How Customer Success Changed in the AI Era (2026)

This is the foundational note for an ongoing series on Customer Success in the AI era. I wrote the May field notes assuming the thesis here was already obvious. It isn't — at least not in the rooms I've been in for the last few months. So this is the longer argument I should have written first.

The core concept of Customer Success has not changed. CS still exists to do what it has always existed to do: make sure the customer actually realizes value from the product.

What has changed is every load-bearing piece underneath that mission. The definition of value. The way value is measured. The way it's priced. The way it's delivered. The ratios. The role. The org chart. The career path. None of those look the way they did three years ago, and the gap is widening fast enough that most CS leaders I talk to are running a 2022 playbook against a 2026 product and wondering why their retention numbers are sliding.

I want to lay the full thesis out in one place, because the series I've been writing — and the playbooks we're shipping on this site — all sit on top of it. If you disagree with the foundation, none of the rest will be persuasive. Good. I'd rather be argued with than ignored.

The one-line version

The mission of CS hasn't changed. The definition of value has. Everything else is downstream of that.

That's the whole thesis. The next 2,000 words are why I believe it, with numbers and sources, and what I think it means for anyone running post-sales right now.

What "value" used to mean

For roughly fifteen years, B2B SaaS CS operated inside a stable definition of value:

Value = adoption. Did users log in? Did they touch the feature set? Did seat utilization grow quarter over quarter?
Proxy for value = renewal. If they renewed, the value was real enough.
Proxy for the proxy = relationship. A trusted CSM, regular QBRs, executive sponsorship — those were the moats against churn.

It was a coherent system. It scaled. It produced the Gainsight playbook, the ChurnZero playbook, the 1:25 CSM-to-account ratios, the health scores weighted on login frequency, the entire CS-as-profession arc that gave us the function we have today.

That system assumed three things about the product:

The product was deterministic — it did the same thing on Tuesday as it did on Monday.
The product was priced per seat — more users = more revenue, full stop.
The product had switching costs — once it was implemented, it was hard to rip out.

All three assumptions are now broken for AI products. Not weakened. Broken.

What "value" actually means in 2026

Every major industry source landed in roughly the same place this year, from different starting points. TSIA's State of Customer Success 2026 calls it "Value Management." Forrester calls it the year AI gets real for service. ChurnZero calls it the agentic CS transition. They're all describing the same shift in slightly different words.

The shift, in plain English: value is now the specific business outcome the customer can attribute to the product, measured continuously, often priced directly, and increasingly delivered by software rather than by a human relationship.

Concretely:

1. Adoption is a noisy signal, sometimes inverted

When the AI agent gets better, the customer's user base often interacts with it less. Fewer turns per conversation. Fewer follow-ups. Fewer escalations. Sessions get shorter.

The 2022 health score reads all of that as a churn warning. The 2026 reality is the opposite — it's product quality compounding.

I have personally argued with my own dashboards on this. A customer's "engagement" graph slides down for three weeks. The health score turns yellow. I dig in. The reason engagement is dropping is that the agent has gotten good enough to resolve things in one shot. That's not a churn signal. That's the win.

Multiply that by every dashboard, every QBR slide, every renewal forecast built on usage as the primary proxy. The signal is broken. Most CS orgs haven't repaired it yet.

2. The contract is moving from seats to outcomes

Sierra prices on resolved customer interactions. Decagon charges per resolution. Adobe announced outcome-based pricing for its CX Enterprise suite. Per Bessemer's pricing playbook and Monetizely's 2026 guide, fewer than 10% of AI companies use outcome-based pricing today — but the trajectory is clear, and the pure agentic category is moving there first.

Why this matters for CS:

The renewal conversation is no longer "did you use it?" It's "did the AI complete the tasks we charged you for, at the quality you needed, and can you attribute the resulting business outcome to us?"
The expansion conversation is no longer "do you want more seats?" It's "you got X% deflection on tier 1 — let's go after tier 2 and re-price the outcome ceiling."
The CSM's day-to-day overlap with Revenue Operations and Finance is now permanent. Outcome attribution is a commercial function, not a relationship function.

The traditional Account Manager / CSM split doesn't survive this. Either the CSM absorbs commercial ownership of the outcome (which is what TSIA's "Value Manager" reframe is pointing at) or the role gets squeezed into something administrative and eventually optional.

3. The retention math has changed underneath everyone

This is the part of the thesis nobody likes saying out loud, but the data is public.

a16z's Retention Is All You Need analysis of hundreds of AI-native companies put the numbers on paper:

AI-native overall: ~40% gross revenue retention. ~48% net revenue retention.
B2B SaaS median: ~82% NRR.

