Build trusted data. Then let AI act on it.
Most analytics work stalls because nobody agrees what the number means. INFOC builds governed data products and semantic models on Microsoft Fabric first, then puts agents on top of definitions the business has already accepted.
DATA-TO-ACTION ARCHITECTURE
Six layers that must work together
AI reliability is decided beneath the model, in the six layers under it.
Business Central, Dynamics 365, Microsoft 365, files, applications, devices and external data.
Fabric Data Factory, notebooks, pipelines, event streams and data-engineering practices.
OneLake, lakehouse, warehouse, data quality, lineage, security and product ownership.
Semantic models, Power BI, metrics, planning, forecasting and self-service governance.
Copilot Studio, Microsoft Foundry, models, knowledge, tools, workflows and agent applications.
Tracing, quality evaluation, red teaming, security, latency, cost, human review and incident response.
MICROSOFT FABRIC
Governed, reusable data products
Built once, then used by BI, integration and AI alike.
DESIGN
Data strategy & architecture
Prioritise decisions, domains, ownership, sources, quality, security and operating responsibilities.
BUILD
Data engineering
Ingestion, transformation, orchestration, testing and monitoring across structured and unstructured sources.
SERVE
Lakehouse & warehouse
Select storage and serving patterns based on workload, governance, performance and consumption.
ACT
Real-time intelligence
Event and streaming patterns where the business decision cannot wait for batch processing.
Platform architecture
POWER BI & SEMANTIC MODELS
One definition of performance
One controlled definition of every measure, with drill-through intact.
Management intelligence
- Finance, sales, inventory, purchasing, warehouse, manufacturing, project and service measures
- Certified semantic models and reusable calculations
- Row-level security and audience-specific experiences
- Data-quality indicators and reconciliation to source systems
- Deployment pipelines, testing, monitoring and ownership
- Copilot experiences grounded in governed models and definitions
ENTERPRISE AI & AGENTS
Choose the right build surface
Different scenarios need different build surfaces.
Use work context, Microsoft Graph and approved knowledge for productivity and collaboration scenarios.
Low-code agents, topics, actions, connectors, channels and business-process automation with governance.
Build, deploy and operate custom agent applications with model choice, tools, tracing, evaluation and monitoring.
Modern AI architecture must define what the agent can know, what it can do, whose identity it uses, when a human must approve, and how failures are detected and contained.
LICENSING AND CAPACITY
The licence model decides who is allowed to read the report
Below F64 every viewer needs a paid licence. At F64 and above they do not.
PER USER
Power BI Pro
Publish and share in the service. Every author and every viewer needs one. Simple, predictable, and the right answer while the audience is small.
PER USER
Premium Per User
Adds larger models, more frequent refresh and advanced features for the people who build. Often used for a small author group alongside Pro for the rest.
CAPACITY
Fabric F SKU
Compute bought by capacity unit rather than by head. At F64 and above, anyone in the tenant with a free licence and a viewer role can read published content.
Fabric capacity is shared across every workload, rather than Power BI alone. Organisations moving from a Power BI Premium P SKU routinely size their F SKU against historical Power BI usage and undersize, because pipelines, warehouse queries, notebooks and Copilot inference now draw on the same pool. Instrument real consumption before reserving, and size to sustained load with bursting for peaks rather than to peak.
OneLake storage is billed separately from compute. Reserved capacity is materially cheaper than pay as you go for steady production load, while a small pay-as-you-go capacity for development can be paused when idle. If you still hold a P SKU, check the renewal date now rather than at renewal: workspace reassignment and testing take real time in a large tenant.
MODEL GOVERNANCE
Six rules that stop a reporting estate becoming forty versions of revenue
Governance is cheaper to install than to retrofit.
Every semantic model has a business owner who approves definition changes, rather than only a developer who deploys them.
Promoted and certified are applied under a written standard. If everything is certified, nothing is.
Access rules live in the model, tested with the roles they protect, not bolted on per report after someone sees the wrong region.
Development, test and production are separate workspaces with deployment pipelines. Publishing straight to production is how definitions drift.
Source to lakehouse to model to report, traceable, so the answer to "where did this number come from" takes minutes.
Reports are deprecated on a schedule with usage evidence. An estate that only grows is an estate nobody trusts.
WHERE IT PAYS
Four applications that survive contact with production
Each one has a named owner and a measurable before.
FINANCE
Close and variance
Ratio analysis across activity, profitability, liquidity and leverage, refreshed against the ledger rather than rebuilt each month. Variance explained while it can still be acted on.
SALES
Customer segmentation
Recency, frequency and monetary scoring applied to the whole book, so retention effort goes to accounts that are slipping rather than accounts that are loud.
SUPPLY CHAIN
Stock position and reorder
Under-stock, over-stock and ageing by SKU, with reorder signals derived from actual movement rather than a static minimum set three years ago.
SERVICE AND OPERATIONS
Grounded question answering
Natural-language access to approved definitions and documents, scoped to the user's own permissions, with the source cited so the answer can be checked.
STARTING POINT
Where to start, and how to trust the output
Do we need Fabric to start?
Not always. If the question is "one dashboard for the leadership team on data that already lives in one ERP", Power BI on a governed semantic model answers it without Fabric. Fabric earns its place when you have several sources to reconcile, volumes that make refresh windows painful, or a need to share compute across engineering and reporting. We say which case you are in before you buy capacity.
How is AI kept trustworthy?
Every agent is grounded in approved data, runs under the requesting user's permissions, and cites its source so the answer can be checked. Tools it may call are enumerated, not open-ended. Evaluation runs before production and continues after, because a model that was accurate at launch is not automatically accurate six months later.
Can this use our Business Central data?
Yes, and it is the most common starting point. Business Central exposes data through APIs and a direct Fabric link, so the ledger, the item master and the sales history land in OneLake without a bespoke extract. The work is in the semantic layer above it, where the business agrees what margin and on-time delivery actually mean.
Who should own the semantic model, IT or finance?
Both, in different roles. Finance owns the definition: what counts as revenue, which entities consolidate, how a return is treated. IT owns the implementation, the refresh, the security and the release. Projects fail when IT is asked to invent definitions, or when finance is asked to maintain DAX.
We already have Power BI and nobody trusts it. Where do we start?
With an inventory, not a rebuild. Usage telemetry usually shows a small number of reports carrying almost all the traffic, and a long tail nobody opens. Certify and fix the few that matter, retire the tail on a schedule, and put the survivors on shared definitions. Rebuilding everything is slower and reproduces the same problem.
How do you handle data quality problems in the source system?
We surface them rather than clean them downstream. Cleaning in the reporting layer hides the defect and guarantees the two systems will disagree forever. Where the source cannot be fixed quickly, the rule is documented and the affected measure is flagged in the model, so the business knows what it is looking at.
What happens if Microsoft changes the platform under us?
It will. Power BI Premium P SKUs moved to Fabric F SKUs, and that pattern repeats. We design for it: capacity choices reviewed at renewal, deployment pipelines so a platform change does not mean hand-editing production, and no dependency on a preview feature in a business-critical path without a stated fallback.
NEXT STEP
Start with the decision the data has to support
Tell us the decision that is currently made on instinct. We work back to the data it needs.