Mentomy | Private AI for Businesses

Many companies take the step of implementing an AI-powered virtual assistant, upload all their documents... and find themselves with confusing responses, contradictory information or outdated data. The problem? It's not the technology, it's the lack of data preparation.

In this article we show you the essential best practices so your AI implementation is a success from day one. From data curation to role design, through the differentiation between public and internal knowledge.

The most common mistakes when integrating AI

Before talking about solutions, let's look at the mistakes 80% of companies make:

"Upload everything and done": Loading documents without review generates responses based on contradictory or outdated information
Disorganized information: Mixing current policies with old versions, drafts with final documents
No access control: The entire team has access to all information, including sensitive or confidential data
Mixing public and internal: Internal documentation appears in the public customer chat, or vice versa
"Set and forget": Not updating knowledge when policies, products or services change
Lack of maintenance: Nobody reviews response quality or identifies areas for improvement

💡 Remember: An AI assistant is only as good as the data you provide. "Garbage in, garbage out" has never been more true.

1. Data curation: The foundation of success

Data curation is the process of reviewing, cleaning and organizing information before feeding your AI assistant. It's the most important step and the one most companies skip.

What to do in data curation:

✅ Eliminate contradictory information

If you have two documents saying different things about the same topic, the assistant won't know which is correct.

Example: One version says "30-day warranty" and another says "90 days". The assistant could give either response.

Solution: Identify conflicts, determine which information is correct, delete or update the rest.

✅ Remove obsolete versions

Drafts, old versions and "deprecated" documents should be removed from knowledge.

Example: Keeping the 2023 product catalog when the 2025 one already exists generates confusion.

Solution: Implement a versioning system. Only the latest approved version should be available.

✅ Unify terminology

If different departments use different terms for the same thing, the assistant loses context.

Example: Sales talks about "prospects", Marketing about "leads" and IT about "potential customers".

Solution: Create a corporate glossary and ensure documents use consistent terms.

✅ Structure information

Well-structured documents with clear titles, sections and hierarchy make it easier for AI to find the correct information.

Example: A 200-page PDF without index vs. a document with chapters, subtitles and clear organization.

Solution: Use headings (H1, H2, H3), lists, tables and hierarchical structure in your documents.

✅ Remove unnecessary sensitive information

If a document contains information that shouldn't be widely shared, edit it before uploading.

Example: A sales report with personal customer data that only needs to show aggregated statistics.

Solution: Redact sensitive information or split the document into public and internal versions.

💰 ROI of curation: Companies that invest time in curating their data before launching their assistant reduce poorly answered queries by up to 70% and increase user adoption by 85%. The time invested pays back quickly.

2. Continuous updates: Keep your AI relevant

Once your assistant is implemented, the work doesn't end. Your company's information evolves constantly: new products, policy changes, regulatory updates, etc.

Update strategies:

📅 Scheduled audits

Establish periodic knowledge reviews (monthly, quarterly, semi-annually depending on your industry).

🔔 Reactive updates

When you change a policy or launch a product, immediately update related documents.

👥 Area owners

Assign each department the responsibility to keep their documentation updated.

📊 Quality monitoring

Review assistant conversations to identify incorrect or outdated responses.

🗂️ Recommended versioning system:

1. Use clear naming: Vacation_Policy_2025_v3.pdf
2. Mark obsolete documents: [OBSOLETE] Vacation_Policy_2024.pdf
3. Maintain a change log: what was updated, when and why
4. Archive old versions outside the assistant (but keep them for legal/historical reasons)

3. Role design: Who accesses what

Not all information should be available to everyone. Role-based access control is fundamental for security, privacy and relevance.

Recommended role architecture:

Level 1: Public information

Who accesses: Customers, website visitors, any external user

What it includes: Product catalog, FAQ, contact information, return policies, service conditions

Example: Chat on corporate website accessible without login

Level 2: General internal information

Who accesses: All employees

What it includes: Vacation policies, benefits, org charts, corporate calendar, general procedures

Example: Chat on corporate intranet with basic authentication

Level 3: Departmental information

Who accesses: Employees of specific departments

What it includes:

• Sales: Playbooks, approved discounts, pricing, pipeline
• HR: Payroll, personal data, evaluations, contracts
• IT: Technical documentation, access, infrastructure
• Legal: Contracts, NDAs, litigation

Level 4: Confidential information

Who accesses: Senior leadership, executive committee

What it includes: Corporate strategy, expansion plans, M&A, sensitive financial data, executive compensation plans

Example: Separate assistant with multi-factor authentication and complete traceability

🔐 Principle of least privilege: Each user should only have access to information necessary to do their job, nothing more. This not only protects sensitive data but also improves experience by reducing information noise.

