The AI Implementation Checklist We Use Before Recommending Anything to a Client

We turn down about 40% of prospective AI projects. Not because we don't want the business — because the AI implementation wouldn't deliver the results the client needs. This is the checklist we run before recommending AI, and it's the same checklist that saves our clients from expensive failures.

Why We Say No (And Why It's Good for Everyone)

The AI hype cycle has created enormous pressure on businesses to "do something with AI." We've seen companies spend $80,000 on AI customer service implementations that were solving the wrong problem. We've seen marketing teams adopt AI tools that produced content their customers hated. We've seen businesses automate workflows that didn't need automation and left the workflows that actually needed help as manual as ever.

Our checklist exists to prevent waste. Both wasted money and wasted opportunity. If an AI implementation won't deliver 3x ROI within 18 months, we won't recommend it. That's our bar.

The 12-Step Checklist

Phase 1: Problem Validation

Step 1: Can you quantify the current cost of this problem?

Before we recommend AI, the client must be able to answer: "What does this problem cost us per month?" Not in vague terms — in dollars and hours.

Good answer: "We spend 25 hours/week on customer support for order status inquiries. At $30/hour fully-loaded, that's $39,000/year. Plus we estimate we lose $20,000/year in customers who ask about order status and don't get a response quickly enough."

Bad answer: "Our customer service is kind of overwhelmed."

If they can't quantify it, we don't move forward. You can't measure AI success without a baseline.

Step 2: What percentage of this problem is automatable?

We've found that 70-85% of typical customer service and operational workflows are automatable. The other 15-30% requires human judgment, empathy, or relationship management.

We ask clients to categorize their problem workflows by:

• Routine (same response every time): ~40% of most support loads
• Semi-structured (same process, different variables): ~30%
• Complex (requires judgment, relationship, novel problem-solving): ~30%

AI handles Routine and Semi-structured well. It struggles with Complex. If Complex is more than 40% of the workflow, AI is probably not the right solution.

Step 3: What would perfect automation look like?

We ask clients to describe the automated future state in specific terms, not vague aspirations. "Better customer service" is not a goal. "AI handles 80% of order status inquiries without human intervention, with 95%+ accuracy, and response time under 2 minutes" is a goal.

If they can't describe what success looks like specifically, they haven't thought through the problem deeply enough to solve it.

Phase 2: Data Readiness

Step 4: Is your data accessible?

AI agents need data. If the information needed to resolve a workflow is locked in email inboxes, PDFs, or someone's head, AI can't access it.

We evaluate:

• Where does the data live?
• Is it in structured systems (databases, APIs) or unstructured (emails, PDFs, scanned documents)?
• What's the API situation — does the system have a modern API or is it held together with screen scrapers and CSV exports?

If data is inaccessible or unstructured, we either scope a data-cleaning project first or recommend against AI until the data infrastructure is in place.

Step 5: Is your data accurate and consistent?

AI on bad data gives confident wrong answers faster than humans. We've seen implementations fail because the client had 30% duplicate customer records, inconsistent product descriptions, and inventory numbers that were off by 20%.

We run a data quality audit before recommending AI. If data quality issues would cause AI to fail more than 10% of the time, we fix the data first.

Step 6: Do you have enough training data?

Custom AI agents learn from examples. If a workflow has 10 examples per month, there isn't enough signal to train a reliable agent. We typically want 100+ examples of a workflow before building a custom agent.

For common workflows (customer service, lead qualification), most businesses have plenty of data. For rare workflows (handling a specific type of dispute that happens twice a year), there isn't enough to learn from.

Phase 3: Business Readiness

Step 7: Do you have a process owner?

AI implementations fail when nobody owns them. We require clients to designate a process owner — the person responsible for:

• Monitoring AI performance weekly
• Approving changes and updates
• Handling escalations that AI can't resolve
• Identifying new automation opportunities

If a client says "everyone will use it" without a designated owner, the implementation will drift and eventually fail.

Step 8: Can your team work with AI, not just around it?

Some teams adapt to AI workflows well. Others treat AI as a threat and work around it — routing AI-resolved issues back to themselves, ignoring AI recommendations, or reverting to manual processes because "the old way was better."

We assess team readiness through:

• Have they been consulted on the problem, or is this being done to them?
• Do they understand the goal (freeing them for higher-value work) or do they see it as a threat?
• Are there performance metrics that would make them feel evaluated by AI?

If the team isn't ready, we either do change management work first or don't proceed.

Step 9: What's your escalation strategy?

Every AI implementation will fail on some percentage of cases. The question is: what happens when it does?

