Handing consequential decisions to something you cannot fully watch is the oldest management problem our species has. Agentic artificial intelligence is a new occupant of a very old seat.

A customer service line is an odd place to look for the frontier of artificial intelligence, and that is exactly where it is landing first. McKinsey reports that 41 percent of AI deployments in customer-facing functions have fully scaled, roughly 3.5 times the rate seen elsewhere in the enterprise (Rodriguez et al.). The more useful question is why this corner of the business, rather than finance or supply chain, is where autonomous agents cross from pilot to production. The answer runs deeper than plumbing.

Why customer experience went first

There are two clean explanations, and they are worth keeping apart.

The first is structural, and it is the one McKinsey names. Over the past decade, customer experience leaders built the exact foundations an agent needs: observable workflows, explicit decision rights, structured data, and established human-in-the-loop models (Rodriguez et al.). On this reading, agents scaled in CX because CX already had the wiring.

The second explanation is behavioral. Customer service is the domain where a mistake is usually cheap to reverse and a damaged relationship can be repaired. The decisions are high-volume, repeatable, and low-stakes in any single instance. That profile matters because the machinery humans use to extend trust — reputation and the ability to correct a bad call after the fact — works best precisely where errors are recoverable and interactions repeat.

These are not competing claims so much as two layers of the same one. The scaling gap is real and reported. The infrastructure is the proximate mechanism, the thing that made CX ready, while the recoverability of a customer interaction is the deeper reason CX went first. A test would discriminate between them. If agentic AI scales just as fast in functions with equally mature data but irreversible stakes — clinical decisions or binding financial commitments — then infrastructure alone explains the ranking and the trust argument adds nothing. If those functions lag despite good data, the recoverability story is doing real work.

The delegating animal

Humans are the animal that routinely hands consequential decisions to agents it cannot fully watch. Scouts sent ahead of the band, kin trusted with the harvest, hired hands, guild members, and eventually whole institutions. Every one of these arrangements is a version of a single problem: how do you let someone act on your behalf without being cheated, misrepresented, or quietly betrayed. Economists later named it the principal-agent problem, but it long predates economics.

The solutions our species settled on are reputation, reciprocity, and correction. Cooperation holds up among self-interested parties when they interact repeatedly and the future casts a long enough shadow to make betrayal expensive (Axelrod). At the scale of a community, that logic runs on reputation — the shared memory of who dealt fairly and who did not (Alexander). An AI agent resolving a billing dispute or recommending a plan is a new occupant of that very old seat. The design questions leaders are now asking — what the agent may decide alone, when it must escalate, and how a wrong call gets undone — are the same questions a band asks about a scout. The vocabulary is new. The trust math is not.

The same coalitional instincts that make humans attentive to loyalty and defection inside groups also shape how customers respond when an agent acts on their behalf. For the deeper evolutionary machinery behind those instincts, see The Coalitional Brain: Why Humans Default to “Us vs. Them”.

The constraint that just moved

Here is the part worth sitting with. A firm could never hold a genuine relationship with millions of customers, because attention and headcount do not scale that way. Relationship-tracking is cognitively bounded; there is a ceiling on how many real ties any individual can carry, and gossip and reputation evolved in part to stretch that ceiling as far as it goes (Dunbar). So firms substituted the static journey — a script, a decision tree, a rule of thumb — for the relationship they could not actually maintain.

Agentic orchestration is the first tool that lets an institution behave, at scale, as if it remembered each customer. McKinsey’s operational name for this is letting context travel to every decision point, backed by shared identity and a unified customer record (Rodriguez et al.). That is memory, rebuilt as infrastructure. It is a real lift, and it comes with a caution that deep history makes obvious: lifting the constraint scales the rules of trust, it does not repeal them. The same machinery now reaches millions of interactions, including the ways it fails.

Key terms

• Agentic AI: software that pursues a defined goal by interpreting context and taking actions across systems, rather than following a fixed script.
• Principal-agent problem: the risk that someone acting on your behalf will serve their own interest instead of yours when you cannot fully observe them.
• Decisions in motion: the live choices that determine what a customer actually experiences — what offer, what policy, what help, when a human steps in — as opposed to the mapped journey.
• First-contact resolution (FCR): the share of issues resolved in the first interaction, a standard proxy for both cost and trust.
• Network reciprocity: cooperation that holds because acts of good faith are linked across a connected network rather than settled one exchange at a time (Nowak).

Governing decisions in motion

The operational shift McKinsey describes is a move from designing journeys to governing decisions in motion, and it arrives in three horizons of widening authority (Rodriguez et al.).

In the first, a single well-defined workflow runs end to end under strict guardrails. This is where most production deployments sit today, and the numbers are concrete. A United Kingdom energy retailer put a natural-voice agent on high-volume, low-complexity calls and reported a 6 percent lift in customer satisfaction and an annual run-rate impact above ten million dollars. A premium automotive manufacturer used an agent to augment human reps and saw first-contact resolution rise 24 percent with a 30 percent productivity gain (Rodriguez et al.).

