All questions

Yes, in a natural Hungarian voice, and it switches to English or German if the caller starts that way. It knows your own texts and prices, not generic answers.

It politely hands over to a colleague or promises a callback and sends the call summary immediately. It never guesses a price or a deadline.

Typically one lost job. The missed-call SMS goes out within 5 seconds with a booking link and 30–40% of callers book from it.

On the website, Messenger, WhatsApp and Instagram the same knowledge base answers, and the booking or order starts from the conversation.

Missed-call SMS and the booking chatbot in 1–14 days. The full phone receptionist with voices and calendar from day 36.

Yes. Twenty minutes, no obligation. You leave with a written scope and a fixed price whether or not we end up working together.

A first working version usually goes live in two to three weeks. Larger systems are split into phases so something is running early and paying for the rest.

It stays in your systems. We connect to your tools with your own credentials, process data in the EU wherever possible, and sign an NDA and data processing agreement before we see anything.

Whatever you already use: Google Workspace, Microsoft 365, HubSpot, Pipedrive, Billingo, Számlázz.hu, Shopify, WooCommerce, Slack, WhatsApp and most things that have an API.

A fixed price for the build, then a monthly fee that covers hosting, monitoring and improvements. No hourly surprises.

No. We need someone who knows the process. We handle the technical side and train your team before go-live.

Each automation carries a typical range of hours saved per week for the people doing that work by hand, taken from the 2026 AI Automation Market Report: email handling 2–3 hours a day, invoice admin 6–12 hours a week, quoting 5–10 hours a week, scheduling 15–25% of working time. The configurator sums the midpoints; the discovery call replaces them with your real numbers.

We do not publish list prices. Every project gets a fixed price after the free 20-minute call, then a flat monthly fee for running it. No hourly billing.

They follow the report's 35-day roadmap: days 1–14 for cheap, visible quick wins such as missed-call text-back and the review funnel; days 15–35 for the automations that give back daily hours such as inbox triage and booking reminders; day 36 onward for larger systems such as voice agents, invoice processing and lead generation.

No. Tabs only group the catalog; you can tick items from any tab and filter by function or level across all of them. The AI treats the whole selection as one package.

You get a summary, a tier, phases with every selected item placed, suggested additions and a next step. You can copy the list, share a link that preselects the same items, or send it with your details from the contact form. Nothing is stored on our side beyond the request.

Cost is driven by three things: how many processes you automate, how many systems have to be connected, and how much clean-up your data needs. We price a scoped project, not seats, and the first call gives you a concrete figure. Payback is measured in hours: one automated function typically returns 2–15 hours a week.

Yes, and usually sooner than at a large company, because in a small team one person carries several roles. A typical five-person business loses 6–12 hours a week to admin. Recovering even half of that is a full working day of capacity without hiring. Start with one process, measure it, then extend.

Most first projects pay for themselves within three to nine months. The clock starts at go-live, which is 1–14 days for a simple automation, 15–35 days for a connected workflow, and 36+ days when several systems and approvals are involved. Payback is counted in hours recovered per week, not in licences.

Four factors, in order of weight: the number of systems that must be connected, whether those systems have a usable interface, how clean your data and documents are, and how many exceptions the process has. The AI model itself is the smallest line. Scope one process and the figure becomes concrete on the first call.

Take one process, count how many times it runs per week and how many minutes each run takes. That gives weekly hours. Multiply by the loaded hourly cost of the person doing it. Compare that to the build plus monthly running cost. Most single functions land between 2 and 15 hours a week.

Running cost has three parts: consumption-based model usage, channel costs such as telephony minutes or SMS, and monitoring. It scales with volume, not with headcount, and you set a cap. For most SMEs it is a small fraction of the salary cost of the work the automation replaced.

Start where work arrives, not where you imagine AI is exciting. List the three channels that bring you the most repeated work, usually phone, email and web enquiries, and automate the one with the highest volume and the clearest rules. That is a 1–14 day project and it produces the numbers for everything after it.

A one-hour process walkthrough, not a demo. We follow one piece of work from the moment it arrives to the moment it is finished, writing down every click, every decision and every exception. That walkthrough produces the scope, the baseline hours and a go-live date in one of the 1–14, 15–35 or 36+ day bands.

No. You need someone in your company who knows the process and can approve decisions, typically two to three hours in the first week and under an hour a week afterwards. The building, the connections and the monitoring are ours. Day-to-day use happens in tools your team already has: email, chat and your calendar.