We require clients to design escalation paths before implementing AI. Questions we ask:

• What does the customer experience look like when AI can't help?
• How quickly will a human see the escalated issue?
• What context does the human receive about what AI tried?
• How do we prevent the customer from having to repeat information they've already given?

If escalation is an afterthought, customers will have terrible experiences when AI fails.

Phase 4: Technical Readiness

Step 10: What integrations are required, and are they feasible?

AI agents are only as good as their access to systems. We map the technical integrations required:

• CRM (sales/lead AI)
• Order management (customer service AI)
• Inventory system (operational AI)
• Email/messaging platform
• Knowledge base / help docs

For each integration, we assess: Does the system have an API? Is the API documented? Are there rate limits? Are there sandbox environments for testing?

If a required integration doesn't have an API, we either scope API development or don't proceed.

Step 11: What's your security and compliance posture?

AI agents handling customer data must meet security requirements. We evaluate:

• What customer data will the AI access?
• Does the AI store this data? For how long?
• What's the data retention and deletion policy?
• Are there compliance requirements (HIPAA, SOC 2, PCI) that apply?
• Who has access to AI interaction logs?

If compliance requirements aren't met, we scope a security implementation before AI deployment.

Step 12: What's your measurement framework?

Before we implement anything, we define success metrics:

Primary metrics (what we're optimizing for):

• Volume handled by AI (target: 70-85%)
• Accuracy rate (target: 95%+ for automated resolutions)
• Customer satisfaction for AI-resolved issues (target: match or exceed human CSAT)

Secondary metrics (what we monitor for side effects):

• Escalation rate (should be < 15%)
• Time to resolution for escalated issues (should not increase)
• Human agent CSAT (should improve as routine work decreases)

Business impact metrics (how we measure ROI):

• Labor hours saved
• Revenue impact of faster response times
• Customer retention impact of improved service

If clients can't commit to measuring these metrics monthly, we don't proceed. You can't optimize what you don't measure.

What Happens When We Say No

When a prospective client fails the checklist, here's what we tell them:

"Your data isn't ready." We scope a data-cleaning project first. This might take 2-3 months but without clean data, any AI we build will fail.

"The workflow isn't automatable enough." We explain what percentage is automatable and what the human-AI hybrid model would look like. If they're not comfortable with the hybrid model, we don't proceed.

"Your team isn't ready." We recommend change management work — communication, training, involvement in design — before any technical implementation.

"The ROI doesn't work." We show them the math. If an AI implementation would cost $40,000 and save $15,000/year, the payback is almost 3 years. We recommend they wait until volume increases or scope decreases.

"This isn't an AI problem." Sometimes the real issue is process design, organizational structure, or a broken product. AI would automate the wrong thing. We explain what we'd actually recommend instead.

The Checklist in Practice: A Rejection Example

Here's a real example of a project we turned down:

A boutique hotel chain (8 locations, $12M/year revenue) wanted an AI concierge to handle guest inquiries. They imagined a chatbot that could answer questions about restaurants, local attractions, and hotel services 24/7.

Our assessment:

• Step 1: They estimated 200 guest inquiries/week. Most were "what time is breakfast?" and "can I get extra towels?" Cost of current handling: essentially zero (front desk handled it while doing other tasks).
• Step 2: 60% of inquiries were automatable. 40% were relationship-based (guests asking for personalized recommendations, handling complaints about noisy neighbors, coordinating late checkout with cleaning).
• Step 3: They wanted AI to "feel like a helpful human concierge." We don't have AI that does that reliably.
• Step 10: They wanted integration with their legacy property management system from 2009 that had no API and required a DOS-era interface.

Our recommendation: Don't build AI concierge. The volume doesn't justify it, the legacy system can't support it, and the expectation doesn't match current AI capability. Instead: build a simple FAQ page and app integration that would cost $8,000 and handle 80% of inquiries without AI.

We turned down a $45,000 project. They went to another vendor. We heard 6 months later that the implementation was abandoned after $30,000 spent.

Practical Takeaways

• If you can't measure the current cost of a problem, you can't measure AI's impact on it. Start with measurement.
• The question isn't "can AI help?" — it's "can AI help at a cost that delivers ROI within your timeline?"
• Data readiness is often the gating factor. Bad data + AI = fast, confident wrong answers.
• AI implementations require process owners, escalation paths, and measurement frameworks. Without these, they'll drift and fail.
• Be willing to say no to AI projects. The opportunity cost of a failed AI implementation is higher than the cost of not doing it.

Ready to set up OpenClaw? Contact Cuanto Labs at setclaude.com