In the second, agents coordinate multiple workflows inside a domain so that handoffs carry context and the journey holds together. A large European telecommunications provider connected signals across channels to decide not only which offer to make but when and with what supporting action, and reported roughly forty million euros in margin impact (Rodriguez et al.). This is where the older logic of network reciprocity becomes practical: value compounds when good decisions are linked across a connected system rather than optimized one exchange at a time (Nowak).

In the third, still ahead of any organization operating fully today, agents share memory and objectives across functions and partners so the whole life cycle is optimized against common goals (Rodriguez et al.).

Underneath all three horizons, the governance moves are the modern version of how a coalition trusts a member it sends out on its behalf. Four of them carry most of the weight.

Decide what each agent is optimizing for, and which tradeoffs it may make, before it goes live. An agent handed a vague goal will optimize the wrong thing efficiently. Name the objectives — retention, trust, cost to serve, time to value — and make the tradeoffs explicit.

Keep consequential actions reversible, and keep a human hand on the ones that are not. Trust is asymmetric. It accumulates slowly through many correct, observable, correctable acts and collapses quickly through a few irreversible wrong ones. Concentrate autonomy where a mistake is cheap to undo, and keep judgment and empathy human where an error is not.

Let context travel, but treat the shared record as the asset it is. Without unified identity and record, an agent acts on partial snapshots, and the main thing that scales is fragmentation, delivered faster.

Log what the agent decided and why. The audit trail is not a compliance nicety. It is how trust gets rebuilt after the inevitable wrong call, and it is how the system improves rather than repeating the same mistake at volume.

For a related analysis of how automation changes the relationship between human judgment and institutional accountability, see Why Automation Still Needs Human Judgment.

Where the trust math turns against you

McKinsey is direct about the risk: because agents act across systems and data, small errors — pulling the wrong record or misapplying a rule — can compound into privacy failures or broken promises, which makes CX a higher-stakes environment than most other functions (Rodriguez et al.). Deep history sharpens the warning rather than softening it. The governance you write today is not a static rulebook. Culture is a second inheritance system layered on top of biology, and it spreads fast through copying and success (Henrich). Your agents will replicate the norms you encode across millions of interactions, faithfully, whether those norms protect the customer or quietly cut a corner.

So the concrete implication for a leader is narrower than the horizon talk suggests. Before scaling any agent, decide the two things that govern its trust exposure: what it is allowed to do without a human, and how any action it takes gets reversed when it is wrong. Get those two right in the high-volume, recoverable workflows first, measure decision quality rather than throughput, and expand authority only into places where a mistake stays cheap. That sequence — and not the ambition of the end state — is what separates the firms that build durable trust from the ones that automate their existing inconsistencies at speed.

The same cultural transmission logic that governs how norms spread through organizations is examined in depth in Dual Inheritance Theory: Bridging Cultural and Biological Evolution.

What would change my mind?

• If agentic AI scales just as fast in high-data, irreversible-stakes functions — clinical, legal, binding financial commitments — as in customer service, the recoverability argument is wrong and infrastructure alone explains where AI lands first.
• If scaled CX agents produce lasting trust and loyalty gains even when a visible error cannot be reversed, then reversibility matters less than argued here.
• If firms that let context travel across domains show no retention or relationship lift over firms that do not, the “restored memory” reading is decoration rather than mechanism.

Key takeaways

• Put autonomy where mistakes are cheap to reverse, and keep humans on the decisions where trust is fragile and errors compound.
• The shared customer record is the real asset; without unified identity, agents scale fragmentation faster instead of fixing it.
• Decide what each agent optimizes for, and which tradeoffs it may make, before deployment — or it will efficiently optimize the wrong thing.
• Log every agent decision and its rationale; the audit trail is how trust is rebuilt after a wrong call and how the system learns.
• Treat the governance you write now as the norm your agents will copy across millions of interactions, so encode the customer’s protection into it from the start.

References & further reading

Alexander, Richard D. The Biology of Moral Systems. Aldine de Gruyter, 1987.

Axelrod, Robert. The Evolution of Cooperation. Basic Books, 1984.

Dunbar, Robin. Grooming, Gossip, and the Evolution of Language. Harvard University Press, 1996.

Henrich, Joseph. The Secret of Our Success: How Culture Is Driving Human Evolution, Domesticating Our Species, and Making Us Smarter. Princeton University Press, 2015.

Nowak, Martin A. “Five Rules for the Evolution of Cooperation.” Science, vol. 314, no. 5805, 2006, pp. 1560–1563.

Rodriguez, Alex, et al. “Rewiring Customer Experience for the Agentic Era.” McKinsey & Company, 14 July 2026, www.mckinsey.com.