AI assistance
AI that works like part of the team.
It works alongside your agents, you decide how much each one does on its own, and it answers customers directly so fewer questions become tickets. Runs on your choice of models, including Anthropic Claude and GLM. Included in every plan. Off whenever you want.
Two jobs, one AI.
An AI Agent that helps your team, or a customer-facing assistant that deflects work before it lands. Both optional, both included, both grounded in your own knowledge.
Agents on your team
Named workers that triage, reply, run workflows, fill forms, and more. You decide what each can do, and whether it acts alone or waits for a person.
AI facing your customers
On the chat widget and in the portal, it answers from your knowledge base, cites the article, and hands off to a human the moment it's unsure. Every answer is a ticket you never open.
AI Agents
An AI Agent is a teammate you configure.
Give it a name, a role, and a set of skills. It picks up tickets or chats like a colleague would, signs its replies, and shows up in the activity log by name. Entirely opt-in: none exists until you create one, and everything else here works without one.
You make two decisions
Do you want one? And how much should it do alone?
That's the whole model. First, whether you want an agent at all, nothing runs until you set one up. Then, per skill, whether it acts on its own or waits for a person. Let it triage and run a workflow by itself, but hold every customer reply for review. Everything starts untrusted, so autonomy is something you grant.
- Per-skill trust: auto-apply or propose for approval
- Mix freely, every capability set on its own
- Untrusted by default
- The agent never sees the trust flags, the system enforces them
Agent · per-skill trust
The skills you hand it.
An agent only does what you switch on. Each skill has its own trust setting, and never does more than the person it stands in for could.
Triage & route
Set priority and move a ticket to the right team. Your rules still win; the agent fills the gaps.
Reply, or draft a reply
Write a grounded, cited customer answer. Send it when trusted, or leave it as a draft to check.
Change status & priority
Move a ticket forward or close it out. Every change lands in the activity log, by the agent's name.
Write internal notes
Leave an agent-only note on what it found and did, context for the next person on the ticket.
Run workflows
Trigger an automation with the ticket's context, limited to the workflows you allow-list.
Fill in forms
Complete and submit a permitted form on the ticket's behalf, the same validated pipeline a person uses.
Add rows to a list
Append a record to one of your Lists, validated against its schema and limited to the lists you permit.
Draft knowledge articles
Turn a resolved ticket into a first-draft KB article. The agent writes it; you review and publish.
And you decide when it acts.
Separate from what it can do is when it does it. Pick the trigger that fits the job.
Watch a queue
React automatically to new tickets and customer replies in the teams it's bound to. Hands-off, opt-in by design.
Only when assigned
Acts only on tickets you assign to it, alongside a human owner. Work you hand it, nothing more.
Only when you ask
Nothing automatic. An "Ask AI agent" button runs it on demand, on the ticket in front of you.
Customer-facing AI
Answer customers before a ticket exists.
A deflection assistant on your live chat widget answers from the same knowledge base that grounds your agents, and cites the article it used. Set per channel how forward it is: greet first, or only step in when no agent is online. The instant it's unsure, or a customer asks for a person, it hands off with the full transcript. Never a dead end.
- Live-chat bot answers from your knowledge base, grounded and cited
- Per channel: bot-first, online-only-when-no-agent, or off
- Hands off to a human on low confidence or on request
- Escalation always fires, the customer is never stuck with the bot
Deflection bot · live chat
"How do I reset my API key?" The bot answers from your "Managing API keys" article, links the source, and offers a human if the customer still needs one.
Self-service in the portal
Help customers help themselves.
In the portal, an "Ask AI" assistant answers conversationally, grounded in the articles that customer can see. As someone raises a request, relevant articles surface while they type, so they often find the answer before submitting. Still need a person? One tap mints a ticket carrying the conversation.
- Conversational "Ask AI" panel in the authenticated portal
- Pre-submit article suggestions on raise-a-request
- Escalate to a real ticket with the transcript attached
- Turn each surface on independently, or leave them off
Ask AI · customer portal
A customer starts typing "I was charged twice." Before they submit, the portal surfaces your "Duplicate charges" article and offers to answer, or to file the request anyway.
Grounded in your words, and only the right ones.
Customer-facing AI answers from your published articles, never the open web, with the source cited. Audience rules are enforced where the AI reads, so it only ever sees what that customer is allowed to.
Cited, not invented
Every answer carries the article it drew from. That's RAG in practice: grounded in what you've written, not the web, so a person can verify it.
Never leaks the internal stuff
Restricted to articles its audience can read. Internal-only knowledge stays internal by construction, not by prompt instruction.
Assistive AI, no agent required.
Built-in on every plan. Most of these propose, and a person approves, so nothing changes a ticket without a human's say-so.
Auto-triage
New tickets are scanned for routing and priority on arrival. The AI proposes; a person confirms before anything applies.
Thread summaries
Long conversations get a plain-language summary, so a new agent can pick up without reading the whole thread.
Resolution summaries
When a ticket closes, the AI writes a short summary of what happened, handy for knowledge drafts and reporting.
Similar resolved tickets
Before an agent replies, Sysflows surfaces tickets resolved the same way before. Semantic matching, not keyword lookup.
Semantic knowledge search
Your knowledge base is indexed with vector embeddings, so search finds the right article even when the words don't match.
