June 2024.
I was obsessed with AI. I'd successfully built CrewFlowAI (my manifest processing tool). I'd implemented ChatGPT for SOP training. I was convinced: AI could transform our entire front desk operation.
Answer any guest question (via text/chat)
Handle routine reservations
Process simple requests (late checkout, extra towels, wake-up calls)
Provide operational guidance to staff
One AI to rule them all.
I hired a developer. We spent 3 months building it. Cost: $8,000.
Launch day: I was excited. This was going to revolutionize our operations.
Total cost: $8,000 + 40 hours of my time.
Total value delivered: Approximately $0.
This was a catastrophic failure.
But it taught me more about AI implementation than any success could.
Here's what I built wrong, why staff rejected it, and what I built instead that actually worked.
The AI Hotel Assistant (I called it "Atlas")
GPT-4 API (conversational AI)
Custom training on our hotel data (SOPs, guest profiles, local info)
Web-based interface (accessible on desktop and mobile)
Reservation modifications (change dates, add requests)
Operational Q&A ("How do I process a crew manifest?")
Policy lookups ("What's the late checkout policy?")
Guest history lookup ("Has this guest stayed before? Any preferences?")
Connected to PMS (Opera)
Pulled data in real-time
It was technically impressive.
And it failed spectacularly.
Mistake #1: I Built What I Thought Was Cool, Not What Staff Needed
I was in love with the idea of a comprehensive AI assistant.
What I thought: "Staff will love having one tool that does everything!"
I never asked them.
I assumed I knew what they needed because I'd been a front desk agent before.
"What's your biggest operational pain point?"
"Slow PMS—everything takes too many clicks"
"Inconsistent information from different managers"
"I wish I had an AI assistant to answer questions."
I built a solution to a problem that didn't exist (in their minds).
Faster PMS workflows (automation, shortcuts)
Clearer, consistent policies (better training, not AI)
I built the wrong product because I didn't validate the need.
Mistake #2: The Interface Was Too Complex
Chat window (for asking questions)
Guest lookup panel
Reservation management panel
Report generator
Settings/preferences menu
What I thought: "More features = more value!"
Staff opened it, saw the complexity, and immediately closed it.
During a busy check-in rush, agents need simple, fast tools.
Click to open app
Navigate to correct feature
Type query or select options
Wait for AI response
Interpret response
Total time: 45-90 seconds
Ask manager standing 10 feet away
Atlas lost on speed every time.
Complexity is the enemy of adoption.
If your tool requires more than 3 clicks or 10 seconds, people won't use it during busy times.
Mistake #3: It Didn't Integrate Into Existing Workflow
Atlas was a separate application.
PMS open on monitor 1
Email/Slack on monitor 2
Phone at desk
Open new browser tab
Navigate to Atlas
Log in (if session expired)
Perform task
This was a context switch.
Humans resist context switching. Every new tool = cognitive friction.
An integration within the PMS (pop-up panel, not separate app).
Or even better: A Slack bot (staff were already in Slack all day).
Your tool must fit into existing workflow, not create new workflow.
Mistake #4: Staff Didn't Trust the AI
AI occasionally gave incorrect answers (about 8% error rate in my testing).
Agent: "What's the cancellation policy for non-refundable rates?"
Atlas: "Guests can cancel up to 24 hours before arrival for a full refund."
This was wrong. Non-refundable means no refunds, period.
The AI had confused it with our standard flexible rate policy.
Agent used this answer with a guest. Guest later disputed when we didn't refund. Created a whole mess.
After that incident, the agent never used Atlas again.
And she told 5 other agents: "Don't trust Atlas—it gave me bad info."
AI needs to be 99%+ accurate in operations. 92% isn't good enough.
One bad answer destroys trust permanently.
Atlas: "Based on our policies, I believe the answer is [X], but I'm only 60% confident. Please verify with a manager."
Or even better: Only answer questions where confidence > 95%. For everything else: "I'm not sure—please ask your manager."
In operations, accuracy >> capability.
A tool that answers 50 questions with 100% accuracy is better than a tool that answers 500 questions with 92% accuracy.
Mistake #5: I Didn't Manage Change Properly
Sent email to all staff: "We have a new AI assistant tool. Here's the link. Try it out!"
Held one 15-minute training session
Get staff input before building
Pilot with a small group first
Collect feedback and iterate
Create champions (early adopters who advocate for it)
Most staff tried it once, got confused, and gave up.
No one was there to help them succeed. No one was cheerleading for it.
I treated Atlas like a feature (just add it, people will use it).
I should have treated it like a transformation (requires training, support, champions, iteration).
Technology adoption is 20% tech, 80% change management.
Negative feedback from staff ("too complicated," "don't trust it," "slows me down")
No measurable operational improvement
I pulled the plug.
I gathered the 3 agents who actually used Atlas regularly and asked: "Why did this fail?"
"Honestly? I don't need it. I can just ask my manager. It's faster."
"It's cool, but it's one more thing to remember to check. I'm already juggling PMS, email, phone, and Slack."
"I used it a few times, but once it gave me wrong info, I stopped trusting it."
Atlas solved my problem (I thought our staff needed better access to information).
It didn't solve their problem (they needed faster workflows, not more tools).
Three months later, I tried again. But this time, I started with staff interviews.
"If you could wave a magic wand and fix one thing about your daily work, what would it be?"
"I hate how many clicks it takes to do simple things in Opera. Like, adding a wake-up call is 7 clicks. Adding a note is 5 clicks. It's so slow."
