorizing communication.
Source · prospect context
- Selected service
- Example service path
- Original inquiry
- Privacy-safe inquiry summary placeholder
- Submission state
- Validated and ready for interpretation
Interpretation · recommendation
- Covered topics
- Visible to the reviewer
- Missing topics
- Shown as distinct items
- Selected questions
- Limited to useful gaps
- Proposed response
- Prepared for review
- 01
Approve
Use the reviewed draft
- 02
Request revision
Guide the AI again
- 03
Manual override
Replace the draft
- 04
Cancel
Stop the action
System boundaries
The model interprets. The workflow controls. The record persists.
A generated paragraph is useful to a person, but it is a weak contract between system components. R.A.I.N.E. separates the customer-facing draft from structured interpretation such as the proposed subject, covered topics, missing topics, and questions selected.
This makes the model’s recommendation inspectable without turning it into operational truth:
- The AI layer interprets: what appears to be known, what may be missing, and what response could help.
- The workflow layer controls: routing, review state, revision, cancellation, and authorized execution.
- The system of record persists: prospect details, selected service, workflow status, and outcome.
The revision path also needs the original inquiry, the previous draft, reviewer feedback, the current qualification state, and the behavioural constraints of the selected service. Passing conversation content alone is not enough if the next step loses the rules that shaped the original interaction.
- AI layer
Interprets
Covered topics · missing context · relevant questions · proposed response
- Workflow layer
Controls
Routing · review state · revision · cancellation · authorized execution
- System of record
Persists
Prospect details · selected service · status · response outcome
Evaluation and iteration
Testing the boundaries, not just the writing
Early testing happened iteratively while I built the prototype. I later ran a structured retrospective evaluation using synthetic inquiry scenarios across the four service paths and the main control branches.
The evaluation examined more than writing quality. It asked whether the system behaved correctly where deterministic rules, AI interpretation, human intervention, and workflow state meet.
Scenarios included partially answered inquiries, vague requests, fully qualified inquiries, meaningless input, human-directed revision, manual takeover, and cancellation.
The strongest results were adaptive qualification, distinct behaviour across service paths, avoiding redundant questions, manual takeover, and cancellation. The most useful failures appeared at system boundaries and led to the iteration work below.
This was a prototype evaluation, not a production study, and it did not measure business impact.
Retrospective evaluation · synthetic scenarios
| Scenario | Expected | Observed | Status |
|---|---|---|---|
| Partially answered inquiry | Ask only for missing information | Skipped topics already present | Pass |
| Vague inquiry | Clarify without inventing a solution | Asked relevant, bounded questions | Pass |
| Fully qualified inquiry | Avoid redundant questions | Moved toward a next step | Pass |
| Meaningless input | Exit before AI processing | Exposed a validation boundary defect | Pending verification |
| Reviewer-directed revision | Preserve context and reviewer direction | Context repaired; behaviour needs refinement | Needs refinement |
| Manual override + cancellation | Support takeover or stop the action | Both remained distinct control paths | Pass |
Iteration 01 · deterministic boundary
Redesign pending public verificationMeaningless input reached the AI path
A synthetic submission containing only punctuation exposed a weakness in the deterministic intake boundary. It was technically non-empty and initially reached the AI qualification path.
The investigation identified three related risks: inconsistent input shape between steps, a validation rule that treated any non-empty value as meaningful, and mismatched value types in branching logic.
The resulting design direction was to normalize input before validation, then check for meaningful content before an inquiry can enter the AI path. That revised sequence still requires final implementation verification and retesting.
The design lesson was clear: a safeguard is not part of the product until its boundary conditions have been tested. This case study therefore presents the change as an open verification item rather than a completed production safeguard.
Observed
Non-meaningful input crossed into interpretation.
Target boundary
Normalize, validate meaning, then route.
Iteration 02 · human-review handoff
Behaviour needs refinementContext survived; constraints still drifted
A second synthetic scenario tested whether reviewer direction survived a pause-and-resume review step. The reviewer asked for a shorter response, a softer next step, and a more focused conversation.
The first revision attempt received the feedback but lost earlier workflow context at the handoff. After the context references were repaired, the revision preserved the original inquiry and applied the requested direction.
One behavioural issue remained: the revision introduced broader discovery questions than the reviewer requested. The next refinement is to carry the selected service’s qualification state and behavioural constraints into the revision step, not only the previous text.
Context continuity does not automatically guarantee behaviour continuity.
InquiryPrevious draftReviewer directionBehaviour constraints
+
Image placeholder · privacy review required
Invalid-input evaluation evidence
The source pair that exposed the validation-boundary defect and shows the recovery response used in the retest.
Add only after cropping, redaction, and publication approval
+
Image placeholder · privacy review required
Revision-handoff evaluation evidence
The source pair used to review context retention and the remaining behaviour-constraint drift.
Add only after cropping, redaction, and publication approval
Practical prototyping
Testing the interaction where it actually lives
I implemented R.A.I.N.E. as a working operational prototype rather than stopping at a conceptual flow. Building the workflow made it possible to test the interaction at the points where AI products often become fragile: input validation, specialized behaviour, structured outputs, review state, context handoffs, recovery, and human intervention.
The prototype connects an inquiry experience, an orchestration layer, specialized AI behaviour, human review, and an operational record. The public case study intentionally omits code, prompts, configuration, identifiers, private data, and unsanitized implementation captures.
The interactive architecture diagram is a privacy-safe abstraction of the implementation. Source artifacts remain local review material until each has been redacted and approved for publication.
Experience
Inquiry capture
Orchestration
Rules, routing, and review state
AI
Interpretation and preparation
Operations
Human review and persistent record
+
Image placeholder · privacy review required
Planning to implementation
The planning board is the clearest visual of how the service model, interaction states, and workflow layers developed.
Add only after cropping, redaction, and publication approval
Inquiry is normalized and validated before routing. Meaningful input moves to AI interpretation and structured preparation, then human review. A person can approve, revise, override, or cancel. Only an approved or human-authored response can reach authorized communication and the persistent record.
Outcome and reflection
Designing the boundaries around AI
The difficult part was not getting a model to produce a plausible response. It was designing the boundaries around interpretation, authority, recovery, and control.
The project reinforced four principles:
- Use AI for ambiguity, not certainty. Explicit user choices and known workflow state should remain deterministic.
- Give reviewers enough context to exercise judgment. Fluency alone is not evidence that an interpretation is correct.
- Design correction, takeover, and cancellation as first-class interactions. Human control must extend beyond approval.
- Preserve behaviour as well as information across handoffs. Context is incomplete when the next step loses the rules that shaped it.
R.A.I.N.E. demonstrates a human-led approach to AI-assisted service design: the system executes what is known, AI prepares what requires interpretation, and people remain accountable for consequential action.
- Human need
- Deterministic system
- AI interpretation
- Human judgment
- Authorized action
Review, failure, or new information can return the process to AI revision, human takeover, or cancellation.