AI is altering how we work, quick. With the rise of enormous language fashions and agentic instruments, anybody can now immediate their approach to solutions that when required an engineer, an analyst, or a BI group. This shift, usually known as “vibe coding” (a time period coined by Andrej Karpathy in early 2025) is actual, highly effective, and right here to remain.
However there’s a catch.
Vibe coding works nice when AI has the precise context. When it doesn’t, you don’t get an error message, you get a assured, improper reply. And in complicated enterprise knowledge environments, that’s a major problem.
On this article we share a concrete instance of what goes improper when AI brokers question uncooked programs with out grounding context. We then present how licensed knowledge, AI Expertise, and Tableau MCP repair the issue.
The Drawback: AI With out Context Produces Inconsistent Outcomes
After we ask an AI agent to tug a metric, the agent doesn’t simply retrieve knowledge – it generates a probabilistic response. Because of this even with the identical query, totally different customers can get totally different solutions, relying on which knowledge sources the agent finds, which logic it applies to retrieve knowledge, and what prior context it has that guides its selections and actions.
In enterprise environments, this downside is amplified. Years of collected knowledge, various levels of knowledge high quality, various enterprise definitions, and inconsistent knowledge area logic give AI brokers loads of room to go improper with no built-in approach to know they did.
Instance: What Goes Mistaken — The MQL Reporting Drawback
After Q1 FY27 closed, our Advertising and marketing group was getting ready for a quarterly pipeline evaluation. A easy query got here up: “What number of advertising certified leads (MQLs) did we generate final quarter?”
Entrepreneurs requested Slackbot this similar query two totally different instances. Right here’s what occurred.
| Description | Slackbot Outcomes 1 | Slackbot Outcomes 2 | Leads Tableau Dashboard |
|---|---|---|---|
| MQLs | 378,591 | 291,256 | 492,298 |
| Knowledge Supply | Org62 Lead Object | Org62 Lead Object | Licensed dataset in Advertising and marketing Knowledge Warehouse (MDW) |
| ‘Behind-the-Scenes’ Logic | MQL Marketing campaign ID is just not null
|
MQL Date in Q1 FY27 |
Snap Sort = ‘Quarterly’
|
Entrepreneurs acquired two totally different solutions. Identical query. Identical underlying system (Org62 is Salesforce’s inside occasion of Gross sales Cloud). Neither Slackbot reply matched the worth from the authoritative supply of fact: the Leads Tableau Dashboard.
Why have been the Slackbot outcomes improper?
Each Slackbot queries discovered the Org62 Lead object, however neither used the licensed MQL definition. The important thing points:
MQL Marketing campaign IDandMQL Dateusually are not aligned with the official MQL definition — they’re associated fields, not the right ones.- Org62 gives real-time knowledge, that means the Q1 FY27 quantity can shift retroactively as leads change standing as time progresses.
- A lead can transfer out and in of MQL standing, making Lead object queries inherently unreliable for period-over-period comparisons.
Why will we belief Leads Tableau Dashboard?
The Leads Dashboard is the authoritative supply for MQL statistics. The dashboard pulls from the Advertising and marketing Knowledge Warehouse (MDW) licensed dataset, which incorporates a quarter-end snapshot with specific enterprise logic that captures the official enterprise definition of MQL:
Snap Sort="Quarterly"ensures point-in-time accuracy.Funnel Stage Title="leads_qualified"makes use of the ruled MQL definition.Nonlinear Funnel FYQ = 'FY2027 Q1'anchors the consequence to the right fiscal interval.
This question pulls numbers from an information layer constructed particularly for correct, constant reporting.
| Issue | Clarification |
|---|---|
| Non-deterministic AI conduct | Brokers generate probabilistic responses, so inherent variability exists even with an identical prompts |
| Transient lead knowledge | Leads can change MQL standing inside or throughout quarters; real-time knowledge doesn’t mirror period-end actuality |
| Totally different historic context | Prior Slackbot conversations and Slack content material affect LLM outputs in a different way for every consumer |
| Various immediate interpretation | Slight variations in phrasing may result in totally different logic being utilized |
| Totally different LLMs / Expertise | Superior vs. fundamental fashions and which Expertise are lively produce totally different outcomes |
The Resolution: A Verified AI Talent Grounded in Licensed Knowledge with Semantics
The repair is to floor the Slackbot agent correctly. Right here’s the three-part method Salesforce makes use of to make sure Slackbot returns outcomes that match the authoritative Leads Dashboard each time.
Step 1. AI Talent: Give the Agent the Proper Steerage
An AI Talent is a structured set of directions (i.e., a Markdown file) that tells Slackbot precisely how one can reply particular questions. For lead funnel metrics, the Leads_Skill defines:
- Which knowledge supply to make use of – e.g., the MDW licensed desk, not the uncooked Lead object
- Which filters to use – e.g.,
Snap Sort,Funnel Stage Title,Nonlinear Funnel FYQ - What enterprise guidelines govern MQL classification
With out this, Slackbot improvises. With it, Slackbot follows a ruled, repeatable path.
