Skip to content
Nithya Parepally
Enterprise Data & AI Platform
Under progress

AI Experience Design Framework

A unified, intelligent, self-service entry point for all data & AI platform interactions, discover services, request access, provision resources, read documentation, monitor requests, and get AI-assisted guidance from a single experience. This board is the design process behind it.

01

Empathize

Understand the human, the context they work in, and the intent behind building this at all.
01

Human

Who is the user?

Data Practitioners

People who need to discover, request, access, and consume data platform services to accomplish technical or analytical work.

Business Stakeholders

People who need visibility into platform capabilities, adoption, roadmap, and the broader data ecosystem to support planning, investment, and business decisions.

What are they trying to accomplish?

Data Practitioners

Discover the right data/platform service, understand whether it meets their need, request access or resources, and get to a usable state with minimal dependency on platform teams.

Business Stakeholders

Understand what the platform provides, how it is being used, where demand exists, and how the data ecosystem is evolving so they can make informed business and investment decisions.

User journey — where AI actually touches it

The same Intent → Discover → Decide → Request → Consume → Feedback flow named under “Where do they need assistance?”, walked stage by stage against what AI actually does at each one — and what stays human.

01Intent

User does

States what they want to do, in their own words — not a form field, a sentence.

Human-only moment

None yet. This moment stays entirely human — AI is listening, not deciding.

02Discover

User does

Sees the catalogue narrowed to what's actually relevant, instead of searching it alone.

AI touchpoint

Interprets the stated intent, searches the catalogue, cross-references usage, and surfaces options worth considering.

03Decide

User does

Chooses a path, with the reasoning behind each option visible.

AI touchpoint

Explains the recommendation and guides the next step — the decision itself stays human.

04Request

User does

Reviews an already-assembled request and sends it in.

AI touchpoint

Prepares the request from what's already established and validates it against policy before a human submits and approves it.

05Consume

User does

Gets what they asked for, and can see it's actually working.

AI touchpoint

Provisions the approved change and monitors it afterward.

06Feedback

User does

Finds out quickly if something's gone wrong, and who's handling it.

AI touchpoint

Detects the deviation and raises it — a human still decides what happens next.

What are their pain points?

Discovery

Difficult to find the right service within a dense catalogue.

Service names and technical terminology don't always map to the user's mental model.

Users may know what they want to accomplish, but not which platform service enables it.

Decision-making

Users need help determining which service is appropriate for their use case.

Comparing services requires understanding technical capabilities, prerequisites, limitations, ownership and policies.

Business visibility

Platform capabilities, adoption and ecosystem information may be fragmented.

Business stakeholders need a higher-level view rather than navigating individual technical services.

Requesting

Users have to translate their intent into technical request details.

Users may not understand why certain information or approvals are required.

Request forms can require information the user doesn't naturally think in terms of.

Progress & visibility

Once a request is submitted, users may not understand what is happening.

When something fails, users may not know what failed, why, or what they can do next.

Dependencies, approvals and provisioning steps can be opaque.

What decisions do they make? (Who & What?)

Data practitioners decide
  • What data/platform capability they need.
  • Which service best fits their use case.
  • Whether a service meets their requirements.
  • What resources/access they need.
  • Whether they are eligible to use a service.
  • What configuration is appropriate.
  • Whether to submit a request.
  • Whether to accept an AI recommendation.
  • What to do when a request fails or requires intervention.
Business stakeholders decide
  • Which capabilities are available and relevant.
  • Where platform adoption/demand is growing.
  • Where gaps exist.
  • What capabilities may need investment.
  • Which initiatives or roadmap priorities matter to the business.

What do they need to delegate?

Data Practitioners need to delegate

information discovery and service selection to platform experts, while request creation, submission and follow-through remain largely manual.

Business Stakeholders need to delegate

ecosystem discovery, platform insights while synthesizing that information for planning and decision-making remains largely manual.

Where do they need assistance?

Data Practitioners

IntentDiscoverDecideRequestConsumeFeedback

Business Stakeholder

DiscoverUnderstandGain visibilityIdentify gapsPlan
02

Context

CO-CREATING/COLLABORATION WITH OTHER STAKEHOLDERS

What environment are they working in?

A standalone enterprise self-service application that acts as a single entry point to the organization's Data & AI platform ecosystem.

What systems do they use?

Data & AI platform services & Data catalogues
Identity and access management
Documentation / knowledge repositories
Usage / adoption analytics
Request / approval systems
Resource provisioning systems
Possibly ticketing or support systems — AI doesn't have access to them?

What data do they have?

User context

  • Role
  • Team
  • Permissions
  • Existing access
  • Previous requests

Service knowledge

  • Capabilities
  • Requirements
  • Availability
  • Ownership
  • Documentation
  • Policies

System state

  • Request status
  • Provisioning status
  • Errors

Organizational context

  • Platform roadmap
  • Adoption data

What constraints exist?

