Skip to content
MCP AI Business Leader

AVOID COSTLY AI MISTAKES: What Every Business Leader Should Know

Tim Wetmore
Tim Wetmore
You know AI could transform your business. Your competitors are using it. Your team is asking about it. But when you start researching implementation options, you're hit with a wall of technical jargon: APIs, RAG, LLMs, embeddings, vector databases...
Sound familiar?

Here's the truth: You don't need to become a technical expert to make smart AI decisions. But you do need to know the right questions to ask your technical team—and understand the business implications of their answers.

I've approached AI from a business leader's perspective, focused on solving problems and improving processes rather than chasing technology for its own sake. In this guide, I'll walk you through the essential layers of AI implementation in plain language, highlighting what really matters for your business success.

AI deployment Schematic
AI Deployment Schematic

Click on the arrows below to see details on each layer

UNDERLYING DATA SOURCES
UNDERLYING DATA SOURCES
What are they?

Real-world scenario: Your sales director asks your AI agent, "What's the total lifetime value of Acme Corp?" The AI gives three different answers because your CRM uses "Acme Corporation," your accounting system uses "Acme Corp," and your support system uses "ACME." This isn't a technical problem—it's a revenue intelligence problem.

Inaccurate data = bad decisions, whether human or AI-powered

Your underlying data sources could be the separate systems that help you to manage your business – think accounting, invoicing, CRM, manufacturing control, logistics, ecommerce etc. Within each there could be specific information such as identification numbers and text from emails and notes. You may have data sitting in Google drive word and excel documents. You may have standalone databases of information and maybe some of the information from your systems is consolidated and pushed into a central database like Snowflake.

What’s Important?

Cleaning up data issues AFTER AI implementation can cost 3-5x more than addressing them upfront.

The underlying data in your business needs to be accurate and integrated. AI thrives on context and the more information it has access to the better your answers will be. Imagine you have an account in your CRM and the same account in your accounting system but the account name is different in each – you need to be able to connect the 2 data sources. There are many ways to achieve this without adding work and this is a typical scenario for automation. i.e. create an automated flow that auto-creates an account in your accounting system when an opportunity is closed-won in the CRM and make sure a common identifier is placed in both systems.

Key fact: With modern integration and automation platforms you don’t need to throw out your existing systems – you can connect them.
THE CONNECTION LAYER: HOW AI ACCESSES YOUR DATA
What is it?

Think of this as the plumbing that connects your AI to your business systems. Just like you need the right pipes to get water from the main line to your faucet, you need the right connectors to get data from Salesforce, QuickBooks, or your manufacturing system to your AI.

The Model Context Protocol (MCP) is the emerging standard for these connections—similar to how USB became the standard way to connect devices to your computer.

What’s Important?

Many software vendors are rushing to release their own AI connectors, but here's what business leaders need to watch for:

  • Control: Can you access all the data fields you need, or only what the vendor decided to include?
  • Cost: Are you paying per connector? Some businesses end up with 15+ separate connections, each with its own fee.
  • Maintenance: Who's responsible when a connector breaks? Your IT team or the vendor?
  • Security: Each connection point is a potential vulnerability. More connections = more risk.

The mistake we see businesses make: Letting each department pick its own AI tools and connectors, creating a management nightmare. One client had 23 different AI connections, each with different security protocols and maintenance requirements.

Key Fact: The connection layer is invisible to users but will determine your costs, security posture, and how quickly you can adapt as AI evolves. Choose a platform that can manage multiple connections centrally.
INTEGRATION & AUTOMATION
What is it?

Integration to AI is different to integrating between 2 systems where we are sampling moving and possibly manipulating data in some way. We focus below on 3 particular concepts which are important to AI integrations.

What’s important?

When AI Should Give Opinions vs. When It Needs Certainty:
Some business questions are judgment calls: "Which customers are at risk of leaving?" AI can analyze patterns and give you its best assessment—and being wrong isn't catastrophic. You'll investigate and make the final call.

Other questions require 100% accuracy: "Create an invoice for Customer A for $15,247.32." There's no room for AI to "assume" the amount or guess the customer. For these situations, your AI system needs to be designed to STOP and ask for clarification rather than forge ahead with assumptions.

Business impact: Companies that don't design for this difference end up either with AI that's too timid to be useful, or AI that makes expensive mistakes with critical data.

Human Oversight: Safety Valves for High-Stakes Decisions: Even when AI has perfect information, some decisions are too important to automate fully. Your system should make it easy to add approval steps where needed—like requiring a manager's sign-off before AI sends a proposal over $50,000 or makes a purchase order.

Company-Specific Knowledge: Your Secret Weapon: ChatGPT knows what's on the internet. It doesn't know your WiFi password, your internal product codes, or your company's negotiation strategies. A properly configured Knowledge Base (sometimes called RAG) ensures AI looks in your company's documents FIRST before searching the web.

Real example: A client's customer service AI was giving outdated return policy information from old web pages. By setting up a Knowledge Base with their current policies, they eliminated this issue entirely.

Key fact: Integration of data and integration of data with AI are the same but in a different way.
SECURITY AND AUTHENTICATION
What is it?

