Why AI Automation Often Breaks at the Handover

You may already use software for enquiries, appointments, stock updates, customer records or internal approvals. The difficult part is not having separate tools. It is getting them to work together correctly when a real customer request is incomplete, unusual or spread across several systems.

An AI assistant might understand a request but choose the wrong action, skip a required step or send an answer before checking the latest information. For a small team, one incorrect update can create follow-up work, customer frustration and confusion about who should fix it.

Google Research’s ToolGrad approach offers a practical idea: prove that a sequence of tool actions works first, then create the user request that matches it. The research reports a pass rate of 99.8%, compared with 63.8% for a query-first approach on the ToolBench test database. Source

TL;DR

ToolGrad builds and verifies the tool-use workflow before writing the customer question. For your SME, the lesson is simple: test the process and its checkpoints before asking AI to handle conversations automatically.

This can help you create safer automations for sales enquiries, bookings, support, stock checks and admin approvals.

What This Means

When an AI system uses tools, it does more than generate text. It may search a database, call an external service, update a record, create a task or send a message. These actions form a tool-use chain.

For example, a customer may ask whether an item is available and request delivery. A reliable workflow could be:

1. Identify the product and required quantity.
2. Check current stock.
3. Confirm the customer’s delivery area.
4. Calculate the applicable delivery option.
5. Create a sales order only after confirmation.
6. Send a clear reply and assign any follow-up task.

A query-first system starts with a made-up customer question and tries to discover a successful sequence afterwards. If the sequence fails, that example may be discarded. ToolGrad reverses the order. It first creates a working sequence by executing available APIs, then produces a matching query and response.

The research describes four repeated modules: an API Proposer suggests possible tools, API Executors run them, an API Selector chooses the best result, and an LLM Updater adjusts the synthetic request and answer. The default setup uses 10 iterations across 50 sampled APIs for one workflow. Source

For your business, the useful principle is not “let AI try harder”. It is “make the correct process explicit, testable and repeatable before AI handles it.”

How This Applies to Malaysian SMEs

Customer enquiries: Suppose you run a renovation company, air-conditioning service or wholesale business. Customers may ask for availability, a quotation and a suitable appointment time in one message. Instead of allowing an AI assistant to reply immediately, define the verified chain first: read the service type, check the service area, inspect technician availability, prepare the required information and route unusual cases to a person. This reduces the risk of promising a slot your team cannot fulfil.

Retail and distribution: If you sell through a website, marketplace and direct WhatsApp enquiries, stock information can become fragmented. A verified workflow should identify the correct product code, check the approved inventory source, record a reservation only when the required conditions are met and notify the responsible staff member. The AI should not guess from an old message or create an order from an ambiguous product description.

Professional services: Accountants, consultants, agencies and training providers often need to qualify leads before arranging a consultation. You can create a workflow that collects the service requested, company size, preferred timing and relevant documents, then checks whether the request fits your service rules. AI can draft the next response, but the system should clearly separate information gathering from appointment confirmation.

Food, hospitality and events: A booking assistant may need to check date availability, group size, menu requirements and deposit status before confirming a reservation. By testing this chain first, you can prevent duplicate bookings and ensure special requests reach the correct staff member. This is especially useful when one person handles messages while another manages the calendar or operations.

Internal approvals: You can also apply the same thinking to leave requests, purchase requests and maintenance reports. The workflow might validate required fields, check the relevant approval limit, notify the correct manager and record the decision. AI can help interpret natural-language requests, but the approval sequence should remain visible and controlled.

What the Research Numbers Tell You

The ToolGrad evaluation compared its method with a depth-first search approach on the ToolBench API database, which contains more than 16,000 real-world APIs. The figures below come from the research summary. Source

Measure	Query-first DFS	ToolGrad	Practical lesson
Pass rate	63.8%	99.8%	Verify the workflow before generating examples
Successful tool uses per sample	2.1	3.4	Reliable workflows can support more complete processes
Tool-use steps per sample	34.3	20.0	Better selection can reduce unnecessary exploration
LLM invocations per sample	64.5	63.9	Better structure matters more than simply adding calls

The reported 0.2% failure rate occurred when the system could not obtain a successful response from three selected APIs across all 10 iterations. Source For an SME, this is a useful reminder that even a well-designed workflow needs an exception path. A failed tool call should create a clear task for a staff member instead of leaving the customer without an answer.

Practical Takeaways for Your Business

• Choose one repetitive process first. Start with booking, enquiry qualification, stock checking or document collection rather than trying to automate the whole business.
• Draw the workflow before selecting AI. List each system involved, the required input, the expected output and the person responsible if something fails.
• Separate read actions from write actions. Checking a record is lower risk than changing a record, sending a message or confirming an appointment.
• Set approval points. Require human confirmation for unusual requests, sensitive records, refunds, large orders or commitments the business cannot easily reverse.
• Use realistic test cases. Include incomplete questions, spelling variations, mixed languages, duplicate requests and unavailable items.
• Keep an activity record. You should be able to see which tool was called, what information it used and what result it returned.
• Create a fallback response. For example: “I need to verify this with our team before confirming.”
• Review failures weekly. Turn each failed case into a new test example or a clearer business rule.

The Bigger Picture

ToolGrad points towards a more disciplined way to build AI assistants. The strongest systems will not be judged only by how naturally they write. They will also be judged by whether they follow the right process, use the right information and stop safely when they cannot verify an action.

The research team reports that a ToolGrad-500 dataset containing 500 samples improved Gemma-3 performance at 1B, 4B and 12B parameter sizes. Its 12B model scored 83.1 on the Berkeley Function Calling Leaderboard, close to Gemini 2.5 Pro at 83.2 in the reported comparison. Source You do not need to build a model of that size to benefit from the lesson. You can apply the same method when documenting and testing your own automations.

For Malaysian SMEs, the long-term advantage will come from turning everyday business knowledge into clear workflows. Your staff already know the exceptions, approval rules and customer expectations. The next step is to capture that knowledge in a form that software can follow and your team can inspect.

Start small: select one process, verify every action, define what happens when information is missing, and only then allow AI to handle the conversation. Reliable automation is built less by asking AI to improvise and more by giving it a proven path to follow.

Ready to Streamline Your Operations?

Your business should run itself. AutoRunBiz deploys AI agents to automate your daily operations — WhatsApp orders, invoicing, customer follow-ups, and accounting. Book a free 15-min ops audit to see where automation fits your business →