Access and permission boundaries

Data sensitivity

Organizational policies

AI capability limitations

Incomplete or conflicting information

Human approval requirements

Potentially irreversible actions

03

Intent

What outcome are we trying to create?

Move users from knowing what they want to achieve to successfully achieving it with less searching, less technical translation, less manual coordination and greater visibility throughout the journey.

Why AI?

Because users often know the outcome they want, but not the platform capability, service or technical configuration required to achieve it. AI can bridge that gap by understanding intent, synthesizing platform knowledge, recommending appropriate services, preparing actions, and helping users recover when things go wrong.

Why now?

As the Data & AI ecosystem grows, the number of services, dependencies, policies and configuration choices increases. A traditional catalogue and form-based self-service model increasingly places the burden of understanding the platform on the user. AI creates an opportunity to shift from users navigating the platform to the platform helping users navigate it.

02

Define

Defines the Human–AI relationship.
STEP 1

What are the actions? Write them as verbs only.

DelegateUnderstandDiscoverAnalyseRecommendDecideExplainGuidePrepare RequestValidateSubmit RequestApproveProvisionMonitorDetect FailureEscalate
STEP 2 · TABLE A

Who should perform each?

ACTOR
HUMAN
AI
WHY
Delegate
The user states intent and scope — only a human can define what's actually wanted.
Understand
Interpreting free-form intent into structured meaning is what the model is for.
Discover
Searching the catalogue and cross-referencing usage is a retrieval task, not a judgment call.
Analyse
Pattern-matching across data and prior usage is faster and more consistent for AI.
Recommend
A suggestion still needs a human decision after it, so AI proposing it carries low risk.
Decide
Committing to a path is a judgment call with consequences a human should own.
Explain
Surfacing the reasoning behind a recommendation is a language task, well suited to AI.
Guide
Walking a user through next steps is repeatable and doesn't require new judgment each time.
Prepare Request
Assembling the request from what's already been established is mechanical, not a new decision.
Validate
AI checks the request against policy; a human confirms anything the policy can't resolve alone.
Submit Request
Submission is the moment intent becomes action — a human commits to it explicitly.
Approve
Approval is an accountability step; it stays with a named human, not the model.
Provision
Executing an already-approved change is mechanical work, not a new decision.
Monitor
Watching for state changes over time is exactly what AI is well suited to sustain.
Detect Failure
Recognising a deviation from expected state is a pattern-detection task.
Escalate
AI raises the alert; a human decides what happens next when something's gone wrong.
STEP 3 · THE NEW AGE INFORMATION ARCHITECTURE

How does the work actually move between human, AI and the platform services?

Enterprise Data & AI Platform | Self Service Portal

Experience information model

1. Intent layer

What the user is trying to do

What do you want to do today?
Find an enablement serviceGet access to dataSet up a Data & AI workspaceTrack / manage requestCheck platform metricsGet help

2. Entity layer

The platform as connected entities

Platform Core

  • Data & AI Workspaces
  • Compute Scaling
  • Storage Integration
  • Pipeline Monitor
  • AI Compute Limit Increase
  • Cost Intelligence

Data Fabric

  • Federated Teams
  • Technical Catalog
  • Access Intelligence Hub
  • Data Quality Hub
  • Repository

My Requests

  • Track / Manage request

Release Notes

  • Platform Core
  • Data Fabric

Home

  • What is the platform?
  • Platform Metrics
  • Adoption · Usage Overview
  • Platform Health · Key Insights

Help & Support

  • Contact Support
  • FAQ
  • Provide Feedback
  • About the AI Analytics Platform

Relationship layer

How the systems work together

To be mapped with the platform team — dependencies between services, what gates a request, ownership and handoffs across Platform Core, Data Fabric and My Requests.

3. State and action layer

States belong to an entity type

My Requests — lifecycle

SubmittedIn reviewApprovedRejectedProvisionedClosed

Enablement Services — lifecycle

AvailableAvailable on requestDeprecated

Available actions

DiscoverCompareRaise requestTrack / manageMonitorProvide FeedbackContact Support

4. Knowledge layer

What answers are grounded in

Technical Catalog

Service and asset detail

Release Notes

What changed and when

Platform Metrics

Adoption, health, insights

FAQ

Answered questions

About the AI Analytics Platform

Scope and boundaries

AI layer

Two surfaces, one model

Platform AI (Global Assistant)

Traverses the whole graph. Answers across Platform Core, Data Fabric and My Requests.

Service-Specific AI (Embedded AI)

Scoped to one entity and its edges. Same model, narrower traversal.

Terminology unchanged from the base information architecture.

STEP 4 · TABLE C

What boundaries, controls and recovery are needed?

