The Value Delivery Layer
Findings are about to be cheap. The scarce thing was never the finding.
Findings are about to be cheap. The scarce thing was never the finding.
Say you run the correlation on an account and it turns up something good. The customer, a payments company, has six million dollars in next year's infrastructure budget to expand capacity they think they've outgrown. Your data says capacity isn't the constraint. The load concentrates in four queries that run on every page render, and two of the three systems on the expansion list aren't the ones under pressure.
Now send it to your main contact. They thank you warmly. Nothing happens. The expansion goes through unchanged, the renewal comes in flat, and nobody inside the customer ever repeats the finding to anyone else inside the customer.
Nothing in that sequence needed the analysis to be wrong. It's the ordinary outcome, and it's the one the analysis has no opinion about.
There are three things that have to happen between a customer's data and a customer's decision.
Detection. What happened. Sessions, transactions, tickets, API calls, the events themselves. This has been automated for a decade, and nobody has been paid well for it in years.
Interpretation. What it means. The correlation between what a customer says they want, what they're spending on, and what they already own. This used to be the expensive part, because it took a person sitting with three unrelated sources for an afternoon per account. It's being automated right now, and the automation is good. I've written about how to do it. Within a couple of years, doing it by hand will look like formatting a deck by hand.
Delivery. Somebody inside the customer with the authority to act changes what they do. Not learns something. Changes something.
The cost of the first two is collapsing. Nothing that collapsed them touches the third.
The industry treats the third step as packaging on the second. It shows up in job descriptions as "communication skills" and in team meetings as "how do we position this." I think that framing is backwards, and it's about to get expensive. Detection and interpretation produce a claim. Only delivery produces value, and delivery is a different kind of work from the two steps in front of it.
The usual defense of the human in this loop is that people want to talk to people, that trust is built face to face, that empathy doesn't scale. I don't think any of that holds up. Customers will happily take an accurate machine-generated answer over a warm human one, and they already do everywhere else in their working lives.
The reasons the delivery layer resists automation are structural, and there are three of them.
A claim about someone else's business needs someone who can be wrong about it. When you tell a VP that six million dollars of next year's budget is aimed at the wrong cause, you're making an assertion about a business you don't work in, to a person who knows it better than you do. The reason they can act on it is that you're standing behind it. If you're wrong, you carry that into every conversation you have with them afterwards. That exposure is exactly what makes the claim worth listening to. A finding with nobody behind it is a suggestion, and executives get dozens of suggestions a week.
The decisive facts are the ones nobody wrote down. Which sponsor is on the way out. Which initiative the CFO already killed once. Which two directors can't be in the same recommendation. Whose budget the finding threatens. None of this is unstructured data waiting for better retrieval. It's deliberately unrecorded, it gets told to a person rather than to a company, and it gets told to that person because of who they've been to the customer over eighteen months. There's no context window that contains a fact nobody ever typed.
Value isn't discovered. It's attributed, and attribution is contested. This is the one people miss. Any claim that your platform produced an outcome is also a claim about which internal team gets credit for it, and there's always a team that was planning to claim it. Saying your platform caught a regression before it reached customers can mean the incident-response program a director has spent two years defending is finally paying off. Or it can mean the release process the same director owns has been shipping regressions for two years. Which one depends on nothing but who's in the room. The finding doesn't decide that. A person does.
A finding isn't a fact you deliver. It's an argument about who inside the customer gets to be right.
Take a different account in a different industry, because none of this is about the vertical. It's fictional.
Ravenna Markets, a grocery chain, two hundred and forty stores. Nine million dollars committed to an outside consultancy on the premise that their demand forecasts are wrong. Your data says the forecasts are accurate to within two points. The loss sits almost entirely in a manual approval step that holds replenishment orders for forty-eight hours in about ninety of the stores.
Send it to the Director of Demand Planning, your main contact and a friendly face for three years, and the finding dies. It has to. That director commissioned the consultancy and has spent two quarters telling the executive team that forecast accuracy is the problem. To act on this they'd have to walk into a steering meeting and correct a diagnosis they authored, in front of the people who funded it. You've handed a friendly person a fact that costs them something to repeat. They'll thank you for it and put it down.
