The decision to adopt a system often comes down to the automation rate. And it's also one of the most misunderstood metrics in automated customer service. "Automation rate" isn't a single quantity but at least three, distinct from one another and often reported as interchangeable.
The distinction between them is not a technical nuance. It is the difference between a verifiable figure and a declaration. For anyone leading a customer service operation destined to become critical infrastructure, knowing which quantity is being observed is the first act of governing the investment. A number, on its own, is not enough. What counts is the method by which it was produced.
This article does not describe how to build an AI Agent that improves, the architecture of the learning loop is the subject of Self-improving Agents. It defines the level before that, how the automation of a customer service system is measured in a defensible way.
Deflection, resolution and real automation
The three quantities the market tends to confuse are clearly distinct.
Deflection rate
This is the share of requests that do not reach a human operator. It is a measure of containment, it quantifies how much traffic has been diverted from the human queue. By construction, a request can be counted as "deflected" even when the customer abandons it without a useful answer. Deflection counts the interactions intercepted, not the problems closed.
Resolution rate
This is the share of requests in which the customer's need was actually met, ideally verified by an explicit confirmation or by the absence of a return contact on the same topic within a defined window. It is a measure of outcome. It is harder to construct and describes what truly matters for the company and for the customer.
Real automation
This is the share of interactions handled end-to-end by the system, including the actions executed on company systems, not just the answers provided, with a correct outcome and an acceptable level of quality. It is the only one tied to a defensible customer experience and to real value for the organization.
The rule for reading these figures is simple. In production, deflection is almost always the highest of the three, and resolution is the most solid. A number without its definition does not describe a result; it describes an ambition.
Why the same number can describe different realities
The same 90% can correspond to two systems that are very far apart. The difference depends on two methodological choices, both fully legitimate when made explicit, misleading when they remain implicit.
The first is what goes into the denominator. A system designed to handle a well-defined set of common, high-volume requests can legitimately reach a very high resolution on that set. This is not an anomaly; it is the direction of the market. Gartner predicts that, by 2029, Agentic AI will autonomously resolve around 80% of common customer service requests without human intervention. The point, therefore, is not how high the number is, but its specification, on which requests it is measured and by what criterion. A high rate on a clearly stated perimeter is proof of maturity; the same rate without a perimeter is incomplete information.
The second is how "resolved" is defined. Explicit customer confirmation, the absence of a return contact, or simply consulting a self-service content item, these are different criteria that produce different numbers. Here too, there is nothing improper in choosing a criterion, provided it is stated and applied consistently.
The practical consequence is that resolution varies enormously depending on the type of request. Simple, repetitive interactions are by now handled in self-service, while the requests that remain with human customer service are by definition the most complex. A dynamic that Gartner observes in the widespread adoption of Agent assist, where the value is concentrated precisely on the more intricate interactions. An aggregate figure that does not distinguish between these two worlds conceals more than it reveals.
The four measures that make a number credible
A mature organization does not replace a single percentage with another single percentage. It pairs it with a framework of measures that, together, tell the truth.
Resolution rate by type of request
Not an aggregate, but a breakdown. Resolution on order-status requests, on billing complaints, on contract changes, read separately. It is this reading that reveals where the system creates value and where it is simply bouncing the problem back.
Avoidable escalations and correct escalations
Not every handoff to an operator is a failure. Routing an emotionally sensitive complaint or a transactional operation to a person is the correct behavior, not a shortcoming of the system. Gartner's analysis points in the same direction, finding that leading organizations will steer AI toward creating value for the customer, and not toward cost reduction alone. The useful metric, therefore, is not the elimination of escalations, but the reduction of avoidable ones, the cases the system could have closed with adequate knowledge or integration.
Perceived quality after resolution
Customer satisfaction is measured downstream, not during the interaction. If perceived quality falls while deflection rises, the system is containing traffic at the expense of people. A good project keeps growing automation and stable quality together; it is the condition that separates a sustainable result from a number bound to deflate.
