Case Study

An AI Copilot That Takes Real Actions Inside a SaaS App

Merchants on a WhatsApp marketing platform were clicking through screen after screen to do routine setup. DXDasia built an AI copilot inside the app: type what you want in plain English and it does the work through real, permission-checked actions, live on production since September 2026.

Key result

Plain-English Commands, Real Actions, Zero Accidental Sends

Agentic AITool CallingSaaS DevelopmentAI Safety Design

Laptop showing a chat interface on a desk

The challenge

Routine work in the app took far more clicks than it should.

Finding contacts, tagging them and building segments meant several screens per task
Creating message templates required knowing exactly which options the platform would approve
Setting up a campaign was a multi-step form that new users found confusing
A chatbot that only answered questions wouldn't save anyone time

Our approach

We built an AI agent that can actually operate the app, with safety designed into its structure.

A floating command box in the client portal, with a saved command history for each account
Twelve tools the AI can call: contact search, add and block, tag creation and tagging, template listing and creation, campaign drafting, order lookup and usage summary
Each tool reuses the app's existing business logic, so the AI follows exactly the same rules as a person clicking through
A catalogue check the AI runs before offering catalogue templates, so it never drafts something the platform would reject
No send capability at all: the AI can draft campaigns but only a person can send them

The outcome

Merchants now get routine setup done by describing it:

Requests like 'tag everyone who ordered last month as VIP' run in seconds
Templates and campaigns arrive as ready-to-review drafts
Nothing reaches customers without a person pressing send
Every action is tested per tool and was verified on a live production account

Want a system like this for your business?Tell us what you need. You get a free plan and a fixed quote within one business day.

Why it worked

Most in-app AI assistants can only talk. Giving the copilot real tools made it useful, and building safety into what tools exist, rather than into instructions the AI might ignore, made it trustworthy. If there is no send tool, the AI cannot send.

The technical approach

The copilot uses an AI model with function calling. Each tool is a thin wrapper over the same code paths the portal already uses, so there is one source of truth for validation and permissions. Every tool has unit tests, the full loop was tested over HTTP with a mocked AI, and then verified with a live read and write on a production account, with the test data removed afterwards.

Why are businesses adding AI agents that take actions?

Because AI has moved from experiment to everyday tool, and the next step is letting it do work, not just answer questions. Stanford HAI's 2025 AI Index reported that 78% of organisations said they used AI in 2024, up from 55% the year before. Microsoft's 2025 Work Trend Index, a survey of 31,000 knowledge workers in 31 markets, found 81% of leaders expect AI agents to be moderately or extensively integrated into their company's AI strategy within the next 12 to 18 months, and 46% said their companies already use agents to fully automate workflows or processes.

Finding	Source
78% of organisations used AI in 2024, up from 55%	Stanford HAI, AI Index 2025
81% of leaders expect agents in their AI strategy within 12 to 18 months	Microsoft Work Trend Index 2025
46% already use agents to fully automate workflows	Microsoft Work Trend Index 2025

What is the main risk with AI that can take actions?

Giving it too much power. The OWASP Top 10 for LLM Applications (2025) lists Excessive Agency as a key vulnerability, caused by excessive functionality, excessive permissions or excessive autonomy. Its recommended mitigation is human-in-the-loop control that requires a person to approve high-impact actions. That is the principle behind our copilot: it has no send tool at all, and every campaign it creates waits for a person.

How can you apply this in your own business?

Start with the tasks people repeat List the five actions users do most often. Those become the first tools.
Reuse your existing logic Wrap the app's own functions as tools so the AI follows the same rules as a person.
Design safety into the toolset Leave out tools that could cause harm, such as sending or deleting, or make them draft-only.
Check before you create Have the AI verify preconditions, like account setup, before it builds something that would fail.
Test each tool on its own Unit-test every tool, then the full conversation loop, before going live.

Frequently asked questions

What is tool calling?

It lets an AI model choose and run specific functions in your software, such as searching contacts or creating a draft, instead of only replying with text.