Single-channel automations go live in 1–14 days: missed-call text-back, email triage drafts, instant replies to web enquiries. Workflows that connect two or three systems take 15–35 days. Anything touching invoicing, an ERP or a multi-step approval chain takes 36+ days. You get a date in writing at scoping.

The one with the highest weekly volume, the clearest right answer and the lowest consequence when it is wrong. In most Hungarian SMEs that is inbound enquiry handling: calls, emails and web forms. It runs dozens of times a week, a mistake costs a correction rather than money, and it frees 2–15 hours a week.

Three things, none of which takes more than twenty minutes: a rough weekly count of the repetitive work, a list of the systems the work passes through, and ten real examples including two that went wrong. With those we can give you a scope and a go-live band on the call itself.

It picks up after the second ring, in natural Hungarian, using your own opening hours, prices and services. It understands what the caller wants, answers, books into your calendar and sends a confirmation. Anything outside its scope is transferred to a colleague or logged with a callback request. It runs 0–24.

Every incoming email is read, categorised and routed to the right person or queue, with spam and newsletters filtered out. For the categories you approve, it writes a reply draft that is 90% finished, using your own price list, policies and past answers. A person reads it, adjusts a word and sends.

It takes the invoice from email, a scan or a photo, reads the supplier, number, date, net and gross amounts, VAT rate and line items, then checks them against your order and delivery note. Matching invoices are prepared for booking; anything that disagrees goes to a person with the difference highlighted.

The colleague records what the job involves in their own words, straight after the site visit. The AI extracts the items, quantities and conditions, prices them from your own price list, and produces a branded quote with three tiers. From speaking to a sendable document takes about 3 minutes.

Your documents are indexed by meaning, not by keyword. When someone asks a question, the assistant searches the index, retrieves the two or three relevant passages and answers using only those, with a link to the source. If nothing relevant is found, it says so instead of inventing an answer.

Within 5 seconds of the missed call an SMS goes out: who you are, that you saw the call, and a question that starts a conversation. The customer replies by text, the answer lands with your team, and the booking or quote continues in writing. Most replies arrive before the caller has dialled a competitor.

Yes, when three rules are followed: the automation gets the narrowest access that does the job, every action is logged, and business data is processed under an agreement that forbids using it for model training. Most incidents come from over-broad permissions, not from the AI, which is why scope is the first design decision.

It is designed to be wrong safely. Every output has a confidence level; below the threshold the case goes to a person instead of being sent. Anything with money or legal weight is suggest-only, so a human approves before it leaves. Every action is logged and reversible, and corrections become training examples.

Yes. GDPR does not ban AI; it requires a lawful basis, a data processing agreement with every processor, transparency towards the person, and data minimisation. In practice that means a signed agreement, a line in your privacy notice, masking fields the automation does not need, and a defined retention period for transcripts and logs.

Your business data stays in your own systems. What the automation keeps is a working layer: the search index, the logs and the queues, hosted in the European Union by default. Model processing can be pinned to EU regions, and the choice of location is written into the agreement before anything is connected.

No. Business processing runs under agreements that exclude training on your content, and that exclusion is written into the contract rather than assumed. Your documents are used to answer your questions and are kept in an index only you can query. The difference matters, and it is the first thing to check in any supplier's terms.

The work does not disappear, it falls back. Failed items stay in a queue, a person is notified, and the manual route still exists because nothing was deleted during the rollout. Monitoring alerts us rather than you, and most failures come from a connected system changing, not from the AI itself.

Automate the repetitive part first, then decide whether you still need the hire. A new colleague costs recruitment, three months of training and a salary every month; an automation takes over 2–15 hours a week of routine within weeks. In most cases the answer is both: automate the routine, hire for judgement.

For simple, single-step flows, yes, and it is often the right call. The limit appears at three points: error handling when a step fails, cases the rules did not anticipate, and maintenance when a connected system changes. Self-built flows usually work for months and then break quietly, which is the expensive part.

If your input always arrives in the same fields and the decision is a fixed rule, use simple automation: it is cheaper, faster and more predictable. You need AI when the input is free text, speech or a document, or when the decision requires reading context. Most real processes are a mix of the two.

