Notes

Two weeks ago we wrote that B2B contact data decays about 22.5% a year. Then we went looking for where that number actually comes from.

What we found was not a study. It was a citation chain.

Published decay estimates run from roughly 12% a year to 30%, and as high as 70% for individual fields like direct dials. Almost every one of them traces back to a company that sells contact data.

That does not make them wrong. Vendors sit on the largest records of churn in the market, so they are the ones who would know. But a number with that much spread, published by the party who benefits from you believing it, is not a planning input.

And even if 22.5% were exactly right, it would tell you nothing you could act on. It describes a market. You do not sell to a market average. You sell to 140 named accounts and 300 named people, and the only honest way to know how stale that list is, is to measure that list.

So we stopped quoting the average and built the measurement instead.

Context Lab is a ten business day audit of the context your AI agents are about to run on. Five dimensions, weighted to 100 points, one score, one verdict.

Two of them are the data checks you would expect. Are the accounts still a fit. Are the people still in the roles you recorded.

The other three are the ones that surprise people. Do the buying triggers in your ICP actually appear ahead of your recent closed-won deals. Have your reply and meeting rates moved by sequence. Are your personas the titles that actually showed up in the deals you won.

That is the difference between clean data and the right data. A perfectly verified record of the wrong account is still the wrong account.

Four bands come out of it. Fresh means safe to automate. Aging means fix the flagged dimensions first, usually two to three weeks rather than a rebuild. Stale means rebuild the perishable layer before you scale outreach. Expired means do not put an agent on this at all.

$3,500 fixed, credited in full toward a Phase One assessment if you go ahead within 30 days. About two hours of your team's time, because you should not have to run a project to find out whether you have a problem.

And if you score Fresh, we will tell you so, and you will have spent $3,500 learning you were right. That happens. It is supposed to.

There is a free version as well. Context Check: 25 accounts, account and contact drift only, one page back to you. Small on purpose.

We are not going to tell you your data is 22.5% stale. We are going to tell you how stale yours is.

If your ICP was built two years ago, it may not describe your market anymore. It is still describing the market you had then.

An ICP has two layers. The durable core, meaning the problem you solve, who pays, why you win, can hold for years. The perishable layer, meaning contacts, buying committees, trigger signals, messaging and channels, starts decaying the day you finish it.

That used to be a manageable problem. A human SDR works around a stale ICP with judgment. They notice the account no longer fits, the contact has left, the pitch did not land.

An AI agent does not notice. It runs whatever context you give it, at scale.

With a person, a stale ICP means some wasted dials. With an agent, it burns through your market, your sending domains and your reputation in weeks.

So before you put an agent on your pipeline, ask one question: is the ICP it runs on still true?

Agents are only as good as the context they run on. When did your team last check its ICP against the deals you actually won?

Gartner asked sales organizations how many times they had torn down and rebuilt their key account program in the last seven years because it was underperforming. 79% had done it at least once.

Rebuilt. Not tuned. Demolished and stood back up, inside seven years, by the same companies who say they expect their key accounts to grow faster than anything else they own.

This is not a neglected corner of the business. It is the part carrying the highest expectations in the whole revenue org, and it is the part that gets rebuilt most often.

When something is rebuilt that many times and still underperforms, the thing being rebuilt is probably not the problem.

Almost every one of those rebuilds touches the same four things. Coverage model. Account tiering. Comp plan. The account plan template, always the account plan template.

Almost none of them touch this: the account manager is the only seller in the building asked to find opportunity without being given any signal.

Look at how the two halves of a revenue org are actually equipped. The rep working net new gets intent data, enrichment, technographics, a scoring model, and now an agent that researches strangers before the first call. An entire industry exists to tell that rep which unfamiliar company is worth an hour.

The account manager, who owns the accounts the company expects the most growth from, gets a renewal date, a QBR deck and a relationship.

And 66% of key account managers tell Gartner that internal coordination alone limits their ability to strategize for account growth. The people closest to your revenue are spending their strategic capacity navigating their own company.