Practical example of role structure:

Company: TechCorp SaaS

External Assistant (public web)
├─ Product documentation
├─ User guides
├─ Technical FAQs
├─ Plans and pricing
└─ Support policies

General Internal Assistant (all employees)
├─ Employee handbook
├─ Vacation and leave policies
├─ Benefits
├─ Corporate directory
└─ Culture and values

Sales Assistant
├─ All of the above +
├─ Sales playbooks
├─ Competitor comparisons
├─ Special discounts and pricing
├─ Proposal materials
└─ Success stories

HR Assistant
├─ All of the above +
├─ Payroll data
├─ Performance evaluations
├─ Hiring information
├─ Disciplinary policies
└─ Labor compliance

Executive Assistant
├─ All of the above +
├─ Financial reports
├─ Corporate strategy
├─ Investment plans
└─ Confidential information

4. External chat vs. Internal chat: Clear separation

One of the most important decisions is what information to expose publicly and what to keep internally. Here's how to structure it:

External chat (for customers)

🌐 Where: Public website, landing pages, support section

🎯 Goal: Answer questions from potential and existing customers, reduce support tickets, improve user experience

📄 Recommended content:

✓ Product and service information
✓ User guides and tutorials
✓ Technical and commercial FAQs
✓ Return and warranty policies
✓ Contact information and hours
✓ Terms of service
✓ Use cases and testimonials

🚫 What NOT to include:

✗ Internal company policies
✗ Employee information
✗ Non-public special pricing or discounts
✗ Internal technical documentation
✗ Internal operational processes

Internal chat (for employees)

🏢 Where: Corporate intranet, employee portal, internal platform

🎯 Goal: Facilitate onboarding, resolve HR questions, access technical documentation, improve productivity

📄 Recommended content:

✓ Employee handbook
✓ HR policies (vacation, leave, benefits)
✓ Internal operational procedures
✓ Technical documentation and troubleshooting
✓ Compliance and security regulations
✓ Org charts and internal directory
✓ Training materials
✓ Expense and travel policies

🔒 Security:

✓ Requires authentication (preferably SSO)
✓ Role-based access control
✓ Access auditing
✓ Sensitive data encryption

Practical case: B2B software company

💬 External Chat

User asks: "How much does the Enterprise plan cost?"

Response: "The Enterprise plan starts at $5,000/month and includes unlimited users, 24/7 support and 99.9% SLA. I can help you calculate the exact cost based on your needs."

✓ Public information, transparent, conversion-oriented

🏢 Internal Chat

Employee asks: "What's our margin on the Enterprise plan?"

Response: "The Enterprise plan has variable costs of $1,200/month and base price of $5,000/month, giving a 76% gross margin. Sales reps can offer up to 15% discounts without approval."

✓ Confidential information, only for authorized employees

5. Monitoring and continuous improvement

Once your assistant is implemented, it's crucial to measure its performance and improve it continuously.

Key metrics to track:

📊

Resolution rate

% of questions the assistant answers correctly without escalation needed

Target: >80%

⭐

User satisfaction

Average rating users give to assistant responses

Target: >4.2/5

⏱️

Response time

Average time for the assistant to provide a useful response

Target: <5 seconds

📈

Adoption rate

% of potential users who actually use the assistant

Target: >60%

🔄

Repeated queries

Users asking the same question multiple times (signal of inadequate response)

Target: <10%

💬

Human escalations

% of conversations requiring human intervention

Target: <20%

Continuous improvement process:

1. Weekly review: Analyze the week's conversations. Identify questions without satisfactory answers or problem areas.
2. Identify gaps: What common questions can't it answer? Is there missing documentation on any topic?
3. Content creation: Generate documentation to cover those gaps. Don't wait to have everything perfect.
4. Testing: Test the same problematic questions to verify improvements.
5. Iteration: Repeat the process continuously. Improvement is incremental.

6. Implementation checklist: Step by step

Follow this list to ensure a successful implementation of your AI assistant:

📋 Phase 1: Preparation (Week 1-2)

☐ Define objectives: Why do you want the assistant?
☐ Identify audiences: Who will use it? (customers, employees, departments)
☐ Inventory existing documentation
☐ Designate area knowledge owners
☐ Establish success criteria and KPIs

🧹 Phase 2: Data curation (Week 3-4)

☐ Eliminate duplicates and obsolete versions
☐ Resolve contradictions in documentation
☐ Structure documents with clear headings and sections
☐ Unify terminology and glossary
☐ Redact or separate sensitive information
☐ Create metadata (tags, categories, access levels)

