Prompt Grounding in Salesforce: Keep AI Smart & Sane!

What?

This is a plain-English “how-to” on adding grounding rules to your Salesforce AI prompts – guardrails for ethical, content-related, dynamic, and data security.

Who?

Salesforce Admins, Business Analysts, Architects, Product Owners, and anyone who wants to tap their Salesforce + AI to its full potential without AI going haywire.

Why?

To Write better AI prompts. Accelerate Revenue. Slash Grunt Work.

Make AI safer, more trustworthy, and aligned with human values.

What can you do with it?

Common use cases for Salesforce + AI Grounding rules include:

Make Responses Ethical: Prevent Bias, Toxicity, and AI responses that may be offensive
Bring appropriate Content: Reduce speculation, made-up or incomplete responses, and add suitable formatting and tone of voice.
Add Contextual Relevance: Use AI for chatbots, data analysis, process automation, and more – with personalization, context, and user settings.

Kids say the darndest things...So, do AI Chatbots 🙁

A Chevy dealership’s newly deployed AI chatbot agreed to sell a car for $1. Why?

Most parents will tell you that their kid has potential but needs guidance – the same is true with AI.

Chevrolet’s AI lacked the facts-based grounding needed to respond appropriately. Like a naive toddler, it soaked up whatever it was told without discretion.

Until your AI has proper guardrails (a.k.a - grounding), your Gen AI may be a child prodigy - genius, but gullible!

What is prompt grounding?

AI systems are getting downright chatty these days. Without helpful guidelines, they might start spouting nonsense or offending people left and right.

That’s why wise human overlords came up with “instruction manuals” for not being digital nitwits.

Let us break down the deets for you with some handy real-world examples.

Ethical Grounding: Don't Be a Jerk

Think of ethical grounding as teaching AI the societal do’s and don’ts. For instance, these grounding rules instruct your AI to avoid generating content that could be discriminatory or offensive.

Think of these as topics you would avoid on the dinner table with your in-laws.

AI knows some subtle stuff for a walking calculator. Here are some minefields you want your AI to avoid.

Practice Impartiality: Treat all humans as equals, even if they root for terrible sports teams. AI responses should be based on facts, not opinions or biases.
Stay Neutral Like Switzerland: Politics, religion, and stereotyping are not what your business needs AI for. Set it up to avoid toxicity and discrimination.
Avoid Harmful Content: AI Models(LLM) are trained on lots of data; some of it may be offensive or dangerous and can contaminate responses.

Ensure that your AI is aligned with your corporate values.

				
					Ethical Grounding Rules:

- Do not generate content that can be deemed offensive or inappropriate

- Avoid using slang or colloquial terms unless specifically asked

- Prioritize safety and ethicalconsiderations in your answers

- Avoid using emotive language

With these types of ethical grounding rules, major social faux pas can be avoided.
However, how can you ensure that these are always a part of your Salesforce+AI efforts?

3 challenges of incorporating grounding rules

Standard grounding across a team of Prompt Engineers
Consistent enforcement for every Salesforce+AI prompt
Automated inclusion of grounding rules for every AI call, including API ones

Adding grounding to your Salesforce+AI can be difficult - GPTfy standardizes, consistently enforces and automates it.

Content Grounding: Look Ma, no hands!

Completeness, Tone, Formatting – We all expect this when we ask a question from AI. However, a typical user prompt is simply a question, often with the expectation that all of these factors are encoded in their interaction with AI.

Unless you specify what to say, how to say(and present), AI will respond with whatever it likes.

Give your AI clear guidelines in these three areas to ensure it doesn’t start discussing the latest celebrity gossip when you ask for stock market updates.

Well-formed and Accurate Responses

Like a squirrel on caffeine, AI can veer off course without direction. Keep responses laser-focused on the task at hand – no going on tangents!

				
					Content Grounding Rules - Response Accuracy:

- Do not hallucinate

- Address all parts of a multi-part question

- Only answer what has been asked for

Appropriate Tone of Voice

Use clear communication suitable for the audience and setting – Just like you don’t get too chatty with the bank teller or too technical with your grandma!

				
					Content Grounding Rules - Tone of Voice:

- Aim for clarity over complexity

- Avoid using absolutes, such as "always" or "never"

- Avoid using passive voice

- Avoid using slang or colloquial terms unless specifically asked

- Use 3rd person voice for summarizations

Formatting for End User

Rather than letting your AI guess the format your users want, let it know…and leave the hypothesizing to Dr. House.

For example, Auto-generated Emails may need nicely formatted HTML, while a ‘Root Cause Analysis’ can be plain text.