They didn't need AI to answer questions. They needed workflow automation.
The New Solution: PMS Shortcuts (I called it "QuickActions")
A simple Chrome extension that added keyboard shortcuts to our PMS.
Click "Guest Services"
Click "Wake-Up Calls"
Select room number from dropdown
Enter time
Click "Save"
Press Ctrl+W (wake-up call shortcut)
Type room number + time
Ctrl+W → Add wake-up call
Ctrl+N → Add note to reservation
Ctrl+P → Post incidental charge
Ctrl+R → Pull guest history
Ctrl+M → Email confirmation to guest
Automatically filled common notes ("Guest requested early check-in")
Auto-populated email templates
Built with JavaScript. Took me 12 hours (vs. 3 months for Atlas).
Cost: $0 (I built it myself).
| Week 1 | Trained 5 agents on QuickActions. All 5 loved it. |
| Week 2 | Word spread. 8 more agents asked for it. |
| Week 4 | 22 out of 25 agents using it daily (88% adoption). |
| ✅ Simple | One feature (keyboard shortcuts), easy to understand |
| ✅ Fast | Saved time on every interaction |
| ✅ Integrated | Worked within PMS (no context switching) |
| ✅ Trustworthy | No AI = no errors |
Before: 4.2 minutes
Time Saved: 1.1 minutes per check-in
Annual: 1,205 hours
Development Cost: $0 (12 hours of my time, salaried)
Lesson #1: Validate the Need Before Building
What problem does this solve?
How painful is that problem? (1-10 scale)
How are people solving it now?
Is your solution faster/easier/better than current method?
If the answer to #4 is "No" → Don't build.
Lesson #2: Start Stupidly Simple
Your v1 should solve one problem exceptionally well.
Not 10 problems adequately.
Atlas tried to do everything. QuickActions did one thing (workflow automation).
QuickActions won because it was simple.
Lesson #3: Integration > Standalone
Build tools that fit into existing workflows.
Don't create new apps that require context switching.
| Bad | Separate AI chat app |
| Good | Slack bot (staff already in Slack) |
| Bad | New scheduling software |
Lesson #4: Accuracy is Non-Negotiable in Operations
In operational tools, 99% accuracy is the minimum.
If your AI can't hit 99% accuracy → Don't deploy it for critical tasks.
Use it for low-stakes tasks only (content generation, brainstorming, research—not guest-facing answers).
Lesson #5: Change Management is Everything
Interview them (what do they need?)
Pilot with 3-5 champions (early adopters)
Roll out to 20% of team first
Provide hands-on training (not just email)
Collect testimonials from early adopters
Share wins ("Agent A saved 30 minutes yesterday using QuickActions!")
Ongoing training for new hires
Regular updates based on feedback
Don't skip Phase 1. That's where Atlas failed.
"What's your biggest operational pain point?"
If 50%+ mention the same problem → High-priority problem
Only build for high-priority problems.
Build the simplest possible version in 1-2 weeks.
MVP #1: A Slack bot that answers 10 common questions (not 500)
Usage frequency
Accuracy
If MVP fails → Pivot or abandon.
If MVP works → Invest in full version.
Don't launch to everyone.
Launch to 5 early adopters (people who are tech-comfortable and influential).
Daily check-ins: "How's it going? Any issues?"
Immediate bug fixes
Goal: Turn them into advocates.
Adoption rate (% of team using it)
Frequency (how often per day/week)
Time saved (before vs. after)
If impact is weak → Iterate or kill.
If impact is strong → Scale.
Train full team (hands-on, not just email)
Create cheat sheets (quick reference guides)
Assign "super users" (people who can help others)
Framework #1: The AI Suitability Test
Is 99%+ accuracy required? (Yes = Risky for AI)
Is speed critical? (Yes = AI needs to be faster than current method)
Is the task high-stakes? (Yes = Build in human verification)
If you answered Yes to #1, #3, or #4 → AI is risky. Proceed with extreme caution.
Framework #2: The Simplicity Filter
Ask: "Can I remove this and still solve the core problem?"
If Yes → Remove it.
Repeat until you can't remove anything else.
That's your MVP.
Framework #3: The Adoption Prediction Score
FactorScore (1-10)Solves acute pain point/10Faster than current method/10Integrates into workflow/10Simple to use (< 3 clicks)/10High accuracy (99%+)/10TOTAL/50
If total score < 35 → Don't build. It will fail.
Step 1: Interview 10 staff members about their pain points
| Step 3 | Ask: "Is AI the right solution, or is there a simpler fix?" (Often, the answer is: simpler fix) |
| Step 4 | Build the simplest possible MVP (1-2 weeks max) |
| Step 5 | Pilot with 5 people, measure impact |
| Step 6 | Only scale if pilots succeed |
Don't spend $8K and 3 months building something nobody wants.
The Hospitality Insider
Enjoyed this post? Get weekly hospitality insights delivered to your inbox.
No spam, ever. Unsubscribe anytime.
Was this helpful?
Written by
Front Desk Manager at Galt House Hotel, managing 1,300+ rooms daily. Published author of 3 books on hospitality operations, leadership, and personal growth.
AI models are advancing simultaneously across three critical dimensions: intelligence, speed, and extensibility, fundamentally reshaping how businesses operate and creating unprecedented competitive advantages.
Hotels face severe staffing shortages with 14% fewer employees than pre-pandemic levels, prompting the industry to turn to AI chatbots and robotics to maintain service standards.
