BI tooling has been converging on a few durable shifts rather than reinventing itself every year: the semantic layer moving out of the BI tool and into the data warehouse, AI-assisted query generation maturing past novelty, and embedded analytics becoming a default expectation rather than a differentiator. None of this is speculative — each trend is already visible in how vendors like Looker, Power BI, Tableau, and the modern metrics-layer tools (dbt's semantic layer, Cube) are positioning themselves.
The semantic layer moves into the warehouse
Historically, business logic — what counts as "active customer," how revenue is calculated, which currency conversion applies — lived inside the BI tool itself, duplicated across every dashboard tool a company used. The trend is toward defining these metrics once in a semantic layer that sits close to the warehouse (dbt's metrics/semantic layer, Cube, or a warehouse-native feature) and having every downstream tool — dashboards, notebooks, AI assistants — query that single definition. This kills the recurring problem of two dashboards showing different revenue numbers because someone redefined "active" slightly differently in each tool.
# semantic_models/revenue.yml (dbt semantic layer style)
semantic_models:
- name: subscriptions
model: ref('fct_subscriptions')
measures:
- name: net_new_arr
agg: sum
expr: arr_delta
dimensions:
- name: signup_month
type: time
type_params:
time_granularity: month
metrics:
- name: net_new_arr_monthly
type: simple
type_params:
measure: net_new_arr
filter: "{{ Dimension('subscriptions__is_active') }} = true"Natural-language query, past the gimmick phase
Ask-a-question-in-plain-English features have existed for years and mostly produced wrong answers on anything beyond a trivial filter. What's changed is that grounding these features in a defined semantic layer — rather than letting an LLM guess table joins from raw warehouse schema — makes the generated queries meaningfully more reliable, because the model is selecting from a small set of pre-validated metrics and dimensions instead of inventing SQL against untyped tables. This is a real trend, not hype, but it depends entirely on the semantic layer existing first; NL-to-SQL against a messy raw warehouse is still unreliable.
If a BI vendor pitches an AI query assistant without a governed metrics layer underneath it, treat the accuracy claims skeptically. The reliability gain comes from constraining the model to defined metrics, not from a bigger or newer LLM.
Embedded analytics as a default expectation
Customers of B2B SaaS products increasingly expect analytics inside the product itself rather than a separate BI tool they have to log into. This has pushed embedded analytics — dashboards and charts rendered inside a host application via SDK or iframe, often multi-tenant and row-level-secured per customer — from a niche feature to close to table stakes for mid-market and enterprise SaaS products. The technical bar has also dropped: most major BI vendors now ship embedding SDKs and row-level security primitives designed specifically for this multi-tenant use case, rather than treating it as an afterthought.
Real-time and streaming dashboards narrowing to specific use cases
"Real-time everything" hasn't materialized as a general BI trend — most business metrics genuinely don't need sub-second freshness, and the infrastructure cost of streaming every dashboard isn't justified by the value. What has stuck is real-time analytics for specific operational use cases — fraud monitoring, live operational dashboards, logistics tracking — where genuine value exists in sub-minute latency, backed by streaming engines rather than batch ETL. The trend is narrower and more targeted than the "everything real-time" framing from a few years ago.
Before adopting a streaming pipeline to feed a dashboard, check whether the actual business decision it informs happens hourly or daily. Most finance and sales dashboards don't need second-level freshness; the added operational complexity of a streaming pipeline is a cost paid for latency nobody asked for.
Governance catching up to self-service
The self-service BI wave of the last decade produced a real cost: dashboard sprawl, inconsistent metric definitions, and no clear ownership. The current trend is tooling that reintroduces governance without killing self-service — certified datasets, metric ownership, usage analytics on dashboards themselves to find and retire the ones nobody looks at. This is less a new technology and more organizations catching up to a mess self-service created.
Wrapping up
The common thread across these shifts is consolidation around a single source of truth for metrics — a semantic layer that AI assistants, embedded dashboards, and traditional BI tools all query consistently, rather than each tool reinventing business logic independently. Adopt the semantic layer first; the AI and embedding capabilities built on top of it are only as trustworthy as the metric definitions underneath.
Independent software engineer in Nairobi specialising in Acumatica customisations, Laravel backends, and tax fiscalisation integrations across East and Southern Africa.