Step 2. Tableau MCP: Semantic Fashions for Machine-Readable Enterprise Logic
A Tableau MCP server connects Slackbot on to Tableau dashboards and, extra importantly, to the semantic fashions underlying them. The semantic fashions translate uncooked knowledge into ruled, machine-readable enterprise logic, metric definitions, filter guidelines, relationship maps, and many others.
When Slackbot queries via the MCP, it’s not guessing logic – it’s studying it from a licensed supply. Customers can ask questions in plain language and get outcomes which might be constant, correct, and aligned with what the Tableau dashboard exhibits. The info is additional enriched via integration with our Knowledge 360 occasion and Information Graph, remodeling remoted knowledge silos into an interconnected, discoverable info community that any enterprise consumer can navigate via Slackbot, without having technical experience.
Step 3. Licensed Dataset: The Single Supply of Reality
The AI Talent grounds Slackbot within the MDW licensed dataset – the identical basis that powers Salesforce’s Leads Dashboard. This dataset is developed, validated, and ruled by the Chief Knowledge Workplace, in shut collaboration with Salesforce Advertising and marketing.
The consequence: When a Slackbot consumer asks about Q1 FY27 MQLs utilizing the Leads_Skill and Tableau MCP, they get 492,298, precisely matching Leads Tableau Dashboard. Each time.
Knowledge-as-a-Product: The Philosophy Behind the Resolution
The method described on this article isn’t a one-time resolution – it’s a strategic framework. Knowledge certification and knowledge democratization are core tenets of Salesforce’s Chief Knowledge Workplace. Licensed knowledge establishes the usual layer for enterprise reporting. AI Expertise guarantee dependable outputs throughout platforms. MCP connections and Information Graphs make knowledge and insights accessible to everybody, not simply analysts who know SQL.
Knowledge-as-a-product is the concept that knowledge needs to be as discoverable, usable, and dependable as some other product. When AI brokers are constructed on this basis, vibe coding turns into a characteristic, not a danger.
The Outcomes: Enterprise Affect
Grounding AI brokers in licensed knowledge, AI abilities, and Tableau MCP doesn’t simply repair the accuracy downside; it basically adjustments how groups work together with knowledge at scale. Right here’s what that shift appears like in follow:
- Elevated knowledge democratization with out sacrificing accuracy. Enterprise customers throughout Advertising and marketing, Gross sales, and Operations can now self-serve complicated metrics via pure language queries, increasing knowledge entry from a handful of SQL-proficient analysts to lots of of stakeholders who want insights to drive selections.
- Enhanced cross-functional alignment on key metrics. When Advertising and marketing, Gross sales, and Finance all reference the identical MQL definition via AI-powered queries, pipeline development discussions turn out to be extra productive, main to raised useful resource allocation selections.
- Improved audit path and compliance readiness. Each AI-generated metric now traces again to licensed datasets with documented enterprise logic and metadata, creating an automated audit path that satisfies each inside governance necessities and exterior compliance frameworks.
- Scaled institutional information past particular person consultants. Enterprise definitions and analytical approaches that beforehand lived within the heads of senior analysts are actually codified in AI Expertise, making organizational information resilient to group adjustments and accessible to new hires from day one.
Collectively, these outcomes anchor each stakeholder on a constant, auditable definition of MQL, guaranteeing that when AI provides you a solution, it’s one you may act on.
Conclusion
Vibe coding unlocks productiveness and pace. The flexibility to ask a query and get a solution with out writing SQL, constructing a dashboard, or submitting a ticket is a real leap ahead for enterprises. However this productiveness is barely priceless if the solutions are proper.
The lesson from our MQL instance is simple: AI brokers are solely as dependable as the info and context they’re grounded in. Uncooked knowledge programs, inconsistent logic, and the probabilistic nature of agent conduct produce assured, improper solutions. In enterprise analytics this has actual penalties.
The answer is equally easy: licensed datasets, ruled semantic fashions, and AI Expertise that information brokers to generate the right solutions.
To harness the facility of agentic AI with out sacrificing knowledge integrity, organizations should:
- Certify the info – set up a ruled dataset as the only supply of fact
- Construct AI Expertise – present context-rich directions for brokers to comply with
- Join semantic layers — use Tableau MCP and Information Graphs to make enterprise guidelines machine-readable
The long run belongs to organizations that mix the agility of AI-driven improvement with the self-discipline of knowledge governance. Grounded in licensed, contextual frameworks, vibe coding turns into an asset as a substitute of a legal responsibility.
