Folks like to say AI adjustments the whole lot. Does that embrace one in every of tech’s most inflexible domains, BI? At GoodData, we predict it already has. Even when the AI hype dies tomorrow, the expectations it units would stay: solutions needs to be quick, contextual, and reliable. That strain alone is forcing BI to shed its previous pores and skin.

For years, BI meant dashboards refreshed in a single day. It was nice for month-to-month evaluations, not so nice for Tuesday at 3:17 p.m. when one thing breaks. You wanted engines that might chew via enormous datasets and run severe SQL (Snowflake or Databricks). That basis remains to be important, however the job description of BI has modified.

Why AI modified the best way we view BI

The vast majority of “AI for BI” demos stay very related and go away you upset moderately than excited. Not as a result of AI isn’t helpful, however as a result of whenever you hand it the steering wheel and hope for the most effective, you primarily create an enormous information on line casino. Slapping a chat field on prime of a pile of knowledge and calling it insightful is sub-optimal at greatest. AI is an interface, not an oracle.

If you wish to perceive the story behind your information, precision issues, and “Is likely to be proper” will not be a method — particularly when a decimal place can swing thousands and thousands of {dollars}. Sending the whole lot to a mannequin and hoping it computes the mathematics accurately is a chance. The mannequin would possibly summarize, hypothesize, and information, however the numbers themselves should come from deterministic, auditable computation.

For that purpose, if you wish to achieve success when integrating AI, it’s a must to be ready for one that will lose any trivia quiz to a magic 8 ball. And no, I don’t say this as a result of I don’t consider in AI, it’s as a result of I don’t wish to depend on AI not hallucinating with my very own information. Why ought to your organization be any completely different?

Though AI will not be the primary wrongdoer for why BI has modified, it positively helped pace issues up. To grasp what this implies, let’s take a look at what has modified.

What has modified?

Whereas there are numerous components of BI which have modified, let’s deal with one use case: creating visualizations. By way of this, we will see 4 pivotal adjustments:

• Reliability
• Simplicity
• Pace
• Accessibility

Reliability

When AI first began making visualizations, folks normally claimed (totally on LinkedIn) that they might now simply speak to their information. By merely giving AI entry to their database and thus having all of the information they wanted at their fingertips.

Whereas this looks as if an important concept, have you ever tried connecting your database to AI? I did, and whereas the primary impressions have been very optimistic, I out of the blue realized that with increasingly more tables the AI began to have issues understanding my information.

Funnily sufficient, this drawback will not be distinctive to AI; even folks can get misplaced in the entire (typically advanced) schema of knowledge. It’s really a widespread drawback throughout the entire market. At GoodData, we deal with this with our semantic layer (Logical Information Mannequin). It’s not solely in regards to the ease of understanding the entire information schema, it’s moderately about making the whole lot less complicated, abstracting pointless particulars, and focusing solely on the that means of the info.

It primarily helps customers and AI navigate the info very similar to a guide from IKEA helps you construct a chair or a cabinet. Positive, you would possibly wish to try to construct it simply primarily based in your instinct, however to be trustworthy, I wouldn’t actually advocate it.

However even with that, AI can wrestle, so the subsequent greatest step is so as to add much more context and create guidelines. Very similar to you’ll create guidelines on your Cursor, you’ll be able to create guidelines with which the AI abides, and with them, it could perceive the language that’s particular to your discipline or firm, for instance. These guidelines usually are not about overlaying flaw;, it is moderately about tweaking the behaviour to your most popular wants. A bit like what ChatGPT does with its reminiscence, which you’ll at all times entry.

Simplicity

When you’ve gotten your information structured, with just a little elbow grease, the AI can lastly perceive your information, however the battle will not be gained. All of a sudden, you realise that whenever you need the AI to create visualizations, it normally wants to make use of SQL to get your information.

And SQL can get very messy, pricey, and in excessive circumstances may even injury your information. I’m not saying that AI would out of the blue drop all of your tables, however SQL injections are very rea,l and creating optimum and proper SQL is a really exhausting process, and debugging will be even worse than writing it by yourself.

One resolution to that is GoodData’s read-only language, MAQL, working on prime of LDM. MAQL itself makes all of the querying secure and easy. No want to fret about SQL dialects for a selected database, as it’s database-agnostic. You possibly can even join any API to it via FlexConnect (the entire idea got here from the identical developer as FlexQuery). And better of all, you’ll be able to even reuse pre-existing metrics to create new metrics, so that you (or AI) can work iteratively and don’t must create the entire logic in a single step.

Pace

With visualizations being correct and easy to audit, one other urgent drawback has emerged. Prior to now few years, the pace at which customers wish to see the already computed information has gone down considerably. It’s partly as a result of AI making it extraordinarily straightforward to create a PoC and get your outcomes quick, however solely typically right. However you’ll be able to’t actually mock the computations, proper?

For this reason now we have our fundamental engine written on prime of Apache Arrow. Whereas it gained’t assist you with the pace at which you fetch the info out of your database (though optimizations of MAQL would possibly), you’ll be able to positively really feel the distinction as soon as it’s loaded.

Apache Arrow is a columnar format with zero-copy learn assist and intensely quick information entry. On prime of this we created a really bold undertaking, which created a framework for constructing information companies powered by the Apache Arrow and Flight RPC – FlexQuery. If you wish to be taught extra about it, I extremely advocate studying the introductory article to the entire structure.

