Subodh Chawla is Senior Director and Divisional CFO at Scale AI. Before joining Scale, he was VP and Head of Finance at Z League and held finance roles at Google and Uber, including leading finance for part of Google’s global Ads organization.
“We’ve mostly used AI to reconcile a lot of the ‘what happened during the week’ problem, and then the ‘why it happened’ is more of a human-touch element.”
We run an extremely lean finance team for the size of the business, and AI is a big part of how we’re able to do that.
Scale is a consumption-based business, so we look at performance every week. A lot of the work we’ve done with AI has been around reconciling what happened during that week so the team can spend more time understanding why.
Finding what changed each week
There are a lot of inputs behind a consumption-based business. We look at things like billing specifications, hours billed and average handling time, but that information doesn’t always come in at the same time or in the same shape. Agents have been helpful in detecting those disparities.
“Our agents pick up some of these gaps of what exactly is the driver of why there’s a miss, and start highlighting whether billing specs were wrong, billing specs were missing, whether the number of hours, AHTs, what we call average handling time, whether that was off.”
Instead of starting with every input and working through them one by one, we can start with what looks off, which is a huge time saver.
Reconcile the agent before relying on it
A lot of leaders know that the biggest hurdle for AI deployment is data quality.
“The data is obviously garbage in, garbage out. That becomes a problem.”
The team already had workflows running through Snowflake and Redash, so our January/February project was to run Claude alongside them and reconcile the results.
They reconciled for the most part, but where they didn’t wasn’t always Claude’s fault. We’ve found filters missing from Redash and mistakes in workflows we had already been using.
Claude’s been a good check in that sense, and also useful for working backwards when something is wrong. We’ll give it the output we know we’re looking for and ask it to debug itself.
Sometimes it finds something we missed in the prompt or the workflow. From there, we can fix it and move on.
Building with AI is becoming part of the job
There’s been a learning curve for the team too. A lot of people came to us with experience in SQL, Python or workflow tools, but we want everyone on the team to have access to and be comfortable working with these tools.
“There’s a lot of effort in really asking people to use the tool, getting them access, getting them used to just using Claude. And then inch by inch, you start depending more on that.”
That is starting to change how I hire as well. One of my biggest interview questions now is how somebody is using AI. I want to understand what they actually built, where it failed and whether they understand how it works.
What we build ourselves
We have an internal AI enablement team, which gives us more room to build workflows ourselves.
One example is travel and expenses. We wanted to catch out-of-policy Uber trips, duplicate expenses and other T&E control issues, and the workflow didn’t take long to build.
Whether we build or buy really depends on the use case.
“If it’s a high priority and it is highly, highly unique to Scale, we can do this in-house because it’s almost easier for us to do that using tooling.”
We don’t do that for everything. For instance, we implemented Pigment for forecasting because that’s an area where dedicated software makes sense.
“The trade-off is your token cost and your Claude usage is blowing up, and then it’s the equation of: do we want to spend on Claude or do we outsource a lot of that spend to a vendor who can do it cheaply?”
That calculation will keep changing as models get cheaper and we use more open-source options.
More agents means more access controls
Some of the data finance works with is highly sensitive. I’m still cautious about putting individual salary information into these systems, and contribution-margin data also needs to be tightly controlled.
“We’re spinning up so many agents at this point that we also need some sort of control mechanism, so that an agent doesn’t slip up and suddenly show data that they wouldn’t show otherwise.”
We need permissions around the agents themselves, both for what data they can access and who can see the results.
The ROI is a smaller team
I don’t think about the return from AI only in terms of how many hours a workflow saves.
“My benchmark is that we run an extremely lean team.”
We’re hiring fewer people and getting more productivity from the people we do hire. So as AI moves further into accounting, we’ll start measuring things like days to close too.
But today, the biggest impact is in operational finance. AI helps us figure out what happened, so the finance team can spend more time on why.