A general chat tool is excellent for drafting and thinking, and every team should use one. It is not enough when the work has to happen without a person opening a browser: answering the phone at night, reading yesterday's invoices, replying to a web enquiry in minutes. Those need a connected, automatic system.

Automate around them first. A system replacement takes months, disrupts everyone and rarely fixes the actual pain, which is usually the manual work between systems. Automation gives you the hours back in weeks, and the connections you build stay useful if you do replace something later.

Below roughly fifty employees, an outside partner is almost always faster and cheaper, because one internal hire cannot cover integration, model work and monitoring at once. Build internally when AI becomes part of what you sell, not just how you operate. Many companies start outside and bring the running of it in-house later.

Mainly around the phone and the calendar, not the treatment. An AI receptionist answers 0–24, books and reschedules, and sends reminders that cut no-shows. Missed calls get an SMS within 5 seconds. Together these typically give a practice back 8–12 hours a week of front-desk time and fill cancelled slots from a waiting list.

In the two places money leaks: calls missed while on a roof, and quotes written at midnight. An AI agent answers and qualifies every call, missed ones get an SMS in 5 seconds, and a voice note from site becomes a three-tier quote in about 3 minutes, which lifts average order value by 20–30%.

Start with „where is my order”, which is 40–50% of all support contacts. An assistant reads the order number from chat, email or phone, checks the shop and the courier, and answers with the live status. Add returns handled from a photo, abandoned cart recovery and product descriptions generated from supplier data.

On the document intake, where most of the hours go. Client documents arrive as photos, scans and emails; AI reads them, extracts the fields, sorts them by client and period, and flags what is missing before the deadline. Bank statements are matched to invoices automatically, leaving only the genuine exceptions.

Outside the machines, in the paperwork around them. Enquiries and drawings arriving by email are read and turned into structured quote requests, shift and quality documentation is dictated instead of written, and production data becomes a weekly report with anomaly alerts. That typically frees 6–12 hours a week of office time.

At order intake and at collections. Orders arrive as free-text emails, spreadsheets and PDFs; AI reads them, matches the items to your product codes and creates the order, flagging anything ambiguous. On the finance side, overdue invoices are chased on a schedule with the right tone per customer, which shortens payment times.

In three places: quotes produced from a voice note in about 3 minutes, site diaries and handover documents dictated instead of typed, and crew scheduling that reacts when a job overruns. Together they usually return 8–15 hours a week to the owner, most of it evening and weekend time.

It takes tasks, not people. In a small company the work that gets automated is the part nobody wanted: retyping, sorting, chasing, answering the same question for the ninth time. Our clients typically redeploy the 2–15 hours a week into customer work rather than cutting headcount, mostly because they were already short-staffed.

About 30 minutes per person for daily use, plus one shorter session two weeks later once real questions have appeared. There is no new interface to learn: approvals happen in your existing email or chat. The process owner needs more, roughly two to three hours in the first week, because they set the rules.

Start with the process that annoys them most, not the one that saves you most. Let the person who does the work define what „correct” means, and show the weekly numbers openly. Adoption fails when it is announced and works when the first version visibly removes something the team already complained about.

We monitor it and fix what breaks; you own the rules and the priorities. Monitoring alerts us, not you, and most incidents are a connected system changing something. On your side one named person reviews the weekly numbers and the exception queue, typically under an hour a week.

Not the way you work with customers, and usually not your systems either. What does have to change is anything that exists only in someone's head: the rules for deciding, the exceptions and the priorities have to be written down. That is one to two hours of work and it has value regardless of the project.

Answer the real question, which is almost always „is my job safe”. Say it plainly before the project starts, then show them the first version on their own work so the fear becomes a concrete opinion. Involve the person who does the task as the one who defines what correct means.

Four numbers, reviewed weekly: items handled without a person, items sent to a person, how often an automatic output had to be corrected, and hours saved against the baseline you measured before go-live. If the correction rate is falling and the automatic share is rising, it is working.

No, and it is configured not to. It answers hours, prices, address, parking and booking; anything clinical gets a callback from a person.

Calls are disclosed as AI, recordings are kept only as long as you define, and every provider signs a data processing agreement. Health data is a higher-risk category, so we run a DPIA with you before go-live.

Yes; it is tuned on your own price list and confirms extracted quantities back to you before the quote goes out.

The missed-call text-back is the report's number one exactly for solo tradespeople: it replaces the receptionist you do not have.

