Capabilities

Flexivity AI has three capabilities that work together:

Progressive Autonomy decides what the AI does with each ticket, from labelling it to resolving it with the customer, across every channel your customers use.
AI Agent Assist puts the AI’s work in front of your agents, inside the help desk they already use.
Analysis shows where your knowledge base falls short and how well the AI is doing.

Progressive Autonomy

A hierarchy of AI automation

You decide how far the AI may go, ticket by ticket. There are seven levels, from doing nothing at all to resolving the request with the customer. Most teams start at Assist, where an agent reviews every reply. They move a class of tickets higher as they gain confidence in the answers.

Level	What the AI does
Ignore	Nothing. The ticket is left entirely to your team. Use it for automated mail and anything the AI should stay out of.
Label	Works out the ticket’s attributes and writes them to your help desk, so your own triggers and automations can use them. Nothing else runs.
Gated	Labels and summarises the ticket, but drafts no answer. A person handles it.
Assist	Drafts a reply from your knowledge base, and an agent decides whether to send it. This is the default for every ticket.
Reply	Sends the customer one reply directly, with no agent in between. Any follow-up from the customer goes to your agents.
Clarify	When a request is too vague to answer, asks the customer one clarifying question, then answers their reply. A clear request is answered directly.
Resolve	Keeps the conversation going until the request is resolved, within a limit you set. When the customer goes quiet, the ticket is either handed to an agent or marked solved, as you choose.

Answers are built only from your knowledge base. Every claim in a reply must cite a retrieved article, and a reply with an uncited claim is withheld automatically — this check is always on. On top of that, optional Answer Verification uses a second AI call to judge whether each cited article actually supports the claim. On tickets, verification runs before the reply is sent, so a failed reply is withheld; on chat it runs after, so a failed reply hands the conversation to a person. Verification is off until you turn it on, separately for tickets and for chat. A reply that fails either check, or a question your knowledge base does not cover, goes to an agent with the reason stated. Replies the AI sends to a customer say that they were written by AI.

A single switch can stop all replies to customers if needed, acting as an override regardless of what your conversation handling rules allow.

Note that replies are not screened for risk. They depend on the reliability of your knowledge base to recommend the right actions.

See AI Configuration for how the levels, rules and limits are set.

Changes in your systems: four write modes

The seven levels decide what the AI may say in a conversation. When a request needs a change in one of your own systems, such as a refund in Stripe or a new shipping address in Shopify, a second ladder decides who carries it out. Each tool that changes something has a write mode, set per brand, from the least the AI does to the most:

Write mode	Who carries out the change
Manual	A person, in your system. The AI hands the ticket over with what to do, and prepares nothing.
Assist	A person approves it. The AI prepares the change, with every check it ran, on the ticket; an agent approves, edits or rejects it. This is the default.
Hold	A person approves it while the customer waits. The AI tells the customer a colleague is confirming, and once the change is approved it tells them it is done and carries on.
Execute	The AI, within the tool’s limits, with no person approving. It tells the customer once your system confirms the change.

The two ladders work together. Hold and Execute apply only at Resolve, the one level where the AI carries a conversation on after a change; at every other level they work as Assist. A change is prepared only from a procedure, and every change is checked first: the record must come from a lookup for this customer on this ticket, and each tool’s limits apply. Changes run on Zendesk and osTicket tickets.

See Data Connectors for how write modes and limits are set.

Classification, attributes and rules

What the AI does with a ticket depends on what kind of ticket it is. Three pieces decide that:

Classification. As each ticket arrives, the AI reads it and decides a value for each of its attributes, in one step.
Attributes describe a ticket. There are three kinds:
- Flexivity attributes are custom attributes defined by you and maintained only by the AI system.
- Mapped attributes are fields in your help desk, such as the topic or category. The AI can fill a field in for you, or read a field your team owns.
- System attributes are built-in to Flexivity and decided by the AI, like sentiment, or worked out from the ticket itself, like whether an agent has replied.
Rules match on attribute values and choose a level for the matching tickets. For example: topic is Email → Resolve, sentiment is Frustrated → Gated. When several rules match, the more specific one wins.

Rules are set per brand, so each brand can have its own treatment. When an agent corrects an attribute value, the correction is what rules use from then on. Corrections are also recorded, so you can see how often the AI’s values are changed.

Data Connectors and Procedures

Data Connectors link the AI to the systems that hold your customers’ records, so it can look things up, and change them, on the customer’s behalf:

Stripe for charges and invoices, Shopify (in beta) for orders and products, and anything else through a custom MCP server, Zapier included.
Lookups, such as recent orders or a charge, answer questions with the customer’s own records. Every lookup is about the customer on the ticket: Flexivity takes who they are from your help desk, and nothing the AI writes or a customer types can point it at someone else.
Changes, such as a refund, cancelling an order or a new shipping address, are carried out as each tool’s write mode says, from a person approving it to the AI making it within limits you set.
Nothing runs until you enable it, tool by tool, and every tool can be tested first.

Procedures tell the AI how to handle a kind of request, in your own words: what to look up, what to ask, what to change, and when to hand over to a person. Start from a sample (such as Order status, Damaged item refund or Address change), paste the procedure your team already follows, or write your own. Flexivity shows you what it understood, step by step, before the procedure goes live, and test conversations let you check it behaves as you expect. Code, not the AI, enforces the steps that matter: a change is prepared only when the procedure reaches it, its target must come from a lookup for this customer, and a condition you set on a change, such as “only if the order has not shipped”, is checked against that record before the change can run.

See Data Connectors and Procedures.

