# Learns every night. Keeps only what works.

Source: https://aide.app/platform/self-improving-ai-agents · Updated 2026-11-16

Self-improving AI agents

Every night Aide compares its drafts with the replies your team sent, tests each change before keeping it, and undoes it if quality drops. Every change is listed with an Undo.

[Start free](https://my.aide.app/register) [Get a demo](https://aide.app/demo)

## What is a self-improving AI agent?

A self-improving AI agent gets better from the replies your team sends, instead of waiting for someone to edit its settings. In Aide, each night's learning is tested before it is kept, and anything a person wrote or customers can read waits for a person.

Updated November 16, 2026

- When it learns: Every night
- Learns from: The replies your team sent, beside the drafts Aide wrote
- On its own: How it writes, its own knowledge once you allow it, and topics customers keep raising
- Needs a person: Anything a person wrote or customers can read
- Safeguard: Each change is tested first, and undone if quality drops
- Plans: Every plan

## Four steps, every night. Compare, find, test, then keep or undo.

01 · Compare

### Aide reads its drafts beside what your team sent. Sent as is, edited, rewritten or not sent.

02 · Find

### It groups what went wrong into themes. And finds what your team keeps changing.

Picture: A conversation thread: Aide's draft as an internal note, and under it the reply a teammate wrote and sent instead.

03 · Test

### It tests the new wording before keeping it. On conversations it has not learned from, and on every eval it already passes.

04 · Keep or undo

### It keeps what tested better. If your team starts rewriting more drafts, the old wording comes back.

Picture: A learned change on the Quality page: what Aide changed in how it writes, the wording before and after, when Aide applied it, and an Undo button.

## It learns how your team writes. And what your team knows.

### How it writes. Tone and style, matched to your team.

- What changes: Openers, sign-offs, length and tone, in your agent's Tone & identity.
- Where you see it: Under Learned changes on the Quality page, with the wording before and after and an Undo.
- Your edits win: An edit your team makes afterwards stays exactly as your team wrote it.
- If Aide undoes one: Aide tells owners and admins.

### What it knows. Turn on one setting and Aide adds what it learned to its own knowledge.

- What it adds: An internal snippet with a fact your team keeps giving, or how your team handles a kind of request.
- Evidence: At least two teammates said it, in at least two conversations.
- Pace: At most 10 changes a night, newest replies first.
- Dated facts: A sale or a deadline is added with its end date, and ends that day.
- Where you see it: Applied by Aide, with the conversations behind each change and an Undo.

### What your knowledge is missing. Listed every night, with the text and the article it belongs in.

Each gap is checked against your whole knowledge base first, and ordered by how many conversations raised it.

Picture: A knowledge gap card: the question customers asked, how many conversations raised it, the article it belongs next to, and the suggested text with a Copy button.

### New reasons customers write. Aide suggests a topic when none fits, and adds it once customers keep raising it.

A topic Aide adds is badged Added automatically and keeps the AI level those messages already had.

Picture: A customer message no topic fits, with Aide's suggested topic on it: how many messages it was seen in, its name and group, and Add topic.

## See whether it is getting better. On a scorecard, beside your own team.

### Aide, graded beside your team. On the same scorecard, next to your median and your top quarter.

Each night a sample of your teammates' replies and of Aide's is graded the same way, on topics both replied to.

Picture: The Compared with your team band on the Quality page: Aide's share of acceptable replies beside the team median and the top quartile, each with its range.

### The Quality page. A written bar, and a pass or fail against it.

- Scorecard: Meets the quality bar or below it, with how many more acceptable replies would reach it.
- Evals: Test conversations built from what went wrong, replayed against your agent every night.
- Draft acceptance: How often your team sends Aide's drafts as is or with small edits.
- Topic by topic: Whether Aide's drafts are good enough for it to send replies itself.

### The Impact page. Your key numbers, with each learning marked on the trend.

- Five numbers: AI Resolution Rate, Acceptable replies, Drafts sent, CSAT and Median first response, each against the period before.
- Markers: One on each day a learning was applied, by Aide or by a teammate.
- Where it was used: The conversations a learning was given to Aide in, and how they ended.
- How to read it: A marker is a date, not a cause.

## You decide how far it goes. And you can take any of it back.

### A person decides. Every time, on each of these.

- A correction to an article
- Publishing a draft
- A change to a snippet someone wrote or edited
- Anything for your agent's Instructions
- A question where your team's replies disagree

### You set each topic's AI level. AI Agent, Drafts or None.

Quality shows when a topic's drafts are ready to send. An owner or admin turns sending on, and can turn it back.

