Case study · HR & Workforce

We built the AI that files timesheet approvals. Nobody switched it on.

MITS Global Consulting filed every client’s timesheet approval by hand. We built an AI agent that does it in the user’s own browser, with a person confirming every match. It works. It was never switched on, and the reasons are the lesson.

  • ClientMITS Global Consulting
  • SystemTimesheet approval filing
  • Delivered10-Apr-2026, parked
  • Stack.NET 8 · GPT-4o
The outcome

Delivered in 24 days, then parked

The system worked on real approval emails. What it lacked sat on the client’s side.

  • 24 days

    From first commit to delivery
  • 104

    Commits
  • 4

    Manuals handed over
At a glance

Project snapshot

Client
MITS Global Consulting Pvt. Ltd.
Domain
HR & Workforce
Engagement
AI enablement
Build
ASP.NET Core MVC on .NET 8, Edge extension
AI
Semantic Kernel over GPT-4o
Automation
Playwright in the user’s own Edge
The story

What we built, and why nobody switched it on

Context

Aditi IT Software Solutions built a timesheet approval system for MITS Global Consulting, an IT staffing and consulting firm of about 700 people, in March and April 2026. Every month each client’s manager approves each consultant’s timesheet by email, and someone at MITS has to find that email, work out which consultant and which month it covers, save it as a PDF and file it in the right client folder in Google Drive. Several people shared that work.

Constraint

The approvals arrive as ordinary email from the managers of MITS’s client companies, in no common format: forwarded threads, a sender prefix added by the client’s mail system, several months in one thread, a consultant named only in the subject line. The Drive folders and file names follow a fixed convention, and the employee list changes every month and arrives as an Excel file. And this was financial work.

Decision

Two requirements ruled out the API route. MITS needed each approval filed as a PDF printout of the mail exactly as it appears, and the Gmail API returns a message’s content, not the rendered mail. And MITS asked for an agent the user could watch. So the system works inside the user’s own signed-in Edge browser: Playwright drives the Gmail and Drive tabs the user already has open, step by step, on screen, and prints each thread to PDF as the user sees it, instead of an app reading the mailbox and Drive through their APIs out of sight.

The AI never files on its own. A GPT-4o model reads the open thread and proposes each consultant, month and client; the system matches the name against the month’s employee list; the user confirms every match on a card in the browser side panel before anything is saved. Each card shows how sure the match is, lets every field be edited, and offers up to three other likely employees.

Build

Back end: ASP.NET Core MVC on .NET 8, with Semantic Kernel calling GPT-4o through the OpenAI API, the same provider as MITS’s HR letter system.

Edge extension (Manifest V3): a side panel in React 18 and TypeScript, with live progress over SignalR.

Browser automation: Playwright connected to the user’s Edge, for Gmail, Drive, PDF creation and upload.

Matching and filing: fuzzy name matching against the monthly employee list (EPPlus for the Excel upload, validated on upload); one service builds every Drive path and file name from the convention.

Hand-over: user, installer, developer and use-case manuals.

The last three weeks were driven by real approval emails, each odd pattern a fix and a test.

Stack

What it is built on

  • ASP.NET Core MVC
  • .NET 8
  • Semantic Kernel
  • OpenAI GPT-4o
  • Playwright
  • React 18
  • TypeScript
  • SignalR
The result

What happened

Delivered on 10-Apr-2026 after 104 commits in 24 days, with its manuals, and paid for. MITS never took it into the monthly routine, and it is parked.

It was not a technology failure. The reasons sat on the client’s side: no shared, clear picture of how the approval process actually ran from inbox to folder; no implementation team to set it up and test it end to end; little appetite for risk with AI in financial work, on MITS’s first AI project; and the plain weight of trust and habit in a routine people had done by hand for years.

That is the lesson worth more than the code. An AI system in a financial routine needs, before the build starts, the process written down as it really runs, people on the client’s side who will own the switch-over, and a stated appetite for the risk. Without them, a working system stays parked.

Proof

In the client’s words

What we value most about working with Aditi IT Solutions is that they took the time to understand how our business and HR processes actually work before developing the solutions.
Sandeep Dhurka Founder Director, MITS Global Consulting, on the wider engagement (the HR letter and onboarding systems)
65+ systems delivered since 2011

Named where the client has given written permission, described by sector where they have not.

See the full delivery record