Krishna Singh

I build AI solutions for ERP, EPM & FP&A teams. Sharing what I build, learn and test in real projects.

[Krishna Singh @krishnafitech]

AI in workforce planning should not just predict attrition.

It should connect attrition risk back to headcount planning, hiring assumptions, and forecast impact.

I tested this using Adaptive Planning / OneStream-style personnel data with a simple cloud AI architecture.

[Krishna Singh @krishnafitech]

What if finance could catch unusual expense transactions before close? I explored an AI architecture using Workday supplier invoice data: export transactions, score anomalies in Azure/AWS, then surface exceptions for review. Workday stays the system of record. Cloud handles AI.

[Krishna Singh @krishnafitech]

this gives finance teams an early signal on high-risk cost centers, accounts, and variance trends.

https://t.co/cu20vExGHP

[Krishna Singh @krishnafitech]

I explored how AI can support variance risk analysis using Workday Adaptive Planning.
Adaptive exports budget and actual data to SFTP.
Azure or AWS validates the data and runs the AI/ML logic. The prediction output is loaded back into Adaptive for dashboards and finance review.

[Krishna Singh @krishnafitech]

Most AI demos look simple.

Prompt.
Code.
Output.

But enterprise AI with Workday Finance and Adaptive Planning data is different.

In production, the model is only one part of the solution.

[Krishna Singh @krishnafitech]

@grantcrawley Yes. That’s the way it should be

[Krishna Singh @krishnafitech]

People often think AI/ML means prompt, code, and model output. In enterprise ERP/EPM, that is only the demo.
In production, AI/ML needs process knowledge, finance data, planning data, integrations, mappings, security, controls, governance, monitoring, and business adoption.

[Krishna Singh @krishnafitech]

Everyone is asking if AI will replace finance teams.

I think the bigger risk is different:

treating AI demos like production-ready finance systems.

This is Part 1 of a short series on what it actually takes to build AI around ERP, EPM, and finance data.

[Krishna Singh @krishnafitech]

What 15 Finance AI Builds Taught Me: AI Replacing People Is Not the Real Risk https://t.co/kzxBwmu3Mv

[Krishna Singh @krishnafitech]

Spent today inside Workday Extend building an FP&A accelerator. Biggest lesson: don’t fight the tool. Extend is built to surface Workday data, not to be an ETL engine. Prism dataset - WQL - grid is the path that just works.

[Krishna Singh @krishnafitech]

Rolling Forecast with Temporal Fusion Transformer

Many forecasting examples focus on a single time series.

Real planning environments rarely work that way.

In this lab, we built a rolling forecast model across five business units:

• Enterprise
• SMB
• APAC
• EMEA
• Public Sector

The model combines:

• Historical revenue
• Business unit context
• Future calendar information

One of the objectives was to see how forecast quality changes when the model can learn from both past observations and known future inputs.

The next step will be introducing operational drivers such as headcount, pipeline, bookings, and customer growth.

[Krishna Singh @krishnafitech]

Can transaction descriptions predict the correct GL account?

To explore that question, we built a BERT-based classification model that learns from previously coded transactions.

Examples included descriptions related to:

• Travel & Entertainment

• Software & SaaS

• Payroll & Benefits

• Marketing & Advertising

• Rent & Facilities

Instead of relying entirely on keyword rules, the model learns patterns from transaction descriptions and predicts the most likely GL account category.

We also included confidence scores so lower-confidence predictions can be routed for review.

This lab explores one possible approach to reducing manual coding effort while keeping finance review in the process.

[Krishna Singh @krishnafitech]

How W04 Loads Live Financial Data Inside Workday https://t.co/Mv612yb3tH

[Krishna Singh @krishnafitech]

I am building a RAG agent for finance close docs.

Not a generic chatbot.

The test is simple:

Can it answer close process questions using only source-backed citations?

Correct answer matters.

Correct citation matters more.

A confident wrong policy answer is still a failure.

[Krishna Singh @krishnafitech]

I learn more from failed models than successful demos.

https://t.co/asRAr0StwB

[Krishna Singh @krishnafitech]

Models are easy to showcase.

Maintenance is harder.

[Krishna Singh @krishnafitech]

The first AI output is usually wrong.

That is where the learning starts.

[Krishna Singh @krishnafitech]

The hardest part of AI is not building it.

It is fitting it into a real business process.

[Krishna Singh @krishnafitech]

@droidbuilds

[Krishna Singh @krishnafitech]

@AbhiCodes15 No coding. Silence.