How Deevita built the data foundation for an AI response platform that reached 85.1% accuracy.
Author:
Vadivukkarasi Senthamarai
Sr Engineering Lead
- Many AI projects stall because the data is not ready, not because of the model.
- Deevita built the data foundation for a hospitality AI response platform: unified intake, ordered queuing, live reservation context, validation, and a full audit trail.
- With staff reviewing every response, AI accuracy rose from 75.3% to 85.1% in 8 months, with no PII exposure.
Data readiness for AI is where many projects stall, not the model. The data sits in many systems, comes in different formats, is checked unevenly, and is hard for AI to reach at the moment it needs it.
We saw this first-hand while working with an AI startup building an agentic platform that helps hospitality businesses deliver better customer service using generative AI. The AI grounds each response in the customer’s live reservation details. A staff member then reviews every draft, edits it where needed, and approves it. Only approved responses are delivered through the same channel the request came in on. The first deployment was with a multi-property hospitality operator. We believe the lessons apply to any organization that wants AI working on live operational data.
This case study covers the part that made the platform work: how we brought the data together, transformed it, validated it, and made it available for AI.
Deevita’s role. Deevita designed and built the data foundation (multi-channel intake, ordered queuing, reservation integration, context building, validation, and audit layer) and the React-based review interface. The client’s team built the AI response generation.
The problem: Data that was not ready for AI
Before we built anything, we assessed the client’s data. We found four gaps that are common wherever AI meets live operational data:
- Fragmented sources. Requests came in through four channels: email, WhatsApp, Zendesk, and voice. Each had its own format and metadata.
- No ordering or flow control. With requests arriving on several channels at once, there was no reliable way to make sure they were handled in the order they arrived, especially at peak times.
- Disconnected context. The information needed to answer a request correctly, such as room type, stay dates, and booking status, is set in the reservation system, separate from the request itself.
- No quality gate. Nothing consistently checked a response for accuracy, completeness, or sensitive information before it reached a customer.
Putting AI on top of that foundation would only produce confident mistakes faster. So we started with the data.
Our approach: Four stages to AI-ready data
| Stage 1 ▸ Bring together | Stage 2 ▸ Transform | Stage 3 ▸ Validate | Stage 4 Deliver |
|---|---|---|---|
| Unified intake, FIFO queue, reservation sync | Classify the request and attach reservation context. | Input checks, human edit, and approval | Approved reply sent on the original channel |
The platform runs on Microsoft Azure, using Azure Service Bus, Azure OpenAI, Azure App Service, Microsoft Entra ID, Azure Functions, and Azure Database for PostgreSQL. It is the kind of work our data engineering team does, together with our Azure integration services.
What we built
| Capability | What we built |
|---|---|
| Unified intake | Email, WhatsApp, Zendesk, and voice are normalized into one request record (sender, time, channel, content, and linked reservation). |
| Ordered queuing | Azure Service Bus FIFO queue preserves arrival order across channels, with retry and dead-letter handling. |
| Reservation integration | Webhook or 30-minute sync, with requests matched by reservation identifiers |
| Context building | Request, reservation details, and property reference data assembled into a structured prompt. |
| Review interface | React-based review, edit, and approval workflow |
| Audit lineage | Every request, the data used, the AI draft, reviewer edits, and the approval decision logged |
Stage 1: Bring the data together.
We built one intake layer that takes in requests from every channel and turns them into a single standard request record. Whatever the source (an email, a WhatsApp message, a Zendesk ticket, or a voice call), downstream systems receive the same structure: who sent it, when, through which channel, the content, and the linked reservation.
Every request is then placed on an Azure Service Bus queue and processed in first-in, first-out (FIFO) order. This keeps requests in arrival order across all channels, smooths out peaks in volume, and makes sure no request is lost if a downstream step is busy or temporarily unavailable.
At the same time, we connected the reservation system. Reservation data is synced every 30 minutes, so the AI typically works from booking details that are no more than 30 minutes old.
Stage 2: Transform the data into an AI-usable context.
A raw message is not usable context. We added a transformation layer that:
- cleans and standardizes the text from email, WhatsApp, and Zendesk
- identifies the type of request, for example, room type changes, cleaning and housekeeping, early or late check-in, early or late check-out, extending a stay, nearby tourist attractions, and nearby shops
- matches the request to the customer’s reservation and attaches the relevant details, such as room type, stay dates, and booking status
- Combines the request, the reservation details, and property reference information into a structured prompt for the AI
Stage 3: Validate before anything is delivered
We built validation in two places, not one.