At the high end of pricing (above ~$250/mo), AI-native NRR climbs to roughly 85%. At the low end (under $50/mo), GRR collapses to 23%. Cassie Young at Primary Venture Partners has been calling the structural version of this the "gross retention apocalypse" — and the argument isn't cyclical, it's structural. Switching costs are lower than in any software category that came before. Prompts are portable. Models are commoditizing. The customer can leave on a Tuesday.

You cannot run a traditional CSM-led, per-seat-priced, high-touch CS motion when your category-wide NRR is 48%. The unit economics don't work.

That isn't a slogan. It's a budget constraint. If your blended NRR is half of SaaS norms, the cost of every human touchpoint has to be earned twice. The AI-first companies have responded in three ways at once, and they are doing all three simultaneously:

(a) Push ACVs up so the per-account economics support a high-touch motion.
(b) Build genuine switching cost through deep operational integration — which is what Forward Deployed Engineers are for.
(c) Tie revenue directly to outcomes so you only collect when value is delivered.

Anthropic, OpenAI, Sierra, Decagon, Palantir, Glean, Cresta — every one of them is operating some combination of (a), (b), and (c). None of them is running a 1:25 CSM motion on per-seat contracts and hoping NPS holds it together.

4. The role is bifurcating, not just renaming

The "CSM became a Deployment Strategist" framing is too clean. What's actually happening is bifurcation. The post-sales function is splitting into two halves that used to be one:

The commercial-and-outcome half — value management, outcome attribution, renewal commercials, expansion business cases. TSIA's "Value Manager."
The technical-implementation half — embedded engineering, eval design, integration depth, switching-cost construction. Palantir's FDE/DS model, copied by everyone.

On the technical side, the numbers are loud. Forward Deployed Engineer job postings rose more than 800% from January to September 2025. Average US comp: $238k. Staff-level: $630k+. Anthropic's Applied AI team, OpenAI's DeployCo subsidiary, Salesforce's Agentforce Contact Center 100 program, Dust's Founding AI Deployment Strategist, Post-Sales, Mistral's Deployment Strategist (Cybersecurity) — all hiring the same shape of role. EY launched a UK & Ireland FDE practice in April 2026, which is the first major consultancy to formally adopt the model.

The career trajectory, the comp, the equity, the proximity to interesting problems — all of it is concentrating in the FDE/DS lane.

On the commercial side, the change is quieter but just as real. CSM-to-account ratios have moved from roughly 1:25 in 2022 to 1:60 in 2026 (ChurnZero). Seventy percent of CS managers are now using GenAI to analyze sentiment across portfolios. AI is running the administrative load — summaries, drafts, sentiment, early-warning triggers — and the human is reserved for judgment.

The CSMs who thrive in 2026 either move toward the FDE/DS pole (technical depth, embedded work, outcome ownership) or toward the Value Manager pole (commercial fluency, outcome attribution, portfolio orchestration of AI-driven workflows). The middle — the pure-relationship CSM running 25 accounts on intuition and quarterly check-ins — is the squeezed position.

5. The product itself broke the playbook

Even if you ignore everything above and only look at the product, the playbook still breaks. AI products are non-deterministic. That single fact creates a class of CS problems that the existing function doesn't have chapters for.

Quality regressions are CS tickets. A model updates. A prompt changes. A customer's success rate drops four percentage points on a Tuesday. They open a ticket. Whose ticket is it? Engineering's? Product's? CS's? The honest answer at most companies right now is we improvised one. I don't think anyone has it figured out cleanly.
Go-live is a passing score on an eval, not a date. The customer doesn't know how to define the eval. So somebody at the vendor sits with the customer and co-designs it. That work used to be a Product question. Now it's a post-sales question that arrives in week one of deployment and never really ends.
"Health" is multidimensional. Pure product-telemetry health models hit ~55–65% churn prediction accuracy. Multi-signal models that include conversation sentiment, support tickets, billing signals, and outcome metrics hit ~78–85% (ChurnZero). The cost of building the second kind of model is high enough that most CS orgs are still operating the first kind and explaining the misses after the fact.
The "smiling curve." a16z documented this at multiple AI-native companies: churned customers returning as the product improves. Retention is product-trajectory-driven, not relationship-driven. The CSM's job is increasingly to keep the customer in the room long enough for the product's trajectory to do the renewal work.

The cumulative effect of these four product realities is that the post-sales function is increasingly responsible for things it doesn't fully control. Model behavior. Eval definition. Quality drift. Outcome attribution. That's an uncomfortable place to operate from, and it's where a lot of CS leaders are right now.

