We had been at all times the technical outliers in a sea of polished dashboard firms.

For years, BI was judged on what occurred on the display screen: the vary of visualizations, the richness of the information mannequin, how simply an finish person might construct a dashboard. Structure barely entered the analysis. Neither did APIs, analytics as code, or open requirements.

Key Takeaways

• AI brokers generate a special sort of site visitors than folks: one activity can flip into tens or a whole bunch of short-lived queries, and some hundred brokers flip occasional human demand into steady machine demand.
• Conventional BI is squeezed between AI above and warehouses under. What stays worthwhile within the center is serving analytics below actual working circumstances: constant definitions, enforced boundaries, and predictable latency and value.
• Warehouse-native semantic layers outline which means solely inside their very own platform. Enterprises operating a number of engines and functions want a impartial layer that retains metrics constant all over the place.
• An agent’s authority ought to often be narrower than that of the one that requested, and the system ought to refuse to reply somewhat than produce a believable quantity when a query falls outdoors its boundaries.
• The Agentic Serving Airplane is the structure of GoodData.AI. It extends capabilities already in manufacturing, reminiscent of dimension-independent metrics, FlexQuery, multitenancy, and column-level permissions, somewhat than changing them.

At GoodData, our staff stored getting pulled into the engineering beneath. Scale. Efficiency. Safety. Serving analytics by way of APIs. Managing definitions and deployments as code. Constructing on open requirements.

The market noticed the dashboard. However the dashboard was the interface, not the platform. Beneath it we constructed a ruled system for serving information to functions: semantics and licensed metrics, question planning and caching, permissions and tenant isolation. With AI, these cease being engineering particulars. They turn out to be the product.

Analytics inside another person’s product

A lot of that engineering got here from one sort of buyer: firms serving analytics inside their very own merchandise. A lot of their prospects share the identical platform, every with totally different information and permissions. Requests arrive collectively. Response instances turn out to be a part of the product’s person expertise. Question prices turn out to be a part of its economics.

A buyer’s information boundary has to carry on each request. A calculation has to maintain its which means throughout each evaluation. The system has to remain usable because the workload grows from one tenant to hundreds.

These had been the issues we selected.

What that appeared like in apply

A metric ought to imply the identical factor whether or not you take a look at it by area, product, or quarter. In 2013, our Extensible Analytics Engine made that potential: a metric outlined as soon as could possibly be reused as the scale of an evaluation modified, with out rewriting it for every report. That separated enterprise which means from no matter view somebody needed to construct. GoodData’s Extensible Analytics Engine

Quick solutions should not come at an unpredictable value. We constructed FlexQuery on Arrow, Iceberg, and DuckDB so it might learn a buyer’s lakehouse in place, with out one other proprietary copy, and serve heavy site visitors with out routing it by way of the shopper’s manufacturing compute. This yr we added validity intervals to caching: caches now expire on a schedule as an alternative of ready for an API name after each ETL run, and recomputation spreads out over time as an alternative of spiking abruptly. GoodData FlexQuery

Tenants share a platform however by no means one another’s information. Our multitenant structure enforces that boundary on each request. AI Lake, our managed lakehouse on open Iceberg storage, was designed for the bursty, simultaneous queries that embedded functions, dashboards, and brokers generate, with storage and compute scaling independently. GoodData Analytics Lake

Most not too long ago, column-level permissions moved entry management all the way down to particular person fields within the semantic layer. An AI agent working in a workspace operates below the identical guidelines because the individual in that seat, the identical approach row-level safety already works.

Squeezed between two layers

The stack is altering round us. Above, AI more and more handles the interplay with information: decoding a query, investigating a solution, deciding what to have a look at subsequent. Under, Snowflake, Databricks, and the engines constructed on open desk codecs like Iceberg maintain absorbing extra of the compute.

Conventional BI sits between them. When the agent attracts its personal chart and the warehouse runs its personal question, a BI product that was largely a UI has nothing left to supply.