Durability of the resolution
A request may appear closed and then reopen the next day. It is the difference between an answer given and a problem actually solved. The honest measure is not the single completed interaction, but the case that stays closed, first-contact resolution and, above all, the recontact rate on the same topic. A number built on completed interactions flatters reality; a number built on cases that do not come back describes it. The correct unit of measure is the closed case, not the message sent.
Why the leap in quality does not come from changing the model
There is a persistent misconception that a higher resolution is a matter of model power. The data refutes it.
The MIT study, The GenAI Divide. State of AI in Business 2025 (Project NANDA), documents that around 95% of enterprise GenAI initiatives produce no measurable return; only 5% generate real value. The cause is not the quality of the models, but the way the projects are integrated and governed. Two findings of the study are particularly instructive. Solutions purchased from specialized vendors succeed roughly three times more often than in-house developments, and generic tools stall in the enterprise precisely because they neither learn from nor adapt to company workflows.
The reading is clear. Bringing real resolution into the high range does not depend on the most recent model, but on the scaffolding around it, a structured and up-to-date knowledge base, transactional integrations that allow AI Agents to perform actions in the systems and not only to respond, and context management that survives a change of channel. The work is done on the architecture, not on the foundation model. And it is also why indigo.ai has always chosen to remain agnostic with respect to LLM models, technologies, and voice: if it is the scaffolding that makes the difference and not the single model, tying yourself to one vendor is not an advantage; it is a constraint. It is the same reason a rule-based system, however up-to-date, hits a ceiling.
The time variable. A number that moves
There is one last reason to read every percentage carefully. Real automation is not a state; it is a trajectory. It is also why, alongside the resolution figure, it matters to watch how the user's perception moves over time. Trajectory evals, for example, customer sentiment analysis, which checks whether perception improves, worsens, or stays stable over the course of the interaction, measure exactly this: not the quality of a single response, but the direction in which the system is heading.
The resolution rate on the day of go-live is the starting point. A customer service system designed to analyze its own conversations, identify the cases it does not close, and propose, with human approval, the improvements that would close them, moves that number upward over time. It is the compound effect of the continuous learning of Self-improving Agents. Here, the consequence is enough; a mature vendor does not show a snapshot, it shows a curve, and can indicate how much it has moved and why. It is the same reason the initial figure should not be confused with the system's potential, a theme we explore when discussing time-to-value and maturation.
Recognizing these distinctions is not an exercise in caution. It is what separates a project that withstands the test of production from one that stalls at the pilot stage. A credible number is a number built on resolution, broken down by type of request, accompanied by perceived quality and by the durability of the closure, and tracked over time.
It is precisely this discipline of measurement that makes the most ambitious results reliable. An automation that reaches high levels, up to 96% on certain use cases, is a verifiable result when it is built this way: resolution measured, broken down by type of request, monitored over time, with the complex tail routed to an operator. Rigor in measurement does not limit ambitious numbers; it is what makes them sustainable.
FAQ
What is the difference between deflection and resolution?
Deflection measures how many requests did not reach a human operator; resolution measures how many actually met the customer's need. The first is a measure of containment, the second of outcome. In production, resolution is the more solid quantity to reason about.
Is a high automation rate inherently not credible?
No. A high rate on a well-defined perimeter is consistent with the direction of the market. Gartner predicts that by 2029, Agentic AI will autonomously resolve around 80% of common requests. Credibility does not depend on how high the number is, but on its specification, on which requests it is measured and by what resolution criterion.
Does a more powerful model guarantee a higher resolution?
No. The 2025 MIT NANDA study shows that 95% of GenAI projects generate no measurable return, for organizational and architectural reasons, more than for reasons of the model. The leap in quality comes from re-architecting knowledge and integrations, and from favoring specialized vendors, which, according to the same study, succeed roughly three times more often than in-house developments.
