When Microsoft’s CEO predicted that apps as we know them will disappear, we at Akvelon treated it as an engineering brief and set out to test what that future could look like inside a real enterprise environment.
Traditional systems had served organizations well, but came with familiar limits: costly updates, rigid frontends, and multi-step releases just to adjust business logic.
We set out to see whether intelligent agents could manage the same operational tasks by reading and writing directly to governed enterprise data — without a traditional app interface.
Our AI team designed an internal experiment to find out, connecting agents to live company data through a governed analytics layer. The goal was to create a system as flexible as conversation, where business users can ask questions and receive secure, validated, actionable answers in real time.
Challenge
To validate the agent-first concept in a controlled environment, we first applied it to our own HR environment, where data access, compliance, and feedback cycles allowed us to iterate quickly.
In Akvelon’s HR internal system, business rule updates or query changes required full release coordination across development, QA, and operations teams. Release cycles could take a significant amount of time.
The root cause wasn’t team inefficiency, but rather architecture: too many layers between business users and data, with business logic buried deep in application code and UI layers. Every adjustment demanded redeployment, versioning, and testing across multiple environments.
To ground the concept in reality, we set out to explore a model where business logic lives inside intelligent agents, capable of performing actions across systems. We focused on three core goals:
- Preserve governance and RBAC.
- Minimize UI and maintenance overhead.
- Allow HR analysts to query and act in natural language.
Solution: Building and Testing Agent-Driven Operations
Akvelon’s AI engineering group replaced static HR reporting modules with an Intelligence Layer built on the Agentic MCP Platform. This layer orchestrates multiple specialized agents that can interpret natural language, plan data operations, and execute validated queries directly against the HR database.
Instead of interacting with dashboards, users interact with the system conversationally. The Intelligence Layer translates their requests into a series of secure, context-aware actions. It:
- Interprets intent using a business-aware LLM.
- Selects and joins relevant tables automatically.
- Returns consistent, reproducible outputs—tables, charts, or forecasts—without developer intervention.
For the first time, non-technical HR managers could generate insights on their own. This shortened feedback loops and allowed decisions to be made in hours instead of days.
The new system introduced several firsts within Akvelon’s internal systems:
- No UI dependency: Queries run entirely through conversational or API interfaces.
- Instant rule updates: Business logic can be modified on the fly through prompts and
validation rules, no redeployments required. - Cross-system awareness: Agents combine structured (tables) and unstructured
(documents, emails) data.
These capabilities laid the groundwork for Akvelon’s broader “From Apps to AI Agents” initiative, demonstrating that large-scale enterprise systems can operate effectively without traditional interfaces.
Project Roadmap
The transformation of the HR platform unfolded through three architectural waves — each one reducing complexity, shortening change cycles, and bringing business users closer to the data.
Wave 1: API-First Pattern
We started from the safer side of the spectrum, allowing the agent to interact only with documented OpenAPI endpoints.
The model learned which operations were available — listing employees, updating records, creating projects — by reading the API specs.
- What worked:
Standard operations like lookups, updates, and reporting succeeded with ~89%
reliability. The agent followed validation rules and RBAC automatically since those lived
in the backend. - Where it struggled:
When an action didn’t have a matching endpoint (e.g., custom “bench report” logic), the
agent couldn’t proceed. Its flexibility was limited to what the API described
Business takeaway:
Wave 2: SQL-First Pattern
Next, we gave the agent direct access to the HR database using Microsoft’s MSSQL Model Context Protocol (MCP) Server, no APIs or middleware.
Now, the agent could run any query across multiple tables, reading and writing data directly.
- What worked:
Unlimited flexibility. The model could explore and execute ad-hoc queries instantly
without waiting for endpoint coverage. - Where it failed:
Without business context, it guessed relationships, produced ambiguous joins, and
made unsafe write attempts. Even “team availability” queries required step-by-step
guidance.
This revealed the opposite side of the spectrum: flexibility without structure is fragile.
Business takeaway:
Neither approach was enough on its own. API-bound agents handled operations correctly but couldn’t adapt to new business logic. SQL-bound agents had full freedom but no understanding of what their actions meant.
The challenge became clear: the real limitation wasn’t access or syntax; the systems lacked the business semantics needed to reason safely about data.
