Annie Tsai

Even though the headlines this week are truly out of a techno forward sci-fi horror story (yes, I’m talking about ai agents posting on OpenClaw gossiping about their humans), the utter mundaneness of what it actually takes to get AI to work well with your business data and extract the real insights you need to move forward may as well be read from the encyclopedia. I will try to make this as entertaining as possible for you.

Do you remember that old saying that Data Scientists used to throw around? The one that goes “Data Science is like 90% cleaning crappy data up so we can do good data science.”

Same deal here.

So this week, let’s focus on three foundational moves that separate slop from superpowers.

Slop In, Slop Out

If your CRM, analytics, docs, and internal notes are messy, AI doesn’t magically clean it up. It will confidently produce garbage at scale.

What’s actually breaking inside AI organizations is accrued institutional knowledge that quite frankly looks pretty gnarly to AI inside of a database. Things like:

• Duplicate records across systems
• Free-text fields doing the work of structured data
• No clear data ownership or definitions
• Historical data that reflects old processes but is treated as current truth
• Large scale pricing, policy, organizational, or other relevant changes that signal shifts in data trajectory with no explanation

Here are some actionable steps you can take:

• Create a single source of truth per domain for customers, revenue, usage, inventory, and any other category that matters in your organization. Document this in plain language.
• Standardize naming conventions and required fields before introducing AI into workflows
• Remove or archive data that no longer reflects how the business operates today
• Create a timeline that outlines relevant policy, procedural, pricing, and other relevant changes.
• Add lightweight validation rules at data entry points instead of trying to fix these problems again downstream.

Context Training: Teaching AI How Your Business Works

After putting into production dozens of agentic systems, I have learned one important lesson: context is king. Generic models don’t understand your pricing logic, customer lifecycle, internal language, or edge cases unless you teach them. You really do have to treat your AI like it’s a new hire.

Where companies go wrong:

• Dropping AI into workflows without defining goals or constraints
• Feeding documents without explaining which ones matter more
• Expecting the model to infer business logic that only lives in people’s heads

So, in the absence of a magic wand or magic snapping fingers, here are a few steps to take in its place:

• Write a short “business context brief” for AI: how you make money, who your customers are, what matters and what doesn’t
• Curate a high-signal knowledge base (policies, FAQs, playbooks, real examples) instead of dumping everything first. You are more likely to get better output with a smaller but more accurately representative data sample.
• Use exemplars: show AI what a good output looks like using real historical decisions
• Regularly refresh your context as pricing, strategy, or org structure changes

Using AI to Step-Function Data Into Insight

I had to do a lot of this in my early days of building agents. Why? Because it is very hard to turn massive, messy datasets into insights that humans can actually act on.

What AI is uniquely good at:

• Synthesizing across systems and large datasets that humans rarely connect
• Surfacing patterns before dashboards are built
• Turning qualitative data into structured insight

What AI is not good at:

• Knowing which insights matter without understanding priorities
• Taking large amounts of unstructured data and extracting insights that take multiple steps
• Being able to consistently deliver expected outputs without clear definitions that account for context evolution and repetition
• Distinguishing correlation from causation on its own
• Understanding political, legal, or ethical landmines embedded in decisions
• Making tradeoffs when incentives conflict and data is incomplete
• Replacing accountability when decisions go sideways

Here are some actionable things to try:

• Use AI to summarize trends weekly across sales, support, and customer feedback
• Ask AI to generate hypotheses, then pressure-test with humans and create a loop to improve
• Pair AI insights with existing metrics to explain why numbers moved
• Design feedback loops where insights inform what data you collect next
• Keep final decision-making and exception handling with your human team members – stay with the human-in-the-loop model until you have tested out the anomolies

And lastly, this is the big one: Instead of relying on AI to synthesize across multiple steps, create new fields in your CRM or other database of record to enable AI to generate the necessary step-function information it needs in order to get to the desired insights

For instance:

Lead Generation: this is the website > this is who they are > this is what they sell > this is their buyer > this is the product they would use in our portfolio > these are other customers of ours like this > this is how we have successfully closed them > this is what you might say in a cold email

versus:

Lead Generation: this is the website > this is what you should say in a cold email

You see how AI might provide better output with one path over the other?

Your Takeaway This Week

AI readiness is so much more than an IT project, and it takes more than telling your team to “use AI”. Clean inputs, deliberate context, and systems designed to convert data into insights are what allow AI to compound inside a business. Your team can try all they want to integrate AI into their daily flows, but it’s still always going to be slop in, slop out.