Billingo, Számlázz.hu, Kulcs-Soft, RLB60, Odoo and SAP through their interfaces; older desktop systems via file import or a browser agent.

The AI reads and proposes; a bookkeeper approves. Nothing is posted without a human click, and that is written into your engagement letter.

No. It is a checklist-driven pre-read that flags clauses for a lawyer to review; the lawyer decides.

Documents are processed under a DPA with EU residency where available, never used for training, and access is scoped per matter.

With Cal.com, Calendly, Google Calendar and most salon systems that expose a calendar or an API.

Into Google Calendar, Calendly, Cal.com or a planner with an API; otherwise it books into a shared calendar you mirror.

Shopify, WooCommerce, Unas, Shoprenter, eMAG and Alza; GLS, DPD, Foxpost and Packeta on the courier side.

Yes, with a dedicated technical login and two-factor handled; every upload is logged with a screenshot.

Fewer than on voicemail. Callers stay when they get an immediate, useful answer and are told they can reach a person.

Yes: the agent takes calls on no-answer, after hours or when the line is busy, and warm-transfers when needed.

Not when grounded in your price list and told to say 'I don't know' outside it; we also log every conversation and review samples weekly.

Known spam numbers are filtered and no text is sent; unknown numbers get the message, which costs cents.

In SMEs it usually removes the tasks nobody wanted and frees people for revenue work. The report's figure is 30–45% of staff time spent on robot work.

Only if it also scores high on frequency and clarity. A painful but vague process is a bad first project; fix an easy, frequent one first and return to the painful one with credibility.

A fixed price that pays back within roughly three months from the measured manual cost. If the numbers do not support that, the scope is wrong, not the price.

Yes, with a warm transfer that passes the caller's name and reason, or by scheduling a callback if nobody is free.

Yes. Calls are forwarded to the agent on no-answer, after hours, or always, and the number stays yours.

The build can be, but maintenance usually is not: every interface update risks breaking the bot. Over a year, API-based AI automation is typically cheaper to run.

Handled the same way. The model reads German, English or Romanian invoices and returns the same field set; currency is captured and converted according to your accounting policy.

Yes, for judgement and compliance. What changes is that they stop typing and start reviewing, which is a better use of their hourly rate.

Often yes if it records calls or processes health, financial or other sensitive data at scale. For a booking agent at a hair salon, usually no, but documenting the assessment is still good practice.

Missed-call text-back and the review funnel. Both are single-workflow projects in the report's starter price tier: 200,000–500,000 HUF setup and 30,000–80,000 HUF a month.

The first, third, fifth and eighth were built for exactly that: tradespeople, doctors, consultants, salons. They replace a receptionist you do not have.

Mixed. The report blends international practitioner data (Reddit, YouTube, LinkedIn) with Hungarian agency figures; pricing and tool notes are specific to Hungary.

Automations run on servers, call paid AI and telephony APIs, and need someone watching for failures and tuning prompts. The retainer buys that; without it, systems silently rot.

We do not publish list prices; every project is quoted at a fixed price after the discovery call, with a flat monthly fee for running it.

Make to prototype fast without a server; n8n once volume grows or data must stay on your own EU server. Many projects start on Make and move to n8n.

Yes, if you have the bandwidth to review results. The order matters more than the calendar: cheap and visible first, then time savings, then the bigger builds.

Shopify, WooCommerce, Unas, Shoprenter and eMAG/Alza marketplaces via their APIs; the courier side covers GLS, DPD, Foxpost and Packeta.

Yes; that is the report's core example. The agent answers while your hands are busy and books into your calendar, for a fraction of a receptionist's cost.

No. A chatbot is one possible interface. AI automation covers any task where a model reads, decides or writes on your behalf, including back-office work nobody chats with.

A scoped first version typically goes live in two to three weeks, provided the process is documented and the tools it touches have APIs.

It can, which is why we give agents the narrowest permissions that do the job, log every action, and route low-confidence cases to a person before anything is written.

Yes. We recommend and configure a clear disclosure at the start of the call. It is required under the EU AI Act for AI systems interacting with people, and callers respond better when told.

Yes. The agent detects the caller's language in the first sentence and continues in it, or you can pin one language per phone number.

Website chat, Messenger, Instagram, WhatsApp Business, SMS, email and phone. The same assistant logic can serve several channels with channel-specific formatting.

Almost never. Workflow automation works with what you have. Replacing a tool is a separate