What that means for what we measure

If the definition of value has shifted, the metrics have to shift with it. Roughly:

Old (2022)	New (2026)
Adoption % / DAU	Outcome volume & quality (the thing being billed)
Health score weighted on usage	Multi-signal health score weighted on outcome trajectory
Raw retention	M3-rebased cohort retention (filter out "AI tourists")
QBR readiness	Continuous business review the customer can self-serve
Renewal forecast	Switching-cost depth + product trajectory + outcome attribution
NPS	Outcome NPS — would you recommend the result, not the vendor

I am not claiming all of the right side is solved. Half of it is improvised at most companies. But the direction is clear, and the orgs still running purely on the left column are going to find their CS metrics increasingly disconnected from their renewal reality.

What it means for the people doing the job

A few things, in order of how directly they affect a working career:

1. Commercial fluency stopped being optional. If you are running CS and you can't read a contract, model unit economics, or build an outcome attribution case, you're below the new bar. That used to be a senior-CSM nice-to-have. It's now a floor.

2. Technical fluency is the new senior-track skill. Not engineering. Fluency. Can you read an eval rubric? Can you reason about why a model regressed? Can you co-design a deployment plan with a customer's tech team and hold your own in the room? The CSMs who can do this are getting paid like FDEs. The ones who can't are getting paid like CSMs.

3. The middle is the risk position. Pure-relationship CSMs at AI-first companies will get squeezed. Pure-relationship CSMs at traditional SaaS companies whose products are turning into AI products underneath them get squeezed more slowly, in the same direction. The shift is happening under the contract, not at the contract boundary.

4. The bridge from CSM to Deployment Strategist exists but it's narrow. I get asked this constantly. The honest answer: yes, the bridge exists, and the people who cross it are people who can hold both halves — commercial ownership and technical depth — credibly enough to be embedded with a customer's engineering team. That's a small population. If you're in it, the next 18 months are going to be unusual for you.

What I'm still figuring out

I want to be honest about what I don't yet have a clean answer for. Some of these will be future posts in this series. Some are open questions I'll be writing into for the next year.

What does the right post-sales staffing model look like below roughly $1M ACV? Palantir-style 1:1 embedding works at very high contract sizes. The math for mid-market AI is wide open.
Where does eval co-design with the customer actually live, organizationally? Today it's CS-adjacent at most AI-first companies but nobody owns it cleanly.
How do outcome-based contracts get renewed when the AI's capability shifts every quarter? Sierra, Decagon, and Adobe are working this out in production right now.
How do you keep a health score honest when the product itself is non-stationary? This is an open ML problem dressed up as a CS problem.
Does "Customer Success" come back as a separate function once AI-first companies mature past the early-deployment phase, or does the FDE/DS model become the permanent shape of post-sales at this category?

I'll write into all five over the coming weeks. I don't pretend to know the answers yet. I'd rather be early and partially wrong than late and confidently wrong.

What I'd love pushback on

If you run post-sales at an AI-first company — Anthropic, OpenAI, Sierra, Decagon, Cresta, Glean, Dust, Mistral, or any of the others — the view from inside is the one I'd most like to correct against. Two specific things I want to be wrong about, if I am:

Am I overstating the squeeze on the traditional CSM role? Are there AI-first orgs where the function is alive and well under the original title?
Do the a16z retention numbers match your operational reality? Or is the public data lagging where the curve actually is right now?

My LinkedIn DMs are open. Argue with this. Especially if you think the foundation is wrong.

What's next in this series

This is the foundational post. The next six pieces go deeper on specific pieces of the shift. Brief teasers, in the order I plan to publish them:

From CSM to Deployment Strategist — the actual role transition. What the FDE/DS model looks like from the inside, what skills bridge the gap, and where the career arc lands.
Is Agentic AI Actually Delivering Value? How Buyers Are Choosing — the buyer-side view. How procurement, ops, and CFOs are evaluating agentic products in 2026, and what that means for how vendors position post-sales.
The Hybrid Play: Traditional SaaS Bringing Agentic Products — what happens when a legacy SaaS company shifts an AI product line into its existing CS function. Salesforce Agentforce, Intercom Fin, ServiceNow — the canonical hybrid cases, and the org-design traps to avoid.
Customer Lifecycle at AI-First Companies — stage-by-stage walkthrough: handoff, kickoff, eval co-design, go-live, expansion, renewal. What each stage looks like when the product is non-deterministic and the contract is outcome-priced.
The Hybrid Expansion Playbook v1.0 — the first downloadable artifact in this series. How to run expansion in a hybrid CSM + FDE motion without colliding the roles.
The End-to-End Deployment Playbook (Synthesis) — the full operator-seat playbook tying everything above together. The piece I want to exist as the reference document for the category.

If you want each piece as it goes up — and the downloadable artifacts alongside them — subscribe to the newsletter. And if you're operating in this space and have something to say, I'd rather hear from you than write into the void.

How Customer Success Changed in the AI Era