AI Agents & AutomationFree worksheetPDF, Excel, Word export

Chatbot Intent Design Worksheet for ERP and Support Bots

A chatbot is only as good as the list of things it is meant to do. This worksheet makes you define each intent, the data it needs, the identity required and what happens when the bot is unsure, before any model is trained or prompted.

Download:

Intent name	Example user phrase	Data source or API	Action type	Identity needed	Training phrases written	Expected asks per day	Minutes saved per ask	Hours saved per month	Phrase target met (10+)	Fallback when unsure
								117.3333	1
								33	1
								27.5	0
								17.6	0
								7.3333	1

Total hours saved per month

202.7667

Intents meeting phrase target

Intents defined

Share of intents ready to train

60%

Hours saved assumes 22 working days per month and is an estimate for ranking intents, not a business case. Any intent that writes to a system should require approval and identity checks.

Check your intents against your ERP data

Optional: we can look at your list and say which intents your ERP APIs can support today.

Optional. The tool and every export work without this. No spam.

How to use it

Collect real questions from ticket history, email and chat logs instead of guessing. Group them into intents, give each a short name, and paste a representative phrase. Note which system or API answers it, and whether the bot only reads or also changes data. Count your written training or test phrases, and estimate daily volume and the minutes each ask costs a human today. The sheet ranks intents by monthly hours saved so you can start with the ones that pay back.

Read versus write

Start with read-only intents. Order status, stock checks and policy answers carry little risk and build trust. Any intent that creates or changes records, such as approving a purchase order or opening a work request, needs strong identity, a confirmation step and an audit trail. Mark those as write with approval and keep a human in the loop. Intents that touch pay, bank details or personal data deserve a separate security review.

Authenticate before returning customer-specific data
Confirm before any write
Log intent, user, input and result
Never let the bot guess a document number

Confidence and fallback

Every intent needs a defined fallback for low confidence, ambiguous input and system errors. Good fallbacks ask one clarifying question, offer a short menu, or hand over to a person with the conversation attached. With large language models the same logic applies: ground answers in the source system, show where the answer came from, and say when nothing was found. Review unmatched messages weekly and add the real ones as new intents.

Frequently Asked Questions

How many training phrases does an intent need?

Classic intent classifiers usually want ten or more varied phrases each, with more for similar intents. LLM based bots need fewer examples but still benefit from a test set of real questions.

Which intents should I build first?

High volume, low risk, read-only ones with a clean API behind them, such as order status or stock lookup. They show value quickly and give you real usage data.

Can one bot serve WhatsApp and Teams?

Yes, with a channel layer that formats replies per platform. Keep intent logic separate from the channel so you design it once.

What is a safe confidence threshold?

It depends on the model and the cost of a wrong answer. Set a higher threshold for write actions, test against real phrases, and tune from logs rather than picking a number up front.

Netray builds grounded, on-prem chat agents over ERP data with read-only defaults. If a vendor chatbot is enough for your list, we will say so.

Talk to an Expert Explore More Free Tools

Related Tools

AI Agents & Automation

More AI Agents & Automation tools

See all

Back to All Tools