AI will increasingly handle reporting, routine follow-ups, account monitoring and basic product guidance.
The human value of Customer Success moves somewhere more demanding: understanding the customer’s industry well enough to help produce a specific business outcome.
In our world, adoption cannot mean the same thing for a medical-device manufacturer, a financial institution and an industrial company. Their workflows, risks, approval structures and definitions of success are fundamentally different.
The future CSM is therefore not simply a product expert. They understand the customer’s operating environment and can connect the technology to measurable outcomes inside it.
If the same Customer Success playbook works for every industry, it probably goes deep enough for none of them.
Klarna learned this the hard way. In early 2024, the company said its AI assistant was doing the work of 700 customer service agents. A year later, its CEO told Bloomberg: 'As cost unfortunately seems to have been a too predominant evaluation factor when organizing this, what you end up having is lower quality.' His conclusion was just as direct: 'Really investing in the quality of the human support is the way of the future for us.' Klarna didn't give up on AI. It found out that doing the same job with fewer people is not the same thing as doing a better job.
Service leaders are drawing a similar line. In a Gartner survey published in April, only 31% had implemented or planned frontline layoffs in response to AI through early 2027. Most were shrinking their teams slowly, through attrition, while 85% were adding new tasks and responsibilities to their agents' roles. As Gartner analyst Eric Keller put it, 'Organizations that only use AI to reduce costs risk missing a strategic opportunity.'
For most, the money is not paying off yet. Gartner also found that service and support leaders put a bigger share of their budget into AI than any of the other business functions it assessed. Only about one in four can show a return on it:
AI takes the routine work. The savings only show up when someone decides what the people should do next.
Marketing generates the interest, sales understands the problem, implementation learns the workflow, and Customer Success discovers what adoption actually requires. By then, each team stores a different part of the truth.
AI weakens the original justification for this relay race. One person can now operate across a broader part of the journey, supported by shared context and specialists who join when their expertise is needed. Specialization should happen around the customer, without moving the customer from one organizational box to another: a specialist can join the relationship without inheriting it. Expertise should be transferred, not ownership.
Services companies run the same relay race, from sales to the project manager to the delivery team, and the customer's context gets thinner at every step. In our industry, where companies publish regulated content in many languages, the same thing happens to the content itself: when the approved wording, the terminology choice or the risk warning stays with the team that made it, every language downstream inherits the gap.
HubSpot described the same leak inside a single team on September 16, the opening day of UNBOUND. In its Fall 2026 product release, it said sales reps spend more time than ever on admin and less time selling:
If context leaks inside one team, imagine what happens between four. In April 2025, Cursor, the AI code editor, showed what that looks like from the customer's side.
The bot wasn't lying on purpose. It didn't have the part of the truth the rest of the company had, so it filled the gap with a policy that didn't exist.
Writing emails faster, summarizing meetings and generating content are useful productivity gains. They do not constitute transformation.
Builders go further. They turn knowledge into repeatable workflows. They decide what the AI can do, what context it needs, where human judgment enters and how quality is measured.
A clever prompt disappears with the person who wrote it. A well-designed system becomes organizational capability.
The important divide will not be between AI users and non-users. It will be between companies that automate old work and companies that invent a better way of working.
The research is blunt on this. MIT's NANDA initiative found in 2025 that tools like ChatGPT and Copilot are now widely adopted, 'but these tools primarily enhance individual productivity, not P&L performance.'
The few companies that get past this point work differently. In McKinsey's State of AI survey, published in August, nearly nine in ten respondents say their organization uses AI regularly in at least one function, yet only 37% report at least some impact on company-wide operating profit, about the same share as last year. Nearly three in four of the 6% who see significant value have fundamentally redesigned their workflows, against one in four of everyone else.
Microsoft's 2026 Work Trend Index points the same way when it looks at which factors are most closely linked to workers reporting real impact from AI:
Services companies face the same divide, and our own industry shows it clearly. For a services firm, automating the old work means selling the same hours for less. The 2026 Nimdzi 100 sees many language providers 'shifting their services from basic translation execution to comprehensive workflow orchestration, regulatory consulting, and AI data management.' The value is moving from producing words to content intelligence: knowing what the content means, where the risk is and when an expert needs to check it.
The model is rented. The harness has to be ours.
That harness is the brain surrounding the model: customer history, industry knowledge, terminology, brand standards, commercial judgment, risk policies, quality criteria and escalation rules.
Without it, AI scales averages. With it, AI can scale expertise.
This matters enormously in the language industry. Translation output is becoming easier to generate. The defensible value is knowing what the content means, where the risk lies, which industry rules apply and when human expertise is required.
When AI produces something generic, blaming the model is often easier than admitting that the organization never captured what it knows.
Every company now rents the same AI models. Think of the model as a brilliant new hire on day one, fast and articulate, but knowing nothing about your clients, your approved wording, your rules or your standards. The harness is everything you give that new hire so the work sounds like your company and holds up to scrutiny. Skip it, and you get generic output at high speed. That's what people call AI slop.
In May 2025, the Chicago Sun-Times printed a syndicated summer guide that recommended books that don't exist. The section came from King Features, a unit of Hearst Newspapers, and was written by a freelancer who had used AI. As the paper's own newsroom union, the Sun-Times Guild, put it, this was 'a syndicated section produced externally without the knowledge of members of our newsroom.' The paper removed the section from its e-paper and said print subscribers would not be charged for that edition, and King Features said it was 'terminating our relationship' with the writer. The writer later told NPR: 'They trust that the content they purchase is accurate and I betrayed that trust.'
The model did what models do. What was missing was the company around it: the review rules, the quality criteria, and the moment when an expert checks the work before the reader does. For a services firm, that's the whole business: clients pay for the judgment around the output, not for the output itself.
In our industry, a translated drug label or financial disclosure is exactly that kind of deliverable, and clients are learning the same lesson. The 2026 Nimdzi 100 expects the language industry to reach USD 73.4 billion in 2026, barely up from USD 72.6 billion in 2025, while some clients now expect cost cuts of up to 75% through AI. It also reports that some buyers tried to skip human translation entirely, using raw AI output even for content once judged unfit for machine translation, and that this AI-only approach 'has frequently resulted in high-profile quality failures and hallucinations, leading clients to return to language providers for secure, expert-in-the-loop validation.'