The AI agent
It answers from what your business already wrote down.
EzyConn’s agent reads your website, your documents and your price list, and uses them to answer the questions customers actually send. It is not a general chatbot with your logo on it: when the answer is not in your content, it says so and puts a person on it instead of guessing.
In one line
Your content, a fixed check before every message, a search that re-reads its own shortlist, and a reply that names its source — in 30 languages, on the provider you choose.
Why it matters
The same nine questions arrive every day. What do you charge. Do you cover my street. Are you open Saturday. Can I move Thursday. Do you do that at all. Every one of them has a settled answer, written down somewhere, and every one of them still costs somebody four minutes and an interruption.
A generic chatbot makes that worse, because it answers confidently from something other than your price list. The customer then arrives expecting a number nobody in your business has ever quoted.
What this agent does is narrower and more useful: it answers from your material, shows which page it took the answer from, and stops when your material runs out. The nine questions stop reaching a human, and the tenth still does.
How it works
Six things happen between the message and the reply.
None of them is a black box you have to take on trust. Every step below is something you can check the result of in your own account.
- 01
You point it at what you already wrote
Give it your website and it crawls the pages itself. Upload the price list, the terms, the aftercare sheet, the two-page document that answers ninety per cent of your email.
- 02
Long documents are cut into readable pieces
A forty-page PDF is not one answer. It is split so the agent can find the paragraph that matters instead of quoting the whole file.
- 03
Every message passes a check before anything is read
On your website widget, a fixed rule set looks at the incoming message before any search or any model call. No model is asked whether the message is safe.
- 04
It finds candidates, then re-reads them
A first pass pulls the passages that look relevant. A second pass re-ranks them, so a near-match on wording does not beat the paragraph that actually answers the question.
- 05
It answers, and shows where the answer came from
The reply carries the source it used. If nothing in your content supports an answer, it says so and offers a person rather than inventing one.
- 06
On your website, the words arrive as they are written
Real token-by-token streaming, not a spinner followed by a paragraph. The visitor starts reading while the sentence is still being produced.
The safety gate
A rule set, not a second opinion from a model.
On your website widget, every incoming message passes five layers of fixed checks before the agent searches your content. There is no model call in that path. It takes well under a millisecond, it is on by default, and no setting in your account turns it off.
Same message, same outcome
Because it is deterministic, it can be audited. The customer’s text is not written into the logs to make that possible.
It does not get in the way
Across 720 ordinary administrative messages in testing it refused none of them. A patient in pain asking for an appointment is still offered one.
It stops on wording that reads as an emergency
It ends the exchange and points to urgent care. Emergency wording, and direct requests for a quantity of a drug, are recognised in eleven languages.
What it is not. It is not triage, symptom intake, urgency detection, emergency routing or escalation to clinical staff. None of that exists in EzyConn, and we are not going to imply it does. It is not compliance and it is not clinical governance.
What it does not reach. It runs on the website widget and nowhere else — not on the phone line, not on WhatsApp, Telegram or email, and not on the staff test bot. It does not catch every phrasing it has never seen, its classifier reads English, and it is not built to withstand somebody deliberately working around it.
Who uses it
Eight industries, and the same agent behind all of them.
What changes between them is your content, your services and your hours — not the machinery. Three verticals are administrative work only, and we say so on their pages rather than in a footnote.
Home services
HVAC, plumbing, electrical and any trade that drives to the customer.
Automotive
Mobile mechanics and independent workshops.
Aesthetics & beauty
Administrative work onlySalons, spas and clinics booking treatments.
Retail & showroom
Consultations, quotes and purchase appointments.
Dental
Administrative work onlyNew patients, appointment changes and treatment enquiries.
Healthcare
Administrative work onlyAppointment scheduling and patient routing for practices.
Wellness & fitness
Studios, gyms and practitioners booking sessions.
Translation & interpretation
Agencies capturing jobs and booking interpreters.
What it can do
Specific things, each of which you can check.
- Choose your model provider
- OpenAI, Anthropic or Google. Switch without rebuilding anything you have trained.
- Use your own API key
- Paste your provider key and it takes priority over ours. If it stops working mid-conversation, the agent falls back at runtime rather than failing in front of a customer.
- Answer in 30 languages
- The customer writes in theirs, the agent replies in theirs, and it stays in that language for the rest of the conversation.
- Remember a returning customer
- What they told you last time is available this time, so they are not asked their own name twice.
- Hand over cleanly
- Three unproductive turns, an explicit request for a person, or a question it will not answer, and your team gets the conversation with the full transcript.
- Stop before the bill does
- A spend cap and rate limiting are enforced on every path, including the streaming one. The agent stops; it does not quietly keep spending.
- Run one of your own actions
- AI Actions call your endpoint mid-conversation, so the agent can look something up in your system and answer with the result.
- Improve on evidence, not on opinion
- Changes to how the agent answers are scored against a fixed set of questions before they ship, so a tweak that helps one case and quietly breaks four does not reach your customers.
What it works with
The agent answers on every channel EzyConn supports.
One set of content, one set of answers. The gate described above is the exception: it is wired on the website widget only.
- Your website
- Your phone line
- Slack
- Microsoft TeamsBeta
- Telegram
- Zoom
Your model provider is OpenAI, Anthropic or Google — your choice, and your key if you have one. Everything the agent produces can leave through Zapier or a signed webhook.
What plan includes it
The agent itself is on every plan. Four things it can do are not.
The base agent — your content, retrieval, citations, streaming, 30 languages, memory — is included from the free tier up, and metered there like everything else.
- The agent, your content and citationsMetered on the free tier at 500 messages and 1,000 AI credits per period.
- Free
- AI Actions — call your own endpoint mid-conversation
- Starter
- Running a published process end to end
- Starter
- AI Email
- Professional
- AI Insights
- Professional
Prices are on the pricing page and nowhere else on this site, because a price typed in two places drifts. Every tier is metered — there is no tier with an uncapped message allowance, including the free one.
Read this before you buy
What the agent cannot do.
The limit: The safety gate protects the website widget and nothing else. Messages arriving by phone, WhatsApp, Telegram, email or the staff test bot do not pass through it. If that matters for your work, the widget is the channel to use.
It is not clinical, in any form
No triage, no symptom intake, no urgency detection, no emergency routing, no escalation to clinical staff. Healthcare and dental customers use EzyConn for reception work: scheduling, changes and admin questions.
The gate does not catch everything
It refuses a good share of phrasings it has never seen, and misses others. We publish no recall number for it, because any number we could print would be measured on a corpus and not on your customers.
The gate reads English
The deterministic floor beneath it recognises emergency wording in eleven languages, but the classifier above it is English-only. It is also not built to resist somebody deliberately trying to get around it.
30 languages, not more
That is the enforced list. You may have seen “50+” on an older version of this site; it was wrong, and this is the number the product actually supports.
Free-text answers are stored as written
Anything a customer types into a free-text question lands verbatim in the appointment notes. There is no consent record, no retention rule and no age gate, so do not use EzyConn to collect health information.
No certification, and no BAA
EzyConn holds no SOC 2, ISO 27001, HIPAA or PCI DSS certification and does not sign a BAA. There is no single sign-on, no data-residency choice and no self-hosted deployment.
Everything above is published deliberately. If one of these lines rules EzyConn out for you, that is the correct outcome and it cost you five minutes rather than a quarter. The full list of what EzyConn does not do is a page of its own.
Ask it something only your business knows the answer to.
That is the only test worth running. Point it at your own pages and try the question you are tired of answering.