ACTION
WHAT BREAKS?
REVERSIBLE?
AUTONOMY
BOUNDARY
HUMAN CONTROL
DETECTION
RECOVERY
Delegate
User delegates too much or to the wrong capability
Yes
Non-controlling
Ask only for what's delegatable
User defines scope and can withdraw it
Shown when the delegation was made
Modify or revoke delegation
Understand
AI misunderstands intent or context
Yes
Assistive
Must ask for high-impact ambiguity without confirmation
User can correct or clarify
Shown interpreted intent
Reinterpret / ask clarification
Discover
Wrong, incomplete or irrelevant services surfaced
Yes
Assistive
Search only, tracked platform services and eligible scope
User chooses what to explore
Shown sources, eligibility and relevance
Refine search / broaden discovery
Analyse
Misses context, data or important trade-offs
Yes
Assistive
Any plain matching bounded to validated evidence
User reviews data and conclusions
Shown evidence, assumptions and limitations
Re-analyse / provide additional context
Recommend
AI recommends an unsuitable service or configuration
Yes
Recommend
Recommends sourced input and eligibility or overreach
User accepts, rejects or changes recommendation
Explains rationale and supporting evidence
Choose alternative / reassess
Decide
User makes a decision based on incorrect or incomplete information
Yes
Non-acting
AI cannot decide; it only supports the decision
User has final decision authority
Surfaces assumptions, trade-offs and uncertainty
Change decision before execution
Explain
AI provides inaccurate, incomplete or misleading explanation
Yes
Assistive
Explanation must be grounded in auditable sources/precedent
User can request clarification or challenge explanation
Source visibility + confidence scoring
Correct explanation / escalate
Guide
AI gives incorrect or leads user down a wrong path
Yes
Assistive
Guidance stays inside pre-approved patterns, not ad-hoc advice
User chooses whether to follow guidance
Compare + source cited
Backtrack / provide alternative path
Prepare Request
Missing, incorrect or inappropriate request details
Yes
Assistive
Assembly follows a fixed request schema, no free-form additions
User reviews and edits before submission
Highlight non-standard / filled fields
Edit request before submission
Validate
False pass (accepts invalid) or false fail (blocks valid request)
Yes
Automated within limits
Validate against auto-generated and known constraints
User can override false-fail conditions
Show fail reasons and reasons
Correct inputs / request exception
Submit Request
Wrong, duplicate or unintended request submitted
Partially
Human-approval
Cannot submit without confirmation of intent
User confirms and can amend before submission
Flow before + confirmation trail
Cancel / withdraw before execution
Approve
Incorrect or unauthorised approval
Partially
Human-led
Approval remains with authorised human approver
Human explicitly approves
Approver record + audit trail
Revoke / escalate where permitted
Provision
Wrong resource, configuration, access or security exposure
Partially / no
System-controlled
Executes only approved, validated configurations within limits
Real-time provisioning status + logs exposed
Rollback / disable / escalate
Monitor
Important state changes or issues missed
Yes
Automated
Error or anomaly must exceed threshold to raise alert
User can adjust monitoring preferences and thresholds
Continuous monitoring + alerts
Investigate / intervene
Detect Failure
Failure is missed or false alarm generated
Yes
Automated
Detection must be based on defined system behaviour, not guesses
User is notified with severity and impact assessed
Logs, thresholds, anomaly signals
Diagnose / retry / escalate
Escalate
Issue routed to wrong team or escalation happens too late
Yes
Allocated
Escalate based on defined ownership and severity rules
User can request or override escalation
Escalation status and outcome
Re-route / return to self-service
03

Design

The dimensions of the interaction — each one a question the interface has to answer.

Medium

What is the right fit for this experience? UI, Conversational, Embedded AI, Multimodal, Agents, invisible…

Delegation

How intent gets in, and how precisely it can be expressed. Conversational, structured, or both.

Transparency

What it's doing, why, and how sure it is. Uncertainty and explanation live here.

Visibility

Can the user check the result faster than doing it themselves? If not, the automation is decorative.

Control

Interrupt, redirect mid-run, override, undo, recover.

Uncertainty

What happens when AI isn't sure? Ask, flag, or defer to a human.

Recovery

What happens when something goes wrong?

Feedback

How does AI communicate progress, status and outcome?

04

Prototype

Two prototypes, because the interface and the model fail in different ways.

Experience prototype

The interface design itself: screens, states, and the moments that steer a live interface.

Behaviour prototype

How does the actual model actually act?

Co-creating/collaboration with User and Dev Stakeholders.

05

Evaluate

Checks that a question can actually answer with evidence.
DIMENSION
EVALUATION QUESTION
Human outcome
Did the user accomplish what they wanted?
Experience
Was it useful, understandable and efficient?
AI behaviour
Did AI behave reliably?
Verifiability
Could the user determine whether the result was correct?
Trust calibration
Did user trust match AI's actual reliability?
Human control
Could the user intervene, correct, stop or recover?
Adoption
Did users actually incorporate it into their work?
↻ LEARN / ITERATE
Findings feed back into Define: the loop restarts with a sharper read of the human, context and intent.