A Big Four accounting firm sets up its global capability centers with agents

One of the Big Four accounting firms is setting up its global capability centers (GCCs) with agents on the repeated work from the first day. VideoDB records how experts do that work, writes the steps as SOPs, and builds computer-use agents from them. Every agent run is recorded, so the firm can check any output and report on what its agents did.

Customer	One of the Big Four accounting firms
Where	The firm’s global capability centers (GCCs)
First workflows	Invoice processing, basic financial research, company valuation
First agent	About three weeks from the first recordings
Built on	VideoDB Capture Agent, video understanding and search, computer-use agents

The goal

The firm’s global capability centers handle manual work at scale: invoice processing, basic financial research, company valuation. Much of it is the same sequence of steps done again and again in ERPs, spreadsheets and filings. The firm wants agents doing that part from the start, so its people spend their time on the work that needs judgment.

Why it was hard

The steps of this work are written down nowhere complete. They live in how experienced people move through the screens: which field they check first, what they compare, when they stop and ask. Anyone building an agent for it has to learn the work one step at a time.

The firm’s EU clients add a second condition. Before anything is recorded, they want to see exactly where their data goes and what happens to it afterwards.

What VideoDB does

1. Records the work. The VideoDB Capture Agent records the desktops of people doing the task, on Windows and Mac. The person at the desk starts and stops it, or it starts on its own from a calendar. Recordings sit in an encrypted buffer on the machine and upload over HTTPS.
2. Writes the steps. VideoDB describes every step of each recording: the app, the field, the click and the frames behind it. From these it drafts the SOP, which the firm’s experts review. The recordings give the agent its demonstrations and give the firm the evidence to check it.
3. Builds the agent. The agents work through computer use, the browser and Excel. The invoice agent checks each invoice line, marks it ready, exception, hold or needs review, and writes the results to an Excel workbook. It reads only and makes no approvals or payments. A research agent gathers financial information from EDGAR filings, blogs and video into a structured workbook.
4. Keeps every run on record. Each agent records itself as it works, and its runs are indexed in VideoDB. Any output traces back to the moment on screen that produced it, so a reviewer can see why a line was marked an exception, and the firm can report on what its agents did.

Built around the client’s data rules

• Recordings go to the region the firm picks, in VideoDB’s cloud or in the firm’s own.
• Once a recording has been turned into steps and indexes, it can move back to the firm’s cold storage or be deleted.
• VideoDB shares its SOC 2 Type II report and signs a DPA for GDPR. See security.

About three weeks to a first agent

• Week 1: scope and record. Pick the task, agree what success looks like, collect the existing SOPs and onboarding documents, and record a few people doing the work.
• Weeks 2 and 3: build and check. Write the steps from the recordings, build the agent, and check its runs against them.
• After launch: every run on record. The agent’s recordings show what it did and where it needs help, and feed the next version.

What it means for the next center

Each new center can open with agents already doing the repeated steps, and every recorded run adds to what the firm knows about how its work gets done. The same path works for any team that runs the same process at scale: record it once, check the steps, then let the agent do it with the evidence kept.

If your team runs work like this, talk to us, or read how computer-use agents are built on VideoDB.