				
					Content Grounding Rules - Formatting:

- Use data from the provided JSON, linked by 18-digit IDs.

- Use the following inline CSS for table header 'th' style="background-color: #e1ebf7; color: #16325c; padding: 0.5rem; text-align: left;"

- Use the following inline CSS for caption = "font-size: 1.0rem; font-weight: bold; padding: 0.5rem; color: #5C5C5C;text-align: left;"

Applying content-based grounding as part of your Salesforce+AI prompts has the same issue – how do we make it part of the AI-Ops process and not leave it to the whims and fancies of individual prompt engineers?

Consistently and easily apply your content-based grounding rules in Salesforce with GPTfy's built-in support

Dynamic Grounding: Say Please...In 日本語

A Global Salesforce deployment means your users expect to see things in their preferred way – in their locale for dates or even in their language.

Dynamic grounding allows AI to adjust its responses based on the current user, their preferences, and context.

It's like having a GPS that updates its route in real time – always staying relevant and on course.

Know Your Audience: Like a good comedian, tailor your material to your audience. Adjust formality, tone, and pop culture references accordingly.
Locale: Customize units, currencies, time zones, etc., based on user location and preferences. French users dig kilometers, not miles!

				
					Dynamic Grounding Rules:

- When referencing dates and amounts, adhere to the USA locale conventions

- Adapt responses to user context

- Tailor responses to feedback

- Personalize based on preferences

- Implicitly adjust to user cues

Bring AI to Your Salesforce

Compatible with Pro, Enterprise & Unlimited editions. Save on licensing costs.

Get Now Watch Salesforce + AI Demos

2 considerations become important as you consider dynamic grounding

Active User: How do you present AI responses – based on their locale or with a company-wide default?
AI Mass Processing: What locale and preferences do you adhere to during automated AI processing?

You can automate dynamic grounding & personalization in your Salesforce+AI with GPTfy - for logged in users & automation.

This lets you personalize Salesforce user results by incorporating locale-specific formats for dates, currencies, etc.

It also allows you to set clear grounding guidelines based on corporate standards for the rest of the rules.

Data Security & Privacy: Keep Secrets, Maintain Trust!

Trust is the #1 AI adoption issue – particularly concerning PII, PHI, MNPI, trade secrets, and sensitive data.

How do you prevent AI data breaches and unauthorized access to other users or system data with a third-party LLM/AI?

The key is defense in depth - Prioritizing security with grounding rules, multi-layered masking, and zero data retention.

These components together form a crucial part of a comprehensive AI trust layer. Details on Salesforce+AI Data Security are here.

A typical use case is to mask personal data like emails, names, and phone numbers before data is sent to AI, and further add grounding rules to prevent any accidental spillage in its responses.

				
					Data Security & Privacy Grounding Rules: 

- Do not disclose private or personal information without explicit consent. Remove specifics like names, contact details, addresses, IDs, etc. and replace with placeholders.

- If sensitive categories like financials, healthcare, employment, or legal matters are detected, use abstract placeholders like [FINANCIAL_INFO], [HEALTH_INFO], [WORK_DETAILS], [LEGAL_MATTERS] rather than specifics.

- Refrain from revealing any identifiable details about an individual such as date+location of birth, full name of relatives, contact info, employer name+location, school details, etc. Replace with placeholders.

- In case of uncertainty around the sensitivity of certain data, err on the side of caution and use broader placeholder strings without specifics.

- Do not attempt to restore redacted information or regain access to deleted content. Respect and maintain implemented privacy protections.

Prevent Prompt Injection - Don't let your AI spill the beans

AI can be tricked into filling in information from memory, speculation, or assumption.

Add specific grounding rules to exclude private details and safeguard your Salesforce+AI from malicious prompts and chat interactions.

Let AI Scrub data that it recognizes as sensitive - Code of silence

Grounding rules can ask your AI to automatically scrub any data it identifies as sensitive during ongoing conversations, even if it’s anonymized.

TL;DR / Summary

"With great power comes..." Grounding is the responsibility of the Gen AI power - Salsforce + GPTfy lets you automate it, seamlessly.

Grounding rules are the training wheels that align your AI with your business’s ethics, safety, and corporate values.

Including grounding rules in your Salesforce requires a Prompt Engineering Platform for the full AI lifecycle. At GPTfy, we have built it.

Get GPTfy Now: AppExchange Link 🚀

Saurabh Gupta

Saurabh is an Enterprise Architect and seasoned entrepreneur spearheading a Salesforce security and AI startup, with inventive contributions recognized by a patent.