🔧 Phase 3: Configuration (Week 5)

☐ Upload curated documentation to platform
☐ Configure roles and access permissions
☐ Separate external vs. internal knowledge
☐ Define assistant tone and personality
☐ Configure restrictions and response policies
☐ Establish human escalation flows

🧪 Phase 4: Testing (Week 6)

☐ Create bank of 50-100 test questions
☐ Test all questions and evaluate responses
☐ Verify sensitive information doesn't leak
☐ Check role-based access control
☐ Beta testing with users from each department
☐ Collect feedback and make adjustments

🚀 Phase 5: Launch (Week 7)

☐ Communicate launch internally/externally
☐ Provide usage guides and best practices
☐ Conduct live demos for key teams
☐ Integrate chat into website/intranet
☐ Activate monitoring and analytics
☐ Establish user feedback channel

📊 Phase 6: Continuous optimization (Ongoing)

☐ Weekly review of metrics and conversations
☐ Monthly documentation updates
☐ Quarterly complete knowledge audit
☐ Incorporate user feedback
☐ Expand knowledge to new areas
☐ Report ROI and improvements to stakeholders

Common mistakes and how to avoid them

❌ Mistake 1: "We set it up once and forgot about it"

Your company's information changes constantly. An assistant without maintenance becomes quickly outdated.

✓ Solution: Establish a regular update process with clear owners.

❌ Mistake 2: "We uploaded everything without review"

Contradictory, outdated or poorly structured documents generate confusing responses.

✓ Solution: Invest time in data curation before launch. It's worth it.

❌ Mistake 3: "Everyone has access to everything"

Without access control, you expose sensitive information and overwhelm users with irrelevant content.

✓ Solution: Implement roles from day one. It's easier to start restrictive and open access than vice versa.

❌ Mistake 4: "We don't measure anything"

Without metrics, you don't know if the assistant works well or where to improve.

✓ Solution: Define KPIs from the start and review them weekly.

❌ Mistake 5: "We mixed public and internal"

Internal information appears in public chat, or customers can't find what they're looking for because there's too much internal content.

✓ Solution: Clearly separate external and internal assistants from design.

❌ Mistake 6: "We didn't train users"

Users don't know how to use the assistant effectively or don't trust it.

✓ Solution: Create usage guides, do demos, share examples of useful questions.

Success stories: Companies that did it right

🏢 Consulting firm with 150 employees

Problem: New employees took 3 months to become productive. HR overwhelmed with repetitive questions.

Solution: Implemented internal assistant with all onboarding documentation, policies and procedures, curated and structured by departments.

✓ Result: Onboarding time reduced to 3 weeks. HR queries down 65%.

🏥 Private clinic

Problem: Reception collapsed with calls about prices, hours, test preparation.

Solution: External chat on website with service information, test preparation, prices and medical FAQs approved by their medical committee.

✓ Result: 40% fewer calls to reception. Patient satisfaction increased from 3.2 to 4.6/5.

🛒 B2C E-commerce

Problem: High abandonment in checkout process. Many unresolved doubts about shipping, sizing, returns.

Solution: Chat on product page and checkout with information about sizing, materials, return and shipping policies. Curated data from 500+ product sheets.

✓ Result: Conversion rate increased 23%. Returns down 15% (better purchase decisions).

🏦 Fintech

Problem: Sales team wasted time searching for information about products, competitors, pricing.

Solution: Internal sales assistant with playbooks, comparisons, success stories, special pricing. Access only for commercial team with granular role control (junior vs senior).

✓ Result: Sales cycle reduced 18%. Sales increased 31% in 6 months.

Conclusion: Proper implementation makes the difference

AI technology is incredibly powerful, but data quality and implementation strategy are what really determine the success or failure of your project.

Remember the key points:

1. Invest time in curating your data before launching. It's the foundation of everything.
2. Establish a process of continuous updates. Information changes, your assistant must evolve.
3. Design clear roles and permissions from the start. Security and relevance go hand in hand.
4. Clearly separate external and internal knowledge. Don't mix audiences.
5. Measure and optimize constantly. Data will tell you where to improve.

With these best practices, your AI assistant won't just work, it will become an indispensable tool for your team and customers, improving efficiency, reducing costs and increasing satisfaction.

Ready to successfully implement AI in your company? Discover how Mentomy supports you at every step, from data curation to continuous optimization, with tools designed to facilitate best practices.

Best practices for integrating AI in your company: Complete guide for success

Implementing an AI assistant goes beyond uploading documents. Discover how to curate your data, keep it updated, design access roles and differentiate between public and internal knowledge to maximize AI value in your organization.

Sources used for this article