Picture: A topic's page: a teammate changes its AI replies level from Drafts to AI Agent, and the hint under it says what that level lets Aide do.

### Apply, or wait for you

One setting, set by an owner or admin, decides whether Aide adds what it learns to its knowledge or suggests it.

### Undo that respects your work

If a teammate changed the text since, Undo says so and leaves their edit in place.

### A notice the next morning

Owners and admins get one notification after a night in which Aide applied changes.

### In the audit log

Changes to your agent, topics and knowledge are recorded with a before and after, kept two years.

## Ask for the next improvement. In plain words.

### Ask why your team rewrote Aide's drafts. Aide reads them and proposes the smallest fix.

Picture: An owner asks Aide why the team keeps rewriting its drafts. Aide works through the conversations and shows an approval card with one added instruction line, Approve and Decline.

### Approve it, and Aide retests. Then it shows you the new answer.

Picture: The approved change runs, then Aide tests the agent on the same question and shows the new reply.

Aide Superintelligence is for owners and admins. It asks before anything customers would notice.

- [How Aide Superintelligence works](https://aide.app/superintelligence)

## One loop, for every kind of team. It learns whatever your team keeps correcting.

Software

### Agents keep shortening Aide's billing answers. Aide writes shorter.

How it writes

Training provider

### The team keeps adding how much notice a reschedule needs. With the setting on, Aide adds it to its knowledge.

Its own knowledge

Online store

### Customers keep asking about a new product line. Aide suggests a topic, then adds it.

A new topic

Logistics

### The team keeps explaining how to get proof of delivery. Aide lists the gap, with the text to add.

A knowledge gap

## Frequently asked questions

A self-improving AI agent is a customer service agent that learns from the work around it: the replies people send, the edits they make to its drafts, the questions its knowledge does not cover. Aide's agents do this every night, keep a change only after testing it, and restore the previous wording if quality drops.

### How does Aide learn from my team?

Every night Aide reads each draft it wrote beside what your team did with it: sent as is, edited, rewritten or not sent. It groups what went wrong into themes, turns them into test conversations for its evals, and looks for the way your team writes. Aide's own drafts sent unchanged never count as evidence, so it learns from your team and never from itself.

### Does Aide change anything without my approval?

Aide changes two things without asking: how it writes, after testing the change, and its topic list, when customers keep raising a subject no topic covers. It adds to its own knowledge only after an owner or admin turns on Apply what Aide learns automatically. A correction to an article, a change to a snippet a teammate wrote, and anything for your Instructions stay suggestions until a person applies them.

### How do I undo something Aide learned?

Open the change and select Undo. Aide lists its writing changes under Learned changes on the Quality page, with the wording before and after, and the knowledge it added under Applied by Aide, with the conversations behind it. If a teammate has edited the same text since, Undo says so and leaves their edit in place.

### How do I know Aide is getting better?

Aide's Quality page grades replies pass or fail against a written quality bar, runs evals every night, and shows Aide beside your team's median and top quarter on the same scorecard. The Impact page draws your key numbers as a trend with a marker on each day a learning was applied. A marker is a date, not a cause.

### Does it work inside Zendesk, Gorgias or Front?

Yes. Connecting Zendesk, Gorgias or Front turns on Aide's drafts inside your helpdesk, and every night Aide compares those drafts with the replies your team sent there. Your team keeps working in the helpdesk it already uses.

### How soon does Aide start learning?

From day one on a connected helpdesk. Aide works from the replies your team already sent to list its first knowledge gaps and to grade your team on the Quality scorecard, then runs every night on that day's drafts and replies.

### Which Aide plans include self-improving AI agents?

Every plan, the Free plan included. The nightly learning runs on Aide's own budget: it is not counted in your AI replies and never slows a live reply.

## Related

- [Aide Superintelligence](https://aide.app/superintelligence)
- [Customer service analytics](https://aide.app/analytics)
- [Knowledge management](https://aide.app/solutions/knowledge-management)
- [Agent governance](https://aide.app/solutions/agent-governance)
- [Procedures](https://aide.app/platform/procedures)

## An agent that gets better every night

Start free, connect a channel and let Aide draft beside your team. It learns from the replies they send.

Not ready yet? Get product updates by email