The question every board member asks: "What happens if an AI agent makes a mistake that costs us money or exposes customer data? Who's liable?”

This is probably one of the most important questions a business leader has. For many industries like banking and insurance it’s not just about access but being able to keep a log of who is doing what. If an AI agent is being used there is typically a person asking the questions which we need to know about.

What’s important?

There are 2 entities asking for access – the AI agent and the person interacting with the AI agent. You can control the AI agent access and then control access to the agent or you can use a solution like Workato where you can have the AI agent take on the authentication requirements of whoever is using the agent.

If you have multiple AI access points you need to manage multiple security protocols. Consider using a platform with at least SOC 2 compliance to manage agent access.

Ensure your system is able to log every interaction to meet regulatory requirements.

Consider this scenario: Your sales AI agent accesses customer data to prepare a proposal. Should it have access to:
- Only the accounts that salesperson manages?
- All accounts in their region?
- The entire customer database?
The answer depends on your business, but the key is having the ABILITY to control this granularly.

Key fact: Consider using a platform such as Workato to deploy your agents and reduce the number of weak points you need to manage
 
LARGE LANGUAGE MODELS (LLMs)
What is it?

No need to go into too much detail here. We are of course talking about ChatGPT, Claude, Gemini, Perplexity….. The LLM allows us to interact but the output is determined by what the LLM is trained on – it could be the worldwide web or it could be information accessible only to your employees or both. You have the ability to determine this.

What’s important?

LLMs are new and changing all the time. They all work in similar ways and you will find they have strengths and weaknesses – Claude for code, ChatGpt for tone, Gemini for pictures, Perplexity for research etc.

Today, Claude might be best for your use case. In six months, it might be a different model. In a year, there might be an industry-specific model built for your sector.

The mistake we see: Companies that hard-code their agents to work with only one LLM. When that model raises prices or a better option emerges, they're stuck.

What to look for: An architecture that lets you switch models without rebuilding everything. Some platforms let you A/B test different models for the same task to find the best performance.

Key fact: Ensure you have the option to choose and even switch out which LLM works best for your application
 
AI AGENTS
What are they?

Agents are typically designed to do specific functions. This determines who has access to the agent, what information the agents have access to and what skills they need. Examples could include:

  • Customer 360 agent for a consolidated view across all data points
  • CPQ Agent which uses conversation and CRM data to create orders
  • Commissions agent to calculate sales commissions
  • Forecasting agent using CRM data and conversation data
  • New hire onboarding agents provides information to new hires and ensures they have access to all the systems they need
What’s important?

There are multiple tools available for building agents. This should not restrict how you deploy those agents, what LLMs they connect to and what MCPs and security layer they sit on top of.

ROI Reality Check: Before building custom AI agents, calculate the cost of your current solution:

  • That $15,000/year SaaS tool for commission calculations? An AI agent might replace it for $3,000/year in LLM costs
  • That reporting tool that requires a data analyst to generate insights? AI can do it in seconds
  • That new hire onboarding process that takes 6 hours of HR time per employee? AI can handle 80% of it

The ROI case writes itself—but only if your architecture allows you to build and modify agents quickly.

Key fact: AI agents can be created for many purposes. This can allow you to retire some Saas software products that are currently doing those jobs.
USER INTERFACE
What is it?

Typically an LLM interface such as logging in to ChatGPT, or it could be a web based interface or a messaging tool like Slack or Teams.

What’s important?

The interface is the easy part—and that's actually the point. If users have to learn a new system, adoption will suffer. Meet them where they already work:

Key Fact: The best AI implementation is one your team uses. Start where they're comfortable, expand from there.

 

In summary here are some of the key messages.

Making Smart AI Decisions for Your Business

Here's what separates successful AI implementations from expensive failures:

Successful approaches:

  • Start with a clear business problem, not a technology
  • Choose platforms that give you flexibility as AI evolves
  • Design for change—AI is moving fast, and your architecture needs to keep up
  • Focus on integration and data quality from day one
  • Plan for security and compliance upfront, not as an afterthought

Common mistakes:

  • Letting each department pick its own AI tools (integration nightmare)
  • Choosing solutions that lock you into one LLM provider
  • Ignoring data quality issues ("we'll fix it later")
  • Building without thinking about security and audit trails
  • Trying to implement everything at once instead of starting focused
Questions to Ask Your Technical Team

Before approving any AI implementation proposal, ask:

  1. "If we want to switch AI models in 6 months, how much would we have to rebuild?"
  2. "How are we handling data integration—are we fixing our underlying data issues or just connecting messy data?"
  3. "Who manages and secures each connection point to our data?"
  4. Can you show me the audit trail for AI decisions?"
  5. "What happens when [vendor name] raises prices or changes their terms?"

If your team can't answer these clearly, that's a red flag.

How GoNavigate Can Help

At GoNavigate, we help SMEs cut through the AI hype and build implementations that actually work. We focus on:

  • Solving your business problems, not deploying the latest tech buzzword
  • Starting focused: one clear use case with measurable ROI, then expanding
  • Building on platforms that give you control and options

Want to explore how AI can work for your business without the technical overwhelm?

Let's talk about your specific challenges and find practical solutions.

Visit our website

Share this post

Keep reading