Send it to the COO, framed as a question about the approval step rather than a result from your platform, and the same finding becomes nine million dollars they get to not spend and a process they already control. They get to be the one who noticed. Your platform is the evidence in their argument, not the subject of yours.
Identical content. One version is worth nothing and costs a little goodwill. The other keeps nine million dollars in the building. Everything that separated them happened after the analysis was finished, and none of it is written anywhere in the data.
Here's the part that surprised me when I started running this work at volume.
When a person did the interpretation by hand, you got two or three findings a quarter, for the three accounts you were already worried about. Scarcity did the selecting for you. You carried what you had.
When the interpretation is automated, you get forty findings per account, all plausible, most of them true, and the constraint moves. It moves onto the question of which two you're willing to spend your relationship on. Because you do spend it, and the balance is finite. A CSM who forwards every finding the system generates becomes a distribution channel for content the customer stops opening. The fourth unread insight email is what makes the fifth one worthless, including the one that was worth six million dollars.
So the delivery layer is mostly a filter, and the filtering is the expertise. Of forty true things, thirty-eight are true and inert. Knowing which two are live this quarter, for this sponsor, given what happened on the last call, is a judgment the analysis can't make, because the information it depends on was never captured.
The tooling makes this worse before it makes it better. Volume of insight feels like productivity, and it reports beautifully.
Three, in the same spirit as the rules for the analysis itself. All of them narrow the output.
Route the finding to whoever it makes powerful. Before it goes anywhere, name the person whose standing goes up when this is true, and the person whose goes down. If those are the same person, the finding isn't ready. If the second person is your main contact, you've got a routing problem to solve before you've got a value story to tell.
If your champion can't repeat it in one sentence in a meeting you're not in, it doesn't exist. This is the only real test of the delivery layer, and it's brutal about decks. Almost every value story that dies in the field dies because it needed its author in the room. Write for the version of the sentence your sponsor says out loud, to a peer, six weeks later, from memory.
Whoever's name is on the claim has to be able to defend it live, unaided. If you can't take the follow-up question without opening a laptop, you can't send the slide. This is the hard limit on how much of the work you can hand to a machine, and it's the same rule as never freelancing on price. The system can build the argument. It can't be the party accountable for it, and the customer isn't agreeing with an argument. They're agreeing with a person who made one.
If delivery is the human layer, then most of what customer success currently measures is a measurement of the two layers being automated. Touchpoints logged, reviews held, emails sent, decks produced. Those numbers describe throughput on work that's about to cost nothing to produce, which is why they'll look great right as the function gets less valuable.
The things only the delivery layer produces are measurable. They're just harder to count. Objectives confirmed in the customer's own words and written down where the next person can find them. Sponsor changes survived without restarting. Decisions that went differently, named, with the person who made them. An infrastructure expansion redirected, a consultancy cancelled, a hiring plan cut in half. None of those are activity metrics, and all of them are the real output.
The practical consequence for teams: the leverage from AI in customer success isn't fewer CSMs covering the same accounts. It's the same CSMs spending an afternoon a week on the two conversations that decide the year, instead of the four days they currently spend assembling material for conversations that decide nothing.
I want to be honest about the boundary of this argument, because "the human is always needed" is the kind of claim that gets made most loudly by the humans in question.
The constraint I'm describing isn't capability. It's plausible that a model will compose a better message to that VP of Operations than I would, and it's plausible it already can. The constraint is accountability and access. Delivery needs a party who bears consequences for being wrong, and it needs information that people only give to someone they've decided to give it to.
Both of those are facts about how companies work today, not laws of nature. If enterprises start accepting claims about their own business underwritten by a vendor's system with nobody standing behind them, and if executives start telling that system the things they won't put in writing, then this layer compresses too and I'll have been wrong about the timeline. I don't think that happens soon, because both changes ask a company to accept a new kind of risk in exchange for convenience it doesn't particularly need.
Until then, the shape of the job is clear enough. The analysis is becoming free. The part where a finding walks into a room, attaches itself to someone's interests, and survives being repeated without you isn't, and it was never the part we were measuring.
If you try one thing from this, take the last finding you sent to a customer that went nowhere. Don't re-examine whether it was correct. Ask who it made powerful, and whether that was the person who received it. The answer is usually in the routing, and it's usually not close.
If you run that exercise and want to compare notes, I'm happy to.