The uncomfortable part is what is already sitting there unread. For every account you already have, you own the richest dataset you will ever have on that buyer. What they bought and in what order. What they evaluated and passed on. Which product lines went quiet. Who left and who arrived. How this account behaved the last three times it expanded.

It is in the CRM, the ticketing system and the billing platform, and it is read as history rather than as signal.

So we buy data about companies we do not know, and leave unread the data about the ones we do.

A test you can run this week. Pick three of your largest accounts and ask the team what each one did in the last 90 days that it did not do in the 90 days before. Then ask where the answer came from. If it came out of a QBR deck rather than out of your own data, you have found the gap.

It is not a coverage gap and it is not a comp gap, which is why seven years of rebuilding has not closed it. Most companies do not have an account management problem. They have an account intelligence problem, and they keep trying to solve it by moving boxes on an org chart.

The most expensive four words in go-to-market work are "we'll validate in market." Because that means later.

The plan gets built. It gets socialized. Headcount gets requested against it. Somewhere around month four it finally meets an actual buyer, and by then reversing it costs somebody their credibility. So it does not get reversed. It gets defended.

We rebuilt our Phase One around that single failure mode. It now runs as four labs.

The first three produce a hypothesis. ICP Lab: who actually buys, read out of your own closed-won, closed-lost and churn history rather than demographic convenience. Message Lab: what makes them care, built per segment. Channel Lab: how you reach enough of them at your contract value. Standard work, done properly.

The fourth lab is the one that changes the economics.

Signal Lab puts the message in front of real buyers inside the engagement, before anyone has committed budget to it. Real accounts inside the defined segments. Not a survey, not a panel, not a focus group. Buyers who could actually purchase what you sell.

Three things come out of it.

• The segments that stayed silent. Silence is a finding, and it is the expensive one to discover in month nine instead of week five.
• The objections that recur. Not the ones you predicted in the workshop. The ones that actually showed up, in the buyer's words.
• Meetings. The test produces conversations, which means the engagement is built to end with pipeline already in motion rather than a document and an assumption.

The uncomfortable implication is deliberate. If the test says the hypothesis was wrong, we found that out on our time, inside a fixed fee, before you spent a quarter of headcount proving it the expensive way.

A strategy that has never met the market is a hypothesis with a price tag.

For decades the mid-market competed in an uneven arena. Outspent, out-staffed, overshadowed by companies whose main advantage was simply being larger. AI was supposed to fix that. The data says it has not. Not yet.

McKinsey put the question to 1,719 leaders across 97 countries this summer. Among organizations above $1 billion in revenue, 54% are now scaling AI across the enterprise. Among smaller organizations, 33%. The US Census Bureau's business survey points the same way: 37% AI use at firms with 250 or more employees, under 20% at the smallest.

The big companies are pulling ahead.

That is not the story most people are telling right now, and it matters, because the structural advantage genuinely does sit with the mid-market. Enterprises have more money, more people and more infrastructure. They also have more friction, and friction is now the expensive part.

A mid-market company can approve in one meeting what an enterprise routes through four committees and a steering group. It can be in production in 90 days while the enterprise is still drafting evaluation criteria. If you ship faster than they evaluate, you win.

The advantage is real. It just does not collect itself.

Four rules worth holding any mid-market AI initiative to.

• One decision maker per initiative. Shared ownership is how a 90-day project becomes a 9-month one.
• The 90-day production rule. If it cannot reach production in a quarter, the scope is wrong.
• Buy first, build only what is strategic. Custom work is for the thing competitors cannot copy. Everything else is a purchase.
• Tight feedback loops. Measure weekly, kill fast. Learning rate is the only durable advantage.

And the honest caveat, because an equalizer is not a total leveler. Enterprises keep real advantages: reinvestment loops, intelligence infrastructure, the sheer scale of experimentation. The twist is that they frequently move too slowly to use them.

The mid-market has always been underestimated. The tooling is finally on its side. The decision speed has to be too.

If you are in that 33%, what actually stopped the last initiative? Budget, or the number of people who had to say yes?