How it works
How your Zenffy agent works.
Every plan answers from your website and notes, books on your calendar and briefs your team. Enterprise setups add your systems, staff approvals and deployment where you choose. The labels show which is which.
From a request
to a checked result.
A useful agent needs more than a good answer. It needs the facts from your systems, a way to act on them, and a way to check the work is done.
Zenffy brings these together for your business. On every plan it works from your website, your notes and your calendar. Enterprise setups connect it to your own systems, such as a merchant’s payment device or a patient’s records.
One agent. More channels coming.
Knowledge engine Enterprise
Answers from what your company already knows.
On every plan, your agent answers from your website and your notes. Enterprise setups add your documents, tickets, chats and databases, keep each person’s access rules, and cite the sources for every answer.
What’s our refund window for a cancelled video consult?
Refunds for cancelled video consults go back to the original payment method within 7 working days1. If the patient paid by card, finance usually completes it in 2–3 days2.
- 1Refund policy, v3 SharePoint · updated 2 Oct
- 2Billing desk FAQ Confluence · updated 18 Sep
- Searches everything at once
- Every connected source in one search, combining keyword and meaning-based retrieval, then reranking the results.
- Cites every claim
- Answers are checked across sources. Each claim links to where it came from, with a confidence score.
- Respects permissions
- Access is synced from each source, so people only get answers from documents they are allowed to open.
- Reads your databases
- Ask PostgreSQL, MySQL, ClickHouse or SQL Server a question in plain English. Queries run read-only.
- Gets better at the job
- It remembers facts and preferences, and turns runs that went well into procedures it reuses.
- Shows up where people work
- Slack, Microsoft Teams, a browser extension, a web app, email, or a search or chat widget on your own website.
Integrations
Connects to the tools you already use.
Every plan books on Google Calendar and emails your team. Enterprise setups connect to your CRM, helpdesk, databases and your own systems.
How an agent works
The same four steps, every time.
01
Understand
Reads what your customer writes or says, and replies in their language.
Enterprise Voice notes, photos and PDFs on WhatsApp.
02
Look it up
Answers only from your website and your notes, and never makes up a fact.
Enterprise Amounts, dates and IDs read from your own systems.
03
Act and check
Takes the next step, like booking a call on your calendar, then confirms it with the customer.
Enterprise Acts in your systems within rules you set.
04
Hand over
Sends your team each lead with a short brief and the whole conversation.
Enterprise A person approves anything that moves money.
Human in the loop Enterprise
Autonomy you can manage.
On every plan, your team gets every lead with the whole conversation. Enterprise setups add risk tiers, budgets and staff approvals, so consequential actions wait for a person.
Waiting for your approval
Release a held payment
- Merchant
- Corner Street Café
- Held
- $1,200 · case RH-2291
- Merchant was told
- Two payments of the same amount came close together.
- Evidence
- Bill photo received. It matches both payments, at 4:12 pm and 4:17 pm.
- Every tool has a risk tier
- From T0, which runs freely, to T5, which is blocked. Policies can be set for the whole platform, a team or a single agent.
- Consequential steps wait
- Reads run freely. A write that matters pauses the run until a person approves it, then the run picks up where it stopped.
- The right person decides
- Approvals route through your org chart: a named approver, the on-call engineer or a manager.
- Decide from anywhere
- Approve or deny from Slack, an email reply or the web.
- Take over any conversation
- Staff can step into a customer chat with its full history, and the agent waits until they hand it back.
- Spending has a ceiling
- Each run and each team’s day has a cost limit, with a live meter. Automated writes are labelled as written by an agent.
On the record Enterprise
Every step is written down.
An agent’s work is a chain of four kinds of step. Each one is recorded, so your team can see what happened and why. Figures in a reply always come from a system step, never from a model.
- Rule
- A fixed check you set, such as “release a report only after a date-of-birth match” or “update firmware only when the device is idle”.
- Model
- An AI model reads the message, works out what is wrong or drafts the reply. Each step can use a different model, with a fallback if one is slow.
- System
- A call to one of your systems: read a record, book a slot, restart a device, attach a document to a case.
- Reply
- A message to the customer in their language, quoting figures exactly as your systems hold them.
Security and deployment Enterprise
Your data stays where you decide.
On Enterprise, run Zenffy in our cloud, in your own cloud account, on your servers or fully air-gapped. On every plan, each business’s data is kept separate from every other business’s.
- Where it runs
- Our cloud · your cloud · on-premise · air-gapped
- Models
- Bring your own, including local models
- Data leaving your network
- None required
- Sign-in
- OIDC single sign-on · SAML on enterprise plans
- Packaging
- Docker Compose · Kubernetes
- Training on your data
- Never
- Deploy where you choose
- Our cloud, your AWS, Google Cloud or Azure account, your own servers, or fully air-gapped with no outbound data.
- Bring your own model
- Use the AI provider you already trust, or models that run entirely on your hardware.
- Access control to the document
- Owner, admin, editor and viewer roles, plus document-level permissions synced from your identity provider.
- Guardrails on every input and output
- Detects prompt injection, redacts personal data, checks answers against their sources, and screens tool results before the model sees them.
- A full audit trail
- Every search, access and agent action is logged with its reasoning, with sensitive parameters redacted and retention you configure.
- Never trained on your data
- Your data is not used to train AI models.
Sending a security questionnaire? Talk to us and we’ll walk your team through the architecture.
