Case study: How we automated 30% of compliance work with AI (without breaking anything)

A real project. Real constraints. Real lessons.

Financial compliance work looks simple from the outside.

In reality, it’s an enormous amount of careful manual labor.

Every time a financial institution runs an audit, analysts have to review large document packages: bank statements, tax filings, regulatory documents, disclosure schedules, board resolutions.

One order can include 40+ documents. Sometimes the entire package arrives as a single 100-page PDF.

Someone has to read everything.

Check numbers.
Verify structures.
Assign review codes.
Flag potential violations.

And they have to do it thousands of times per day.

Our client provides workflow software for compliance teams. Their platform ingests audit orders, processes them, and generates deliverable packages for financial institutions.

Think of it like a factory for compliance.

Orders come in.
Documents get reviewed.
Reports go out.

Before AI, almost all of that work was manual.

The scale of the problem

Each order required an analyst to:

👉 split the document package into individual files
👉 read every document against a checklist
👉 cross-reference financial figures
👉 assign the correct review codes
👉 flag anything suspicious

Jurisdiction rules made things even harder.

What counts as a regulatory red flag in one country might be irrelevant in another.

There are hundreds of review codes, and the correct one depends on the document type, jurisdiction, and engagement.

Analysts were doing highly skilled work.

But most of it was repetitive pattern recognition.

The goal

Leadership had a simple objective.

Reduce operational costs by replacing repetitive manual work with AI.

But they didn’t want a risky “flip the switch” automation project.

They wanted something safer.
Something measurable.

The key rule: keep humans in the loop

Instead of full automation, the team started with a simple rule.

Let AI do the work.
Let humans verify.
Track accuracy.

If analysts approve 99 out of 100 AI results, that step can be automated later.

Not before.

This principle shaped every decision.

The starting point: a prototype illusion

A stakeholder had already built a prototype.

Front-end only.
No database.
No security.
No integration.

But in their mind it was nearly production ready.

Their argument was simple.

“We built this in two weeks. Professionals should finish it in two days.”

Reality required a careful conversation.

A prototype is not a product.

Production means security, scalability, integration with real data, and reliability under heavy usage.

The only way to prove that was simple.

Ship fast.

Three weeks to production

Instead of long planning cycles, the team leaned heavily on AI tools throughout development.

AI accelerated nearly every stage.

Designers generated interface concepts in hours.
Developers used AI-assisted IDEs for implementation.
An AI agent in the project management tool drafted user stories automatically.

The result was surprising.

The first production version shipped in three weeks.

AI wasn’t just the product being built.

It was part of the development process itself.

Why traditional discovery failed

After the first release, the team ran user interviews to decide what to build next.

But most users had never used AI tools before.

They couldn’t imagine what the system could do.

Feedback was vague.

Discovery took longer than building the product.

So the strategy changed.

Ship something simple.
Let users interact with it.
Observe real behavior.

Real usage produced far better feedback than hypothetical conversations.

What AI now does in the workflow

The AI system targets specific manual steps.

Each capability replaces a clearly defined task.

Document splitting

Orders arrive as large PDFs. The AI analyzes page titles and content and splits the package into individual documents like financial statements, regulatory filings, and tax records.

Previously analysts did this by hand.

Cross-reference verification

The AI extracts financial figures and compares them against source records.

It runs in two stages: data extraction and reconciliation.

Humans still review the result.

Compliance checks

This is the most complex task.

The AI reviews documents against structured checklists for disclosure completeness, unusual entries, related-party transactions, and regulatory gaps.

Jurisdiction rules are embedded into decision trees inside prompts so the model follows structured paths rather than open reasoning.

Review code suggestions

There are hundreds of review codes.

The AI suggests the correct one based on document type and jurisdiction. Analysts approve the final decision.

Order chat

Users can now ask questions about an order and receive answers directly from order data.

Keeping the model reliable

When dealing with financial and legal documents, creativity is dangerous.

Prompts were built using decision trees.

Instead of asking the model open questions, the system guides it through structured checks.

If condition A is present, apply rule set B.
If condition B is absent, move to rule set C.

Predictability matters more than clever reasoning.

Another surprising lesson was about tone.

Early AI responses said things like “this is incorrect.”

Users resisted the tool immediately.

Changing the phrasing to suggestions dramatically improved adoption.

Same information.
Completely different reaction.

How AI communicates matters.

Architecture

The AI system lives in a central engine, separate from the workflow platform.

Systems send a request.
The AI runs a prompt chain.
Structured JSON responses return to the platform.

Data access happens through an MCP server, which exposes order data as callable tools.

Security uses short-lived delegation tokens, limiting what the AI can access.

Some processes take up to fifteen minutes, so the system runs asynchronously in the background.

When the primary model provider experienced a four-hour outage, automatic failover switched to a backup model and the system stayed online.

Results so far

AI capabilities are now active in production.

Every order flows through them.

So far the system has achieved 30% of the client’s annual cost-reduction target.

And the platform continues expanding with new AI capabilities.

What actually worked

👉 Human in the loop first
Accuracy was measured before any step was automated.

👉 Ship fast
Real usage produced better feedback than interviews.

👉 Break complex tasks apart
One task per prompt improved reliability.

👉 Constrain the model
Decision trees worked better than open reasoning.

👉 Pay attention to tone
AI suggestions were adopted more easily than AI accusations.

👉 Use AI to build AI
AI-assisted development dramatically reduced delivery time.

The biggest takeaway from the project isn’t that AI replaced compliance analysts.

It removed the most repetitive 30% of their work.

Which allowed humans to focus on the decisions that actually require judgment.

If someone on your team is building AI into a product right now,
forward this. It might save them three months of wrong turns.