A
Context Quality
How good information gets to the AI · T01–T05
T01
Give AI the right information
Listen for
keeps giving wrong answerscan't trust the results
Say
AI gets things wrong when it gets the wrong information. We control exactly what it sees before it answers.
Ask
How do your engineers look up quality issues today — search, a system, a spreadsheet?
When an AI answer is wrong, who catches it — and how fast?
Technical view
On Fabric
Operational records in a Lakehouse Delta table. The agent reads only relevant rows — ~800 tokens for a lookup, ~8,000 for root cause. Sizing that read window correctly is context engineering. Wrong size = wrong answer or wasted cost.
T02
Connect to your data without moving it
Listen for
data is spread across systemswe'd have to build a connector first
Say
The AI reads from wherever your data already lives — the same quality system your engineers use today. No copy needed, no migration project first.
Ask
Where does your operational data sit today — a database, a data lake, something else?
Has anyone tried to build on top of that data before — what got in the way?
Technical view
On Fabric
Fabric Lakehouse + OneLake shortcuts eliminate ETL. Existing Delta tables are queried directly by the agent at call time. No data duplication, no migration project before you can start.
T03
Keep AI answers current
Listen for
AI doesn't know about last weekworking on old data
Say
An AI working from a snapshot of your data doesn't see this week's new incidents. We make sure it reads live records, not a copy from last month.
Ask
How often does your quality data change — real-time, daily, weekly batches?
Has anyone in your team complained that an AI gave outdated information?
Technical view
On Fabric
Retrieval-augmented generation (RAG) over live Delta tables. No model retraining when data changes. Fabric pipelines update the retrieval index on a schedule or on-write trigger. The model stays fixed; the context changes.
T04
Structure your data before AI touches it
Listen for
our data is inconsistentdifferent systems, different formats
Say
If your quality data has inconsistent formats or different naming across systems, the AI will be inconsistent too. We start by making the data AI-ready.
Ask
Do your operational records follow a consistent format, or does it vary by team, tool, or era?
Has anyone mapped where the same data exists in more than one system with different values?
Technical view
On Fabric
Data profiling in Fabric using Spark notebooks. Schema validation before Lakehouse Delta ingestion. Standardising record IDs, entity codes, and severity levels is a prerequisite to reliable retrieval — not a phase two task.
T05
Find by meaning, not just keywords
Listen for
search doesn't find itdifferent teams call it different things
Say
Traditional search finds exact words. AI search finds meaning — two records about the same event come back together even when each team gave it a different name.
Ask
When an engineer searches your quality system, do they use keywords, filters, or free text?
Have you ever missed a failure pattern because different teams described it differently?
Technical view
On Fabric
Vector embeddings of operational records stored in Fabric Lakehouse. Cosine similarity search finds semantically related records regardless of terminology. Embedding model runs on Azure AI Foundry — no data leaves the tenant.
B
Agent Capability
What AI can do with that information · T06–T10
T06
Build agents that work across steps
Listen for
someone still has to stitch it togethercan't automate the whole flow
Say
The AI can look things up, run checks, compare results, and hand off to the right person — all in one workflow, without a human stitching it together.
Ask
What does your best analyst do in the first 10 minutes of investigating an issue cluster?
Which of those steps could an AI realistically do first?
Technical view
On Fabric
Semantic Kernel tool calling. Multi-hop = 2–3 sequential tool invocations (lookup → correlate → recommend). Class 3/4 decision routing. Agent state managed across steps inside Fabric Data Science.
T07
Match the question to the right tool
Listen for
AI is getting expensivehard to justify the spend
Say
A simple lookup shouldn't cost the same as a full investigation. We route each question to the right tool and track exactly what it costs.
Ask
Does your team know what an AI answer costs today — or is it just part of the cloud bill?
Who's responsible for proving AI saves more than it costs?
Technical view
On Fabric
Decision class routing 1–4. Class 1 (simple lookup) → phi-4, ~$0.000003 Azure equivalent. Class 4 (root cause) → gpt-4o, ~$0.000080. That's a ~27× cost swing. DET tracks this per call in OneLake Delta.
T08
Agents that act, not just answer
Listen for
it tells us what to do but we still have to do itAI can't write back to our system
Say
The AI can update a record, open a ticket, or flag for escalation — without a human in the loop for routine decisions.
Ask
Which actions in your quality process are routine enough that a human mostly just clicks confirm?
What would it take for your team to trust an AI to do that?
