AI Glossary / AI and machine learning / Beginner

Feedback

Definition

Feedback is information about how things went, returned to the system that produced them, so future behavior can adjust: outcomes flowing back to the actor.

What it means

Every learning system, human or artificial, needs to know what happened next. A model predicts, reality responds, and the response, captured as data, is feedback. Without it there is no learning, only repetition.

Feedback comes in speeds and qualities. Instant, clean feedback, the customer clicked or did not, teaches fast. Delayed, noisy feedback, the loan defaulted or not, teaches slowly and ambiguously, because months pass and many causes intervene.

In reinforcement learning, feedback is formalized as reward: the numeric consequence of an action. In ordinary analytics it is the labeled outcome: what the customer actually did after we acted.

The design of feedback collection decides what can be learned: if outcomes are never recorded against predictions, the organization has no idea whether it is right, and no mechanism to improve.

For a business, feedback is the difference between a system and a loop: decisions that generate recorded outcomes become assets, and decisions that vanish become expensive habits nobody can evaluate.

Feedback also has a direction problem: you usually learn the outcome of what you did, never of what you did not do. The road not taken leaves no data, which is why holdouts and experiments are the only way to see the alternatives.

How it works

Decide, for every prediction or action, what outcome would count as the result, and over what window it should be measured.

Capture outcomes against the original decision: the same identifiers must link the prediction, the action and what followed.

Handle delay honestly: late outcomes need pending states and patience, not premature scoring of decisions as successes.

Feed results back to whoever or whatever learns: the model's retraining, the team's review, the vendor's report.

Audit feedback quality: missing outcomes, biased recording and gamed metrics all teach lies fluently.

Close the loop on a schedule: feedback that lands in a quarterly review teaches the organization quarterly; feedback in a nightly retrain teaches it daily.

Three practical examples

• Recommendations: clicks and purchases recorded against the suggestions that produced them.
• Credit: repayment outcomes tracked against the approval decisions, over years.
• Support automation: resolution and escalation logged against each automated reply.

Think of it this way

Think of archery practice. Shooting arrows is action; seeing where they land is feedback. A blindfolded archer shoots all day and improves nothing, because improvement is not in the shooting, it is in the seeing.

The limit: the target can be moved by wind the archer never feels. Feedback arrives through a noisy world, so reading it well is its own skill.

Worked example

closing the loop on collections

Suppose a collections team tries a new contact-timing model but has no habit of recording outcomes against treatments. The figures are illustrative.

They add one discipline: every contact logs the model's recommendation, the action taken and the payment outcome over 30 days.

After two months the data shows the model's timings beat the old schedule by 11 per cent on mid-risk accounts and lose on low-risk ones.

The team restricts the model to mid-risk, a conclusion invisible before outcomes were linked to decisions.

Illustrative case study

A fictional retailer ran personalized promotions for a year with feedback defined as redemption, and celebrated ever-rising redemption, until someone asked what would have happened anyway.

Because outcomes were never linked to a holdout group, the feedback measured appetite, not impact, and the program's true effect was near zero.

The invented lesson: feedback without a comparison is applause, not evidence. Record what happened, and what would have happened.

The invented follow-up: the retailer now defines impact as measured-against-holdout in every program brief, and redemption-only claims no longer survive the first review.

Common misconceptions and mistakes

• Feedback is automatic. It must be designed: outcomes defined, linked and captured, and most organizations discover their loops are broken only when they look.
• More feedback is always better. Wrong or gamed feedback is worse than none, because it teaches confidently in the wrong direction.
• Feedback equals accuracy. It includes fairness, cost and side effects; the outcome you measure is the behavior you get.

Business use

Link every automated decision to its recorded outcome as a non-negotiable data practice; unlinked decisions are unlearnable ones.

Define the outcome window honestly: measuring too early flatters quick bad decisions over slow good ones.

Keep a holdout where you can: impact is the difference between what happened and what would have, and only comparison reveals it.

Assign feedback collection an owner per decision system; loops without owners quietly open, and open loops look like working systems until someone asks for evidence.

People also ask

What is delayed feedback?

Outcomes that arrive long after the decision, like loan defaults, requiring pending states and careful window design.

How does feedback differ from a metric?

A metric summarizes; feedback is the per-decision outcome data that metrics are built from and learning consumes.

Can feedback be gamed?

Yes; measure agents on redemption and they find redeemable customers, not incremental ones. Design for impact.

Check your understanding

Why must outcomes be linked back to the specific decision that caused them?

Answer: Because learning requires knowing which action produced which result, and unlinked outcomes can describe the world but cannot improve the decisions.

Related terms

• Feedback loop
• Reinforcement learning
• Label
• Retraining
• reward function
• holdout

Learn next

AI for Business Owners explains how to close the loop between decisions and outcomes so systems actually improve: https://aiforbusinessandlife.com/academy/ai-for-business-owners. [C2]

More in AI and machine learning

• Federated learning - Federated learning is training a shared model across many devices or organizations without moving their raw data to one place: the training goes to the...
• Feed-forward neural network density model - A feed-forward neural network density model uses a standard layered network to estimate the probability distribution of the data itself, so it can score how...
• Feed-forward neural network regression - Feed-forward neural network regression is using a standard layered network to predict a number, such as a price, a demand or a risk score, from input features.
• Feedback neural network - A feedback neural network includes connections that allow earlier internal activity or later-layer information to influence subsequent processing.
• Filters, random fields, and maximum entropy model - The filters, random fields, and maximum entropy model, usually called FRAME, is a probabilistic model developed for texture analysis and synthesis.
• Finite mixture model - A finite mixture model describes data as coming from a fixed number of underlying groups, each with its own distribution, so one dataset can be read as...

Sources

[S1] Russell and Norvig, AIMA glossary; [S2] Google, Machine Learning Glossary. Full URLs are in the source ledger. Verified 11 October 2026.

Previous: Feed-forward stylization Next: Feedback loop