Spam & silent close
Junk, marketing, and bulk mail flagged with a one-click silent close, no notification to the sender.
Model Context Protocol · Pro
Your AI and your tools, one protocol.
MCP is the open standard for connecting AI to tools, and Sysflows speaks it both ways. Reach out: a trusted agent calls a tool on your own MCP server, inside the same trust and approval it already uses. Connect in: point Claude Desktop or any MCP client at Sysflows over a scoped, revocable token, never seeing more than the person it belongs to.
- Agents call your MCP tools inside the existing trust + approval model
- Workflows call a tool deterministically with the MCP tool node
- External clients connect in with a scoped, instantly-revocable token
- Every call routes through one audited boundary, capped and logged
- Off by default, Pro-only, never billed against your token allowance
MCP · both directions
A trusted agent calls your inventory MCP server to check stock, then drafts the reply. Separately, an analyst's Claude Desktop reads this week's tickets over a scoped token. Every call audited.
You stay in control
Every feature has a switch. So does the whole system.
Want summaries but not triage? One toggle. Want it all off? One master switch. Nothing the AI proposes applies until a human confirms it, unless you've trusted a specific agent skill to act. The defaults are conservative.
- Master on/off switch for the entire AI system
- Independent per-feature toggles: triage, summaries, search, spam
- Every suggestion human-confirmable before it applies
- Agent skills set individually to "propose" or "auto-apply"
- Token usage tracked live, with 80% and 100% alerts
AI settings
Token allowances, in plain English.
Every plan includes a fixed token allowance per seat per month, pooled across your whole team, so a quiet seat shares its budget with a busy one.
Starter · ~30 AI resolutions
Growth · ~80 AI resolutions
Pro · ~165 AI resolutions
One AI resolution (a complete draft-reply exchange) runs roughly 10,000 tokens; the estimates above vary with conversation length and features used.
Overage is opt-in only. You're alerted at 80%, and at 100% AI features pause unless you've enabled a top-up. Top-ups cost $20 per 1M tokens and are never charged automatically. You decide before you spend.
What's where, by plan.
Assistive AI and deflection are on every tier. Configurable agents and inbound MCP step up as you grow.
Every plan
Triage, summaries, similar tickets, semantic search, spam, deflection bot, portal Ask AI, and on-demand drafts. Included from day one.
Growth & Pro
Configurable, autonomous AI Agents with per-skill trust, trigger modes, and allow-lists. One per team or workflow.
Pro
The inbound Sysflows MCP server and API tokens: connect Claude Desktop and other MCP clients over a scoped, revocable token.
Your choice of frontier models.
No proprietary black box, just well-documented models you can read about yourself, and pick between.
Multiple models via Amazon Bedrock
Generative work, agent loops, drafts, summaries, triage, runs on the model you select from a catalogue served through Amazon Bedrock, including Anthropic Claude and GLM. Choose a model yourself, or leave selection automatic and cost-aware: fast models for high-volume jobs, stronger ones for the multi-step agent loop.
Vector embeddings on pgvector
Semantic search (the right article or a similar past ticket) uses vector embeddings in pgvector. It finds meaning, not just matching words.
Common questions about AI
What's the difference between assistive AI and an AI Agent?
Assistive AI (triage, summaries, similar tickets, search, spam, drafts) is built into every plan and proposes for a person to approve. An AI Agent is a named worker you configure on Growth and Pro that can act on its own, within the trust and trigger limits you set. Agents are opt-in, none exists until you create one.
Can an AI Agent send replies to customers on its own?
Only if you trust it to. Each skill is "auto-apply" or "propose for approval" independently. A common setup lets it triage and run a workflow on its own, but holds every customer reply as a draft. Everything starts untrusted, so you grant autonomy skill by skill.
Will the deflection bot ever quote our internal notes?
No. It's restricted to articles its audience can read. Internal-only articles and notes are excluded at the source, not by a prompt instruction, so the bot can't surface them even if asked.
What happens when the bot can't answer?
It hands off to a human. On live chat it re-queues to a real agent with the full transcript; in the portal, one tap mints a ticket carrying the conversation. Escalation always fires, so a customer is never stuck with a bot.
Do the AI drafts actually sound like us?
Drafts are grounded in your knowledge base and the thread, so they reflect your content and tone, and each cites the article it used. The more you build out your knowledge base, the more on-brand they get. A person always reviews before anything sends.
What happens when we hit the token cap?
At 80% you get an alert; at 100%, AI features pause. The platform keeps working normally. Agents just won't see AI drafts or summaries until the next billing period or until you enable a top-up. No surprise charge, no auto-overage.
Which AI models does Sysflows use?
You choose. Models are served through Amazon Bedrock, from a catalogue that includes Anthropic Claude and GLM. Pick a preferred model for your account, or leave it on automatic and Sysflows selects a cost-aware model for each job. Each model has a published weight against your token allowance, shown in the picker, so premium models draw tokens faster and lighter models draw fewer.
Is our conversation data used to train models?
No. Your data is in your own isolated database and is not used to train any model. AI calls go through Amazon Bedrock, which does not use your inputs or outputs to train models and does not share them with the model providers.
Start free. AI included.
No credit card. 14-day full trial. AI on from day one.