What stays worthwhile within the center is serving analytics below actual working circumstances.

Why not go straight to the warehouse?

As a result of agent site visitors has a special form. A single activity can flip into tens or a whole bunch of small, short-lived queries. A couple of hundred brokers flip occasional human site visitors into steady machine demand. Pointing that straight at manufacturing analytical compute is an availability resolution earlier than it’s a value resolution.

Warehouses now ship semantic layers of their very own, and that is good for everybody. However every defines which means inside its personal partitions. Most enterprises run a couple of engine and a couple of utility, and plenty of serve hundreds of shoppers of their very own. One thing impartial has to maintain a metric constant throughout all of them, isolate one workload from one other, and implement who may even see what, wherever the information lives. A single-platform vendor is the mistaken social gathering to play that position.

An agent investigating income

Contemplate an agent requested why income is declining. It compares areas, examines product traces, follows up on specific accounts. One activity turns into a series of requests, every consequence deciding the following.

All through that chain, income has to imply the identical factor whether or not the agent slices by area, product, or account. That’s the drawback our engine solved once we made metrics impartial of the scale they’re seen by way of. Each request has to remain contained in the requester’s boundary, the identical one our multitenant structure enforces for each tenant. And the entire chain has to complete inside limits on time and value, which is what FlexQuery is for.

The agent’s authority ought to often be narrower than that of the one that requested. It could learn the ruled regional income metric however not the shopper rows behind it. It could begin a reforecast however not change a value. And when a query cannot be answered contained in the semantic, coverage, or information boundary, the system ought to say so as an alternative of manufacturing a believable quantity.

Somebody reviewing the reply must see which definitions, sources, and filters produced it. If a definition adjustments, the staff wants to check the implications earlier than they attain reviews, functions, and brokers.

An individual opens a dashboard and picks a filter. An agent retains asking till the duty is finished.

What we’re constructing now

The Agentic Serving Airplane is the structure of GoodData.AI. It extends what we already run in manufacturing; it would not substitute it. Every a part of it begins from one thing our prospects already use.

Information stays the place it’s. Iceberg lets us work on information in place, and we’re including Trino as a local information supply so prospects can question throughout the techniques it federates, with person identification handed by way of so the supply’s personal safety nonetheless applies.

That means is outlined as soon as. Date parameters and semantic switchers let one ruled definition serve many views, which retains the catalog small for folks and unambiguous for brokers. A semantic high quality agent already flags definitions that confuse AI, reminiscent of duplicate or unclear titles. Subsequent, it should suggest the repair and apply it on approval.

Authority follows the article. After column-level permissions, object-level permissions lengthen the identical management to metrics, visualizations, and dashboards. Brokers inherit the identical per-object guidelines as human analysts, enforced within the semantic layer.

Inference runs near the information. Native inference runs core AI options on GoodData-hosted GPUs, with no third-party mannequin supplier within the information path. Subsequent, every activity is routed to a mannequin sized for it, so prospects cease paying frontier-model costs for routine work.

Brokers can name it from anyplace. Analysts more and more work in Claude, Cursor, and ChatGPT. By MCP, ruled metrics and dashboards seem inside these instruments with tenant scoping and entry management intact.

The system watches itself. Utilization analytics already exhibits how AI is used throughout a company, together with failures and timeouts. Subsequent, an observability agent watches the brokers prospects construct on GoodData and names the precise change that fixes a failing sample.

The place we’re going

Dashboards, embedded functions, assistants, copilots, and autonomous brokers all run on the identical basis: reusable definitions, enforced entry, inspectable outcomes, and predictable execution. Analytics is not a legacy enterprise we’re defending. It is the primary mature approach folks consumed this basis, and years of proof that it holds up in manufacturing.

It continues the work that has occupied our staff for years. The following query could come from an agent as an alternative of a dashboard. It nonetheless has to make use of the fitting definition, respect the fitting boundaries, and return on time.