Public Data Is Abundant. Evidence Isn't.

There is an extraordinary amount of information about companies sitting in public view.

You can find out who they're hiring. You can see what technologies they use. You can read what their executives are saying. You can watch their headcount change.

You can find funding rounds, job postings, product launches, partnerships, acquisitions, advertisements, organizational changes, customer announcements, and whatever their CEO decided was profound enough to put on LinkedIn at 6:14 on a Tuesday morning.

If you're sufficiently determined, you can assemble a truly magnificent pile of facts.

We know because we've done it.

There is only one problem.

A pile of facts is not evidence.

At least, not by itself.

Evidence of what?

Suppose I tell you that a software company just hired 14 platform engineers.

That's a fact. Maybe even an interesting one.

But if you're a salesperson, what exactly did I just tell you?

That they're growing? Probably.

That they're investing in engineering? Seems reasonable.

That they need your observability platform?

Hold on.

Maybe they're replacing contractors. Maybe they're building a new product. Maybe they're migrating infrastructure. Maybe they're backfilling people who left. Maybe they already have an observability platform they're perfectly happy with.

Or maybe you don't sell observability software in the first place.

The fact is real. The sales opportunity isn't. Not yet.

If you sell...	The question that fact has to answer
Observability	What does this change about operational complexity?
Compliance	Is engineering scaling faster than security or compliance?
Payroll	Does this fact tell us anything useful at all?

A fact only becomes useful when it helps answer a question.

The fact didn't change. What you're trying to learn from it did.

That distinction has consumed an unreasonable amount of our time. It also turned out to be one of the most important parts of Kairo.

The internet is very good at producing facts

Imagine a software company that has:

• Raised $150 million
• Grown to 500 employees
• Launched three new products
• Hired 40 engineers
• Adopted Kubernetes
• Opened several senior positions
• Appeared in a dozen news articles
• Added a bunch of new technologies to its stack
• Had its CEO give an interview about growth

Congratulations. We have researched the hell out of them.

Should you call them?

No idea.

That's the problem.

It is extremely easy to confuse having a lot of information with knowing something useful. And modern sales software gives us all sorts of ways to make that information feel more certain than it really is.

Put enough facts on a dashboard. Add some colorful badges. Calculate a number between 0 and 100.

Now it feels like we've answered the question.

But what question did we answer?

This is why I get uncomfortable when company activity gets flattened into a generic "intent" score.

We cannot read anyone's mind. We don't know that a company is about to buy something. We don't know that a deal will close. Putting a score next to a company doesn't magically make either of those things knowable.

What we can observe is what companies are actually doing. They're hiring people. They're changing teams. They're launching products. They're adopting technology. They're talking publicly about problems and priorities.

The job is figuring out which of those things actually matter to what you sell.

We learned this by getting it wrong

Recently, we've been beating the hell out of the reports Kairo produces for salespeople.

The rules are pretty simple:

• Assume the report is wrong.
• Check every important claim.
• Try to find reasons the supposed opportunity isn't actually an opportunity.
• If the evidence doesn't support the story, kill it.

The first rounds hurt.

One recent report was a good example. We had genuinely useful information: evidence that the company was expanding parts of its sales organization, with responsibilities around pipeline, forecasting, lead handoff, account scoring and revenue operations.

For the right seller, that's interesting.

And what did an earlier version of Kairo decide to lead with?

Funding. Headcount. Company size.

All true.

All almost completely beside the point.

The data was good. The narrative threw it away.

Finding true things isn't enough. Finding lots of true things isn't enough. Even finding a genuinely useful piece of evidence isn't enough if you bury it underneath five facts that merely sound impressive.

So we changed it.

Then we attacked it again.

A plausible story is not enough

This problem gets considerably more dangerous when you add AI.

Modern AI is astonishingly good at explaining things. Give a capable model enough true facts about a company, ask why that company should buy your product, and there's a pretty good chance it gives you an answer that sounds fantastic.

That's precisely the problem.

"Sounds convincing" is an incredibly dangerous standard for sales research. A system can start with individually true facts and still arrive at a conclusion you have absolutely no business putting in front of a salesperson.

So we stopped asking only whether Kairo could produce a good explanation, and started asking a more important question:

Should it be allowed to produce one at all?

Before Kairo tells a salesperson that an opportunity exists, we want a chain we can defend.

Link	What has to be true
1. Observation	Something actually changed.
2. Evidence	We can support that change.
3. Why it matters	It points to a real problem, priority or need.
4. Seller fit	The seller actually solves that kind of problem.
5. Opportunity	It's worth a salesperson's attention.

If any link breaks: no report.

This sounds obvious when you write it down. Building a system that reliably does it is considerably less obvious.

Sometimes the right answer is "no"

We recently ran ten companies through the latest version of this process.