Every channel

The same levels and rules apply wherever your customers reach you:

Tickets: email and web-form tickets in your help desk, on osTicket and Zendesk.
Chat: conversations in Zendesk messaging. The AI answers in the chat, and hands the conversation to an agent when it cannot help or the customer asks for a person.
Knowledge base AI search: a search widget on your public osTicket pages. Customers ask a question in their own words and get an answer drawn from your published articles, with links to the full articles, often before they open a ticket. On Zendesk, Help Center’s own search fills this role.

Your history, from day one

Flexivity AI imports your existing ticket history and knowledge base articles when you connect, so related tickets and answers are available immediately rather than after months of new tickets. The history sync runs in batches, so your help desk is not slowed down. Its progress is shown on the Help Desk page of the admin console. New and changed knowledge base articles and closed tickets are then picked up as they occur.

AI Agent Assist

Native to your help desk

AI Agent Assist is a panel inside your help desk’s own ticket view: a sidebar app in Zendesk, and a panel on the ticket page in osTicket. Agents do not switch tools or sign in anywhere else. The panel updates as the ticket changes. It has three parts: Take Action, Context and the Activity Log.

Context for decision making

Context answers one question: what should I know before I answer this? It is open by default and has five sections, always in the same order.

Sentiment is shown in the Context header, so it is visible even when the section is collapsed. It tells an agent how the customer feels before they read anything.
Summary is a short account of the whole conversation: the reported issue, what has been tried, and where things stand. It updates as messages arrive. An agent picking up a ticket that has passed between three people over two weeks reads three paragraphs instead of twenty-five messages.
Attributes shows the values the AI recorded for the ticket, such as its topic. An agent can correct any of them here, and the correction is what your rules use from then on.
Knowledge lists the knowledge base articles that match the ticket, with links. Only your knowledge base is used to ground an answer. If no article covers the ticket, the section says so, which is also a signal for the Knowledge Builder.
Related tickets are past tickets similar to this one, with links. They are shown automatically when the knowledge base has little to offer, and on request with Find related tickets otherwise. They show how your team handled something similar. They are not used as the source of an answer, because a closed ticket shows that someone stopped working on it, not that the resolution was right.

Take Action

Take Action appears only when the AI needs an agent, so an empty panel means nothing is waiting. From here an agent can:

Use a draft reply. At Assist, the AI’s draft is ready to review, edit and add to the reply. Draft a new reply writes another.
Pick up an escalated ticket. When the AI hands a ticket over, the card says why: the customer asked for a person, the AI could not verify its answer, your knowledge base did not cover the question, or the conversation reached its limit. Each reason points to a different next step.
Follow automated handling. While the AI is working a ticket at Clarify or Resolve, the panel shows where the conversation stands: what the AI sent, whether it is waiting for the customer, and how many messages remain. An agent who replies takes the conversation over, and the AI stops.

The Activity Log records what the AI did on the ticket. It is collapsed by default.

Agents can rate summaries and drafts with a thumbs up or down. Ratings feed Feedback Analytics.

Analysis

Knowledge Builder

The Knowledge Builder finds the questions your knowledge base does not answer, and drafts the articles.

How it works:

Resolved tickets are grouped into topics by similarity, to find the issues that keep coming back.
Each topic is compared with your published articles. A topic with many tickets and no matching article is a gap.
For each gap, the AI drafts an article from the most relevant resolved tickets, as a knowledge-base topic or as a short FAQ.
Administrators review drafts in the admin console and push them to the help desk: to Zendesk Help Center as unpublished drafts, or to osTicket as FAQ articles. They can be edited there before publishing.
The Knowledge Builder works one brand at a time. Your articles are kept in sync, so a topic you have since covered stops appearing as a gap.

Use case: After analysing six months of tickets, the Knowledge Builder finds that “VPN setup” is a frequent topic with no article. It drafts one from the 15 most relevant resolved tickets.

See Knowledge Builder for runs, settings and drafts.

Feedback Analytics

Agents rate AI-generated summaries and drafts directly in the help desk. Ratings are collected in the admin console, giving administrators evidence about AI quality.

How it works:

Each summary and draft has a thumbs up and thumbs down.
Feedback is anonymous: no information about the agent is captured.
The admin console shows trends over time, per feature.
Classification needs no ratings. The system records whether the values the AI chose were kept or changed by an agent.
AI Activity shows what the AI did, ticket by ticket, filtered by date, channel, level and attribute.

Feedback informs how administrators and the Flexivity AI team adjust prompts and configuration. It does not automatically train or tune the AI.

Use case: An administrator notices low ratings for summaries. Agents say the summaries are too long. The administrator shortens them through the summary prompt on AI Configuration, and the ratings improve.

Deployment Modes

Flexivity AI supports two deployment models to match different infrastructure and compliance requirements.

Direct Connection (Cloud-Managed)

Flexivity AI connects to your help desk directly from the cloud. Cloud help desk solutions like Zendesk always use this mode. This is also an option for internet-accessible osTicket instances.

Characteristics:

No on-premise infrastructure required beyond the help desk application itself
The cloud service connects directly to the help desk for new events
Ticket text is fetched when the AI needs it, with payment card, government and bank identifiers masked as it arrives (best effort), and scrubbed of PII before it is stored
Suitable for cloud-hosted or SaaS help desks where running a Docker agent is not practical

Self-Hosted Agent (On-Premise)

A Docker-based agent runs in your environment alongside your help desk application. It masks payment card, government and bank identifiers before anything leaves your network (best effort), and passes ticket text to the cloud when the AI needs it, where it is scrubbed of PII before it is stored.

Characteristics:

Payment card, government and bank identifiers are masked on-premise before transmission; other ticket text is read by the AI in the cloud and scrubbed there before it is stored
Outbound WebSocket connection only, over TLS on port 443, so no inbound firewall rules are required
Runs as a lightweight Docker container (Docker Engine 20.10+, Docker Compose v2)
Automatic updates via container image pulls