AI readiness assessment
Model-by-model review of what Copilot needs: unambiguous naming, field descriptions, synonyms, complete measure coverage, and no duplicate models competing to answer the same question. Scored, with a remediation backlog.
AI & Copilot
AI, Copilot and advanced analytics
Copilot readiness, custom assistants and forecasting, assessed against a fixed evaluation set.
- Semantic models that answer natural-language questions correctly, with evidence
- Copilot deployed to a defined audience with cost and data boundaries understood
- Forecasts and anomaly alerts measured against a baseline, not just shipped
- An AI governance position you can show a risk committee
Copilot is a multiplier on whatever is underneath it. Pointed at a well-named, well-documented semantic model it is genuinely useful. Pointed at twelve overlapping datasets with cryptic column names it produces confident nonsense, and the organisation concludes AI does not work.
So our AI work usually begins with readiness: naming, descriptions, synonyms, measure coverage, and removing the duplicate models that make a natural-language question ambiguous. Then we deploy the thing, Copilot in Power BI and Fabric, or a custom assistant on Azure OpenAI with retrieval over your own content, with evaluation, logging and cost controls from the start.
We also build the classical statistical work that is often the better answer: demand forecasting, anomaly detection on operational data, propensity and churn models. Less fashionable, frequently more valuable, and far easier to validate.
What we deliver
What the work involves
Copilot enablement
Copilot configured in Power BI and Fabric for a defined audience, with capacity cost implications quantified, data boundaries documented, and user guidance on what it is and is not reliable for.
Custom assistants on Azure OpenAI
Retrieval-augmented assistants over your documents and data, with Azure AI Search, grounded citations, prompt and output logging, evaluation sets, and guardrails against prompt injection and data leakage.
Forecasting and planning
Demand, revenue and capacity forecasting with proper backtesting, confidence intervals, and a published accuracy record against the previous manual process.
Anomaly detection
Operational monitoring on the data itself (unexpected volumes, margin shifts, failed transactions, sensor drift), tuned to a false-positive rate the team will tolerate rather than mute.
AI governance
Use-case register, risk classification, human-in-the-loop requirements, data residency and retention positions, and documentation aligned to emerging obligations including the EU AI Act.
Method
How the engagement runs
Each stage names the artefact it produces, so progress can be checked without a status meeting.
Use-case triage
Candidate use cases scored on value, data readiness, risk and effort. Most organisations have one or two worth doing now and several worth refusing.
Scored use-case register with a recommended first build
Readiness remediation
Model metadata, naming and documentation brought up to the standard the AI layer requires.
Remediated models and a readiness score
Pilot with evaluation
A narrow pilot with a real evaluation set and a defined success threshold agreed in advance, so the go/no-go decision is not a vibe.
Pilot system, evaluation results, go/no-go recommendation
Productionise
Monitoring, cost controls, logging, access boundaries, fallback behaviour and user guidance.
Production deployment with operational controls
Measure and iterate
Accuracy and usage tracked post-launch; the model and prompts revised on evidence.
Performance report and iteration backlog
Deliverables
Deliverables
- Scored AI use-case register with explicit recommendations against some of them
- Copilot readiness report per semantic model
- Remediated model metadata: names, descriptions, synonyms, measure coverage
- Pilot system with documented evaluation set and results
- Forecasting or detection models with backtested accuracy
- Prompt and output logging, with cost monitoring
- AI governance documentation and risk classification
- User guidance on reliable and unreliable uses
Stack
Technologies
- Power BI Copilot
- Microsoft Fabric Copilot
- Azure OpenAI Service
- Azure AI Search
- Azure AI Foundry
- Azure Machine Learning
- Python
- Fabric Data Science
- MLflow
Questions
AI & Copilot: common questions
Almost always because the semantic model does not give it enough to work with. Columns named `amt_2`, no field descriptions, no synonyms for the words your business actually uses, measures missing so it tries to aggregate raw columns, and several models that could each answer the question differently. Copilot is a language interface over metadata; thin metadata produces thin answers. The fix is unglamorous and highly effective.
Not under the Azure OpenAI and Microsoft Fabric commercial terms as they stand: your prompts and data are not used to train the foundation models. That said, the exact boundary, region of processing, retention for abuse monitoring, which features call out to which service, is something we document per deployment so your risk function can review the real configuration rather than a marketing summary.
They solve different problems. Copilot is strong at questions over governed analytics data with no build cost. A custom assistant makes sense when you need retrieval over your own documents, a specific workflow, tight control of prompts and outputs, or integration into an application. If Copilot covers the need, building a bespoke alternative is usually a waste of money.
With an evaluation set built before the pilot: real questions with agreed correct answers, scored for accuracy, groundedness and refusal behaviour, plus a latency and cost budget. The success threshold is agreed up front. If it is not met, we say so and recommend stopping, which has happened, and saved the client the larger bill.
Most analytics and reporting assistants sit in the lower-risk tiers, but the transparency and documentation obligations still apply, and classification depends on use rather than technology. We maintain a use-case register with risk classification, intended purpose, data sources, human oversight and limitations, the documentation set a regulator or customer is most likely to ask for. Legal interpretation stays with your counsel.
Related
Power BISemantic modelling, DAX and report delivery with one agreed definition per measure.Microsoft FabricLakehouse architecture sized against measured consumption rather than headcount.GovernanceOwnership, cataloguing, classification and access evidence, implemented in the tools teams already use.
Sector context: Financial services, Manufacturing, Retail, Healthcare, Logistics.
Next step
AI & Copilot: what the next step looks like
A two-week readiness assessment and a use-case register you can take to a board
Answered within one working day, by the engineer who would do the work