Produced a report

6

Refused

4

PostHog didn't get one. LangChain didn't. PubNub didn't. Mux didn't.

And that was good news.

Not because those are bad companies. Not because nobody could ever sell anything to them. And definitely not because there wasn't enough public information about them.

There was plenty. One of the refused companies alone gave us funding information, customer information, product capabilities, adoption numbers and substantial public usage data.

We could have written something. More importantly, an AI system absolutely could have written something, and it probably would have sounded pretty damn good.

But with the evidence we had, we couldn't support the connection we needed to support. So Kairo refused to produce the opportunity. When that happens in the product, the customer gets their credit back.

From a traditional SaaS perspective, this is a slightly bizarre feature. We built software that occasionally refuses to sell you the thing you just asked it to sell you.

I think that is exactly how it should work.

Because the alternative is expensive bullshit

Salespeople do not suffer from a shortage of companies. There are millions of them.

They don't suffer from a shortage of names, email addresses, phone numbers, LinkedIn profiles, company descriptions, news articles, funding announcements or technologies. There are databases full of those.

What they suffer from is a shortage of time.

A salesperson can already find 500 companies they could contact. That's easy. The valuable question is:

Which of these companies deserves their attention today?

And if we're going to tell someone "spend your time here," then "we found some interesting things about this company" isn't good enough. We have to be willing to say "no" when the evidence isn't there.

Otherwise, all we've done is automate the production of more things for salespeople to investigate.

We have enough of those already.

There is no such thing as a universally good lead

Go back to the 14 platform engineers.

To an observability seller, a fast-growing engineering organization might mean new operational complexity. To a compliance seller, the interesting part might be that engineering is scaling and security isn't scaling with it. To a payroll seller, it might mean nothing at all.

Same company. Same public information. Different seller, different question, different opportunity.

That is why I do not think there is such a thing as a universally "good lead." There are companies that are good opportunities for someone.

The important part is "for someone."

So Kairo has to answer both sides of the question:

Is something actually happening here?

Does it matter to what this particular seller does?

If we can't get to yes on both, we shouldn't be telling a salesperson to spend their time on it.

We also need to show our work

There is one more rule we've become pretty stubborn about.

If we can't trace an important conclusion backward, we don't want it.

If we tell you a company appears to be dealing with a particular problem, you should be able to see why. What happened? What evidence supports it? Where did that evidence come from? Why do we think it matters, and why does it matter specifically to you? Who at the company would plausibly care? And where are we uncertain?

Step	Example
Source	A job posting, with the date it was posted
Observation	Hiring 6 platform engineers
Claim	The platform team is expanding
Opportunity	Why this may matter to this seller

This is not because we expect salespeople to audit a research paper every time they open Kairo. The opposite. The goal is for the answer to be immediately useful.

But there's a huge difference between:

Trust us. The score is 87.

and:

Here's what changed. Here's the evidence. Here's why we think it matters to what you sell. Here's where it came from.

The second one can be challenged.

That's a feature.

Because we're going to get things wrong

We already have. Plenty.

• We've had dates that meant "this is when our crawler first saw the page" presented as though they were dates when something actually happened.
• We've had product capabilities mistaken for problems inside the company selling the product.
• We've had true funding information promoted into bogus urgency.
• We've had good evidence connected to the wrong conclusion.
• We've had reports recommend the wrong person to approach.
• We've had repetitive writing.
• We've had reports where the conclusion sounded better than the evidence deserved.

Those are bugs. Some are ordinary software bugs. Some are data problems. Some are AI problems.

And some are much more embarrassing: we hadn't defined what we meant carefully enough.

Every time we find one, we ask the same question: where did the chain break?

Because "the AI wrote something bad" isn't a useful diagnosis.

Did we collect the wrong information? Did we misunderstand what it said? Did we mistake background information for something that actually changed? Did we connect a real change to a problem it doesn't actually prove? Did we correctly identify a problem but connect it to something the seller doesn't really solve? Did we have a good opportunity and simply explain it badly?

Those are different failures. They need different fixes.

Reality is the source of truth

That sentence sits near the top of our internal architecture documentation, and it has become a useful rule for almost everything else.

Not the model. Not the score. Not our database. Not the categories we invented. Not what we hoped we'd find.

Reality.

That's a much harder standard than "does this output look good?"

And we are not finished. Our latest round of red-teaming still found connections we weren't willing to defend. So we're changing them. Then we'll attack it again. Then customers will find things we missed, and we'll change those too.

That's the process.

The goal isn't to build a system capable of producing a persuasive explanation for every company on the internet. Generative AI made that surprisingly easy.

We're trying to build one that knows when the available evidence actually supports the explanation.

And when it should shut up.

Because public data is abundant.

Evidence isn't.

And your sales team's time is far too expensive to confuse the two.

---

Next: We Tested Whether Our Ranking Engine Can Actually Identify Companies Worth Selling To