- On input. Schema and completeness checks and reservation matching confirm the AI is working from correct, current data. Sensitive personal information is protected throughout the workflow.
- On output: human review, edit, and approval. Every AI-generated response goes to a staff member for review. They check it against the request and the reservation details and can edit the response to correct details, add information, or adjust the tone. Once they are satisfied, they approve it, and only then is it delivered. Nothing reaches a customer without a person signing it off. Deevita built the review interface and the data foundation that feeds it and logs its outcomes.
Stage 4: Deliver and learn
Once a response is approved, the platform completes the loop:
- The approved response, including any reviewer edits, is sent back to the customer through the same channel the request came in on: email, WhatsApp, Zendesk, or voice
- For each request, the data used, the original AI draft, any edits made by the reviewer, and the approval decision are logged for audit and improvement
- Reviewer edits show exactly where the AI fell short and are used to improve the prompts and the context the AI receives
How we delivered it
We worked in five phases, and each one had to pass a quality gate before the next began:
| Phase | Focus |
|---|---|
| 1 | Unified intake across all channels, with Azure Service Bus FIFO queuing |
| 2 | Request classification, reservation matching, and context building |
| 3 | Integration of AI response generation with the review, edit and approval interface |
| 4 | Live pilot with reservation data sync and channel delivery |
| 5 | Full rollout with governance and continuous tuning |
What our client said
“The Deevita team did a great job developing and optimizing our database architecture, related platform and connector products, and associated Azure services. They were very attentive to detail, incredibly quick to respond and adapt, and a pleasure to work with throughout the project. Their work played a key role in building our AI platform for hospitality, ensuring a solid and scalable foundation for future growth. Their professionalism and dedication made the entire collaboration a great experience.”
CEO of an AI startup
Results: what changed in the 8 months after full deployment
Measured by scoring AI-generated responses by month:
| Area | Outcome |
|---|---|
| AI accuracy | Response accuracy rose from 75.3% at launch to 85.1% |
| Human review | More than 70% of AI drafts were approved by staff on the first review |
| Data safety | No PII exposure across AI-assisted interactions |
| Efficiency | Escalations fell from 9.4% to 8.4% |
| Customer experience | Satisfaction with AI-assisted responses rose to 4.3 out of 5 |
These gains came from the data foundation, richer context and the reviewer feedbacSk loop working together.
Is your data ready for AI? Six questions to ask
If you are planning to bring AI into your operations, these questions show how ready your data is:
- Unified: Can you capture inputs from every channel in one standard format?
- Ordered: can you process requests reliably and in sequence, even at peak volume?
- Connected: Can AI reach your systems of record, and is that data fresh enough to rely on?
- Enriched: does each request reach the AI with the full context needed to answer it correctly?
- Validated: Do you check the data going into the AI, and does a person review, edit where needed, and approve what comes out?
- Learning: do reviewer edits and decisions feed back into improving the system?
Frequently asked questions
What does data readiness for AI mean?
It means the data an AI system relies on is unified across channels, processed in order, connected to your systems of record, enriched with the context each request needs, validated on the way in and out, and improved by reviewer feedback.
Why do AI projects stall?
Many stall not because of the model but because the data is not ready: it sits in many systems, comes in different formats, is checked unevenly, and is hard for AI to reach at the moment it needs it.
How do you keep AI responses safe to use with real customers?
Validation happens on input and on output. Schema and completeness checks confirm the AI works from correct data, sensitive personal information is protected throughout, and every AI response is reviewed, edited if needed, and approved by staff before delivery. Every step is logged.
How fresh is the data the AI works from?
Reservation data is synced every 30 minutes, so the AI typically works from booking details that are no more than 30 minutes old.
What did Deevita build, and what did the client build?
Deevita built the data foundation (multi-channel intake, ordered queuing, reservation integration, context building, validation and audit layer) and the React-based review interface. The client’s team built the AI response generation.
Conclusion
AI is only as reliable as the data underneath it. By unifying intake and ordering requests, enriching them with live reservation context, and keeping a person in the loop, we turned a manual process into a fast, dependable AI-assisted one. The same foundation applies wherever AI needs to act on real operational data. Explore our AI services or read more case studies.
About Deevita
Deevita is a Microsoft partner specializing in data integration and platform engineering. We help organizations bring data together, make it trustworthy, and put it to work for AI.
Is your data ready for AI?
Talk to Deevita about a data readiness assessment:
Ask Our Experts →
(sales@deevita.com) or
contact us
.