From Understanding to Action: The Next Step in Complaint Management AI

In many complaint management systems, AI’s role is focused on providing text summaries and guidance. These features help complaint handlers understand cases better, interpret procedures and prepare responses, but leave much of the work of progressing the complaint to the handler.

The reality is that these capabilities save time, which is helpful, but their ultimate impact is limited because the handler must still decide and initiate every next step.

At Complyr, our approach to AI in complaint management goes further. We believe significantly greater value is unlocked by connecting automated AI decisions directly to the complaint workflow.

Alongside summaries and guidance, Complyr focuses on using AI to make the decisions that move complaints forward, connecting those decisions directly to workflows that users can design, control and update inside the system. This opens up opportunities for real-time routing, prioritisation and escalation of complaints, with explicit rules governing when the system acts automatically and when a case is sent for human review.

A useful way to understand our approach to AI comes from psychology. In Thinking, Fast and Slow, Daniel Kahneman popularised the distinction between System 1 and System 2 thinking, adopting terms introduced by psychologists Keith Stanovich and Richard West. System 1 describes fast, automatic, intuitive judgements. System 2 describes slower, deliberate reasoning for critical thinking and working through a problem.

AI borrows this distinction as a design analogy. System 1 models are built for rapid, focused judgements inside software. System 2 reasoning describes a more deliberative approach, useful when a task requires extended analysis or several reasoning steps, as often seen in mainstream models such as ChatGPT and Claude.

In complaint handling, summarisation and guidance are typically System 2 tasks, but the quick, often intuitive decisions that human handlers constantly make are System 1 tasks. The AI driving most complaint management systems uses System 2 models, whereas Complyr uses a dynamic combination of System 1 and 2 models to give our AI helpers a true human approach to complaint management.

The technical distinction between them starts with how they produce answers. System 2 models build an answer piece by piece. Given a complaint and instructions, the model uses patterns learned during training to predict a suitable next piece of text, called a token. It then repeats the process, taking account of everything it’s written so far. This sequential process, called autoregressive generation, provides the flexibility to produce summaries, explanations and detailed responses.

Purpose-built System 1 models take a different approach. They receive the relevant state, such as complaint text, contact history and account information, alongside focused questions with precisely defined answer spaces. These models might ask which category a complaint belongs to, how severe it is against a rubric, or whether escalation criteria have been met.

These models can then act as powerful decision makers, returning values with a predefined structure and a probability that complaint workflows can use directly. An example categorisation might assign billing a probability of 0.93, service delivery 0.05 and other 0.02. The workflow then applies its rules to automatically select the next step based on that decision, moving the complaint forward.

Behind the scenes, there’s a highly specialised architecture. Independent questions are evaluated through a single parallel process, without an autoregressive output-generation loop. Category, urgency and missing-information assessments can share the same context and be answered simultaneously, reducing inference time. This means that multiple important decisions are made in just a few hundred milliseconds.

Another great advantage of our System 1 models is that they completely eliminate hallucinated outputs. The model can’t invent a category, action or answer beyond those you provide in the workflow, allowing us to deliver a “zero-hallucination” output, backed at all times by probability levels that can be used to control whether a decision is accepted or sent for human review.

The difference between System 1 and 2 models extends to how they’re trained. System 2 models typically begin by learning patterns from large amounts of text, then undergo further training to improve their responses. One common method is reinforcement learning from human feedback, or RLHF. People compare candidate responses and their preferences train a reward model, and reinforcement learning encourages the language model to produce responses the human evaluators prefer. This can improve usefulness, instruction following and truthfulness, but does not automatically give you confidence or certainty in the responses generated.

On the other hand, our System 1 models lean heavily on reinforcement learning for calibrated decisions, or RLCD. This targets a different output where bounded judgements are accompanied by probabilities that reflect uncertainty. It aims to make the model’s reported probability correspond to actually observed outcomes. Across comparable assessments assigned a probability of 90%, the relevant outcome occurs approximately 90% of the time.

For complaint workflows, the difference between System 1 and 2 models matters. While a clearly written, helpful response supports a handler’s understanding, a decision accompanied by validated probabilities lets the complaint system determine whether to handle the case automatically, gather more evidence or request human review.

Human complaint handlers are constantly making decisions under uncertainty. For example, they may be confident about the appropriate category of a complaint while remaining unsure about its urgency. Much like our AI models, the level of uncertainty they feel influences whether they make a choice, investigate further or consult a colleague.

System 1 models expose the uncertainty in a structured form. Probabilities concentrated on one option indicate a clear assessment; probabilities spread across options indicate ambiguity. A confidence statistic derived from the distribution is also provided as an additional signal to help make a decision. These are the signals that complaint workflows can use to determine when and how to act.

As we move beyond AI text summaries and guidance to making decisions that actually move the complaint forward, this is incredibly valuable. Complyr’s configurable workflows provide the foundation for defining which actions follow particular conditions. Connecting AI assessments to those conditions can extend automation across complaint routing, prioritisation, information gathering, messaging and escalation.

Flexibility is the key here. Several assessments, run in parallel, can be used to automate business rules. A clear billing categorisation might assign an action to a particular person, while urgency determines the deadline for that action. An ambiguous classification could trigger a human review. Missing information could initiate an evidence request without preventing assignment. Each assessment can have one or more roles, with the complaint workflow controlling how they interact with each other.

Consider a customer reporting a duplicate charge after two unsuccessful contacts. A workflow, with a defined AI assessment, could route the complaint to billing, flag the repeated contacts and initiate a transaction check, while a System 2 model provides summaries and policy guidance to give clear context around the case.

It doesn’t stop there, though. Automation can continue throughout the complaint lifecycle. A new message received, a document uploaded, or notes updated could all trigger another assessment where the workflow requires it. Updated assessments may reveal greater detail, heightened urgency or customer vulnerability. This lets the workflow respond as the complaint develops, reducing gaps between information becoming available and prompt action.

Using a workflow to handle uncertainty not only provides complaint automation but also gives consistency and supports auditability. Human uncertainty, when captured, is usually seen in case notes. With a System 1 model, the numerical probabilities and confidence signal become a routine part of each assessment. Complyr’s complaint management software automatically records the context, question, probabilities, confidence, thresholds and resulting decision, creating a permanent record of the assessment and workflow rules that produced each action.

Those records can also be used to improve the complaint process. Repeated referrals for a human review might reveal unclear categories, missing context or unsuitable thresholds. Comparing assessments with verified outcomes helps teams calibrate their workflows and identify suitable, safe automation boundaries. When the workflow is correctly tuned, time to first action drops, unnecessary handoffs decrease, and system performance relates directly to positive complaint outcomes.

For businesses using Complyr, the big opportunity is to connect AI decisions to workflows they have full control over. Fast decisions can move complaints forward immediately, while exposed uncertainty, explicit rules and a full audit trail make those actions easier to review, explain and improve.