Start with match rates. What percentage of customer records are consistent between CRM and ERP? If it’s below 80%, that’s your first cleanup priority. You don’t need 100%. You need “good enough” for your specific use case. 

No. Clean the data needed for your first specific improvement. If you’re targeting Days Sales Outstanding (DSO) reduction, focus on customer and invoice data. Product hierarchy cleanup can wait. Sequence improvements to deliver value continuously rather than waiting for perfection. 

For a targeted cleanup (one domain, one use case): 4-8 weeks to “good enough”, assuming you have one system of record, executive sponsorship and someone who can make decisions about data conflicts.  

Without those conditions, it will take twice as long. For enterprise-wide data governance: 12-18 months before any AI deployment. The targeted approach delivers value faster and reveals what’s actually needed for the next improvement. 

No. AI on messy data produces faster, more frequent mistakes. Data quality fixes require discipline, not technology. Once data is clean, AI can help keep it clean, but it can’t clean what’s broken. 

Operational gaps, not technology. Enterprises tried to automate processes that didn’t work manually, deployed on messy data, skipped stabilization work, and never defined success metrics. Only 21% redesigned workflows; 60% generated zero material value. 

Yes, and that’s the advantage. Budget constraints force focus, timeline pressure forces quick decisions, and smaller organizations force alignment. Every limitation that seems like a disadvantage is actually an advantage when you’re watching enterprises stumble.  

Learn more from enterprise missteps here. 

You can’t afford pilot proliferation, so you pick focused initiatives with clear ROI. You can’t wait 18 months for data perfection, so you fix targeted issues that pay for themselves. You don’t have a data science team, so you use production-ready tools instead of custom experiments. 

Define success metrics before buying tools. Fix operations to improve operational ROI (AI-readiness is a side effect). Deploy where work actually happens. Kill underperformers early. Start with stabilization, not innovation.

Narrow applications with human oversight: data entry automation, predictive analytics with human review, intelligent automation for routine decisions under $10K, workflow-embedded recommendations. The pattern is AI assists, human decides. 

Emerging capabilities are real but not reliable at scale. Enterprises are testing them with mixed results. Vaporware is marketing promises with no production track record. Emerging might work in 2-3 years. Vaporware might never work as promised. 

Ask for three customers who have been in production 18+ months. Ask about error rates and error handling. Ask about data quality requirements. Ask what happens when it fails. If they dodge or make claims that contradict published research, be skeptical. 

Fix your data foundation first (it delivers operational ROI regardless of AI), then deploy narrow, production-ready tools on that clean foundation. Skip the transformation promises. Take the boring wins. 

An AI AUP should have three clear boundaries:  

1. What’s allowed without asking (safe zone) 

2. What requires approval (yellow zone) 

3. What’s prohibited (red zone).  

Plus an escalation path for unclear cases.  

Create psychological safety. Employees hide usage when they fear consequences. Clear boundaries, amnesty for past actions, plus approved alternatives equal visible, manageable AI usage. Prohibition creates shadow AI usage; clarity prevents it. 

Amnesty. If you punish past behavior that happened before policies existed, you guarantee future hiding. Announce AI usage policies, explain why they matter, give amnesty for past actions and enforce going forward. 

Measure AI ROI the same way as any operational improvement:  

• What metric improved?  

• By how much?  

• What did it cost?  

If vendors tell you AI requires special ROI frameworks, they’re dodging accountability. 

No. Did DSO decrease? Did close time improve? Did forecast accuracy increase? Use existing operational metrics. If improvements can’t be measured in business terms you already track, be skeptical of the value. 

• Data quality fixes (customer master, product hierarchies, financial reconciliation consistency).
• Process standardization (documented workflows, consistent execution).  
• System integration (data flows between systems rather than through spreadsheets). 

Only for small, contained experiments with clear learning goals. “Let’s test if this AI tool would help with collections prioritization” is reasonable, with the explicit understanding that production deployment requires data cleanup first.