To address that, the next stage focused on defining what the system needed to “understand” before it could act: table relationships, validation rules, and HR-specific terminology. Structured documentation and natural-language meta descriptions were layered into the environment, linking data to real business context.
Wave 3: MCP Intelligence
The final wave introduced an Intelligence Layer — a business-aware orchestration system that connects AI agents directly to the HR database while enforcing governance and validation rules.
Instead of building new frontends or endpoints, Akvelon’s engineers moved their focus to business logic that now lives in structured prompts and rule definitions. This shift allowed the system to evolve dynamically without requiring full redeployments.
The Intelligence Layer translates each user request into a series of secure, context-aware steps, which specialized agents then perform to:
- Interpret user intent using a business-aware LLM
- Map intent to domain-specific rules
- Validate operations before execution
- Adapt over time through stored corrections and learned patterns
This architecture merged the best of both worlds: the governance of API-first design with the flexibility and speed of SQL first experimentation.
Updates that once required a full deployment cycle can now be made instantly by changing a rule or prompt, turning the HR system into a living, conversational interface for business operations.
Business takeaway:
Key Outcomes and Engineering Insights
Each wave of experimentation demonstrated that no single architecture fits every need; the optimal choice depends on the balance between safety, agility, and business meaning.
Traditional applications can require months of coordination, 2-6 months for a new workflow, and weeks for minor updates.
- With Agents + API, the cycle shortened noticeably. Reusing existing endpoints allowed
new business workflows to appear in one to three months, process modifications in
two to four weeks, and fixes in just a few days. - Agents + Direct Database brought even faster iteration: because logic was embedded
in data queries rather than code, new workflows took about one to two months, with
process changes in two to three weeks and fixes in one to three days. - Finally, under MCP Intelligence, the system reached a new threshold of
responsiveness. Thanks to semantic awareness and automatic rule validation, creating a
new workflow now takes only one to two weeks, existing processes can be modified
within a few days, and most operational issues can be addressed within one or two days.
This acceleration redefines enterprise agility, turning what once required months of coordination into near real-time iteration, while maintaining the governance and auditability standards expected of production systems.
Each architecture also revealed different level of build and maintenance complexity. As RBAC moved from frontend to backend to the data layer, the engineering effort dropped from full-stack releases to lightweight business rule updates under MCP Intelligence.
The project introduced several firsts that later shaped Akvelon’s broader ‘From Apps to Agents’ initiative.
- No frontend required: The “app” becomes a conversation with an intelligent system.
- Governed intelligence: Validation and RBAC live where they belong — in trusted services.
- Progressive adoption: API-first offered safety and observability; SQL-first unlocked flexibility and learning.
- Reusable architecture: The same design can be applied to analytics, CRM, or delivery platforms.
Where MCP Intelligence Makes Sense
MCP Intelligence can be applied across different domains, but it’s most effective under specific operational conditions. Through this experiment, Akvelon confirmed that the agent-first model isn’t meant to replace all enterprise systems; it shines in a focused segment where traditional apps are too slow or too rigid. MCP Intelligence fits best when AI agents can synthesize context, apply business logic, and generate cross-system answers faster than any manual process:- Data is semi-structured or scattered across multiple sources (spreadsheets, APIs, emails, PDFs)
- Latency of a few seconds or minutes is acceptable
- Human review is possible before results are finalized
- Accuracy tolerance is moderate, and errors are not business-critical
- Updates or analyses are needed occasionally, not continuously
Each of these use cases applies the same principle validated by the HR system: AI assistants + business intelligence layer + direct database access = faster change, lower cost, and smarter decisions.
However, for daily, high-volume, regulated, or deterministic operations — such as payroll, billing, or transactional systems — traditional architectures remain more efficient and predictable.
Impact and Next Steps
The HR platform initiative was more than an internal upgrade - it became a proof of concept for the agent-first enterprise.
It proved that intelligent agents, when connected directly to enterprise data through a structured orchestration layer, can replace large portions of traditional application logic.
The result wasn’t just a faster system; it was a new way of thinking about software where conversation replaces interfaces and business rules live inside the data environment itself.
Akvelon continues to expand this approach across industries.
Ultimately, the HR platform serves as the blueprint for how organizations can transition from applications to agents, transforming static tools into living systems of intelligence. This approach now forms the foundation of Akvelon’s agentic strategy — a pathway from traditional applications to intelligent systems that continuously learn, adapt, and reason across business domains.