When creating FlexQuery, it wasn’t nearly glueing “a bunch of applied sciences” collectively and hoping for the most effective. Once we created it, it was a really strategic long-term funding, lengthy earlier than AI.

Accessibility

And now that we may have our information crunched reliably and quick we moved to the notion that we will eat our information insights wherever, anytime. It began with wherever my AI can go, my information can comply with, and now there are even experiments with having your day by day digest as a podcast despatched to your mail every morning so you’ll be able to test your information whenever you sip in your morning espresso.

The benefit of entry to your information is behind all the opposite elements I’ve talked about, as a result of not many BI corporations deal with their platforms like a modular engine in opposition to which you’ll base all of your computations. Fortunately GoodData with its api-first method could be very nicely ready to be hooked as much as nearly any frontend or backend. Take OpenAPI specification for instance, you probably have an excellent and descriptive OpenAPI specification, builders can have a a lot simpler time hooking up your product in addition to AI, which positively wants that additional context.

Something that may be executed in GoodData will be executed via APIs and SDKs as nicely. Whereas they don’t seem to be good (nothing is), they’re open-source they usually have a really wholesome improvement. The power of the modular and API-first method can for instance be seen in a number of the articles like Hand Drawn Visualizations, turning your Dashboard right into a scheduled podcast and Hyperpersonalized Analytics.

New AI-Assisted Options

So aside from the PoC, that’s what would keep if AI collapsed tomorrow, however there are additionally many new AI-assisted options that we couldn’t even fathom earlier than AI. From reactionary KDA to Semantic High quality Checker, there are fairly a number of use-cases that will merely be unimaginable with out AI.

AI-assisted KDA

One of many use-cases closest to me is AI-assisted KDA. The premise is straightforward, think about there’s an anomaly someplace in your information. It might probably occur any time, even when you find yourself asleep. And whereas a notification that your information wants consideration is sweet, there’s solely a lot a easy notification can do, particularly at 3AM.

So you’ll be able to let your notifications set off AI-assisted workflows, similar to KDA. Because of this as a substitute of a really sturdy and infrequently costly exhaustive KDA, you’ll be able to make the most of AI that will help you navigate the search house, thus saving loads of time and computational energy. Even with AI it may be in a magnitude of some thousand queries, however most of them will be cached e.g., via FlexQuery.

MCP / A2A

A function that’s completely AI-driven is the utilization of AI-centric protocols to have the ability to connect with brokers and instruments. Whereas the change within the BI is unquestionably not about chasing the subsequent huge protocol which is likely to be out of date in a number of months, there’s positively no hurt in implementing new methods to connect with your product and that is true not just for BI, however merely for any platform that you can imagine.

When you would possibly surprise why you’ll wish to make your platform ready to connect with AI (or vice versa), take into consideration the convenience of use on your consumer. And bear in mind: Giving an AI hammer and nails whereas hoping it won’t hit any thumbs is way more harmful than giving it a sandbox (instrument) the place you’ll be able to assure the correctness of the outcomes.

Semantic High quality Checker

And lastly a function that’s each enabled and enabling for AI is Semantic High quality Checker. It’s really a small miracle that this function is lastly attainable. Information administration can get very messy and the that means of your information can get blurry.

With regards to the cleanliness of knowledge (or moderately the shortage of it), there are three cardinal sins:

• Unexplained Abbreviations – AI won’t perceive your ASDU with out clarification, or was it simply SDU…?
• Duplicit Names throughout completely different tables –
• Lack of Enterprise context – Is your Income internet, gross or recurring…?

And whereas duplicate names are fairly straightforward to catch programmatically, I wouldn’t dare attempt to programmatically remedy the shortage of enterprise context or unexplained abbreviations. That is the place AI really comes into play, as a result of whereas it may not be good (as you would possibly know, AI by no means is..) however you’ll be able to’t construct neither semantic fashions nor Rome in a single go. It’s important to work on it iteratively and slowly enhance the simplicity or moderately the understandability of your semantics.

With higher semantics the AI will even have a greater understanding of the best way you wish to use your information and out of the blue it could choose up extra minute particulars. And with AI on board you’ll be able to have much less evaluate cycles and decrease onboarding time.

Conclusion

AI hasn’t changed BI, nevertheless it definitely raised the bar for it. The winners gained’t be the groups that hand their information to a chat field and hope; they’ll be the groups that pair deterministic, auditable computation with AI because the interface and accelerator. Reliability, simplicity, pace, and accessibility aren’t nice-to-haves anymore; they’re the scaffolding that lets AI be helpful with out turning your numbers right into a on line casino.

That’s why the form of recent BI seems completely different. A semantic layer (LDM) provides people and fashions the identical map. A secure, read-only, metric-centric language (MAQL) retains logic constant and guards the warehouse. A columnar, Arrow-native runtime and FlexQuery transfer outcomes at interactive pace. An API-first floor lets insights present up wherever folks work, be it dashboards, apps, brokers, even a morning “podcast” of your KPIs. On prime of that basis, AI turns into sensible: guiding KDA workflows to slender search house, checking semantic high quality to maintain that means tight, and talking via agent protocols with out punching holes in governance.

If AI hype vanished tomorrow, this stack would nonetheless matter. The expectations it set (quick, contextual, reliable solutions) are actually everlasting. The trail ahead is incremental: harden your semantics, codify metrics, instrument pace, after which let AI assist with the final mile( explanations, navigation, triage) not the mathematics. Deal with AI as an interface, not an oracle, and BI stops being a once-a-month report and turns into a reliable, real-time resolution companion.