Technical view
On Fabric
Semantic Kernel + Fabric Activator for event-triggered agent actions. Write-back via Logic Apps. A confirmed outcome flag links the AI action to the closed business result — this is what makes ROI measurable.
T09
Agents that remember what worked
Listen for
knowledge walks out when people leavewe keep making the same mistake
Say
Past decisions that led to a confirmed good outcome get used as examples the next time a similar question comes in. The model stays the same — the context gets smarter.
Ask
Do your best domain experts have pattern recognition that isn't in any procedure or runbook?
Have you ever caught a failure pattern months faster the second time — because someone had seen something similar before?
Technical view
On Fabric
Confirmed outcome flags in DecisionRecord feed back as few-shot examples for similar future queries. High-confidence past decisions are promoted into the context window. No model fine-tuning — context engineering handles the learning.
T10
Multiple agents, each good at one thing
Listen for
one AI can't handle everythingneed different expertise for different areas
Say
Instead of one AI trying to do everything, you can have specialist agents — one for clustering issues, one for root-cause patterns, one for recommending next steps — that hand off to each other.
Ask
Are there different types of expertise in a full investigation — process, data quality, domain analysis?
Would you route those to the same person, or to different specialists?
Technical view
On Fabric
Semantic Kernel ProcessFramework for multi-agent orchestration. Each agent has its own context window, model selection, and tools. Agent handoffs are typed — the issue classifier passes a structured payload to the action recommender. Full audit trail across the chain.
C
Accountability
Measuring and governing what AI does · T11–T15
T11
Know what it costs per outcome
Listen for
shows up in the cloud bill but not the ROIcan't tie AI cost to a business result
Say
We measure what it costs to close one issue cluster, complete one investigation, resolve one escalation — not just the monthly cloud bill.
Ask
How do you measure ROI from your quality systems today?
If AI helped close 10 issue clusters this quarter, what would that be worth?
Technical view
On Fabric
DecisionRecord.azure_equivalent_usd summed per confirmed outcome. Cost per outcome = Σ(token cost × step count) ÷ confirmed outcomes. Showback report from OneLake Delta — queryable by Finance or Quality without touching agent code.
T12
Every AI decision can be explained
Listen for
how do we know why it made that callwho's accountable when it's wrong
Say
Every time the AI makes a recommendation, we keep a full record — what it was asked, which model answered, what it cost, and whether it passed quality checks. Your quality team can query it at any point.
Ask
If the AI flags a case for escalation, who approves that — and what do they see?
Is there someone at your organisation who owns AI accountability or compliance?
Technical view
On Fabric
DecisionRecord written to OneLake Delta per call: decision class, model used, tokens, cost, quality flag, checks failed. Fully queryable in Fabric — no separate logging infrastructure needed.
T13
Set the rules before the AI runs
Listen for
who sets the rules for what it can doneed guardrails before we go live
Say
You decide the guardrails upfront — maximum spend per decision, which answers need human review, which actions need approval. The AI works within those rules automatically.
Ask
Who at your organisation would set the rules for when an AI recommendation needs a human sign-off?
Is there a policy today for how much automation is allowed in a quality decision?
Technical view
On Fabric
DET budget mode: calibrate → enforce. per_decision_usd cap in agent config — calls over budget are blocked before the model is called. AGT (Azure Governance Toolkit) layers policy-as-code audit alongside Semantic Kernel for compliance reporting.
T14
Measure whether answers are actually right
Listen for
how do we know the answer is rightwhat if it makes something up
Say
We check that every answer contains what it should. Every response is tested against a standard before it reaches anyone.
Ask
Right now, how would you know if an AI recommendation was based on the wrong data?
Is there a standard today for what a good investigation report must contain?
Technical view
On Fabric
Two-tier quality probe: structural check (always, free) confirms record IDs, action references, and owner titles are present. Sampled LLM yes/no probe (10% of calls) checks factual accuracy. quality_flag and checks_failed fields written per record to OneLake Delta.
T15
Keep humans in charge of hard decisions
Listen for
that decision always needs a humansome decisions always need a sign-off
Say
The AI handles the routine. Hard decisions — a critical escalation, a final approval — still go to a person. That line is drawn in the system, not just in a process document.
Ask
Which decisions would you never want automated, no matter how good the AI gets?
When something goes wrong with an automated decision today, how is it escalated?
Technical view
On Fabric
Decisions above cost threshold or below quality score are blocked and routed to a human review queue. Fabric Activator triggers an approval workflow for Class 4 decisions. The approval itself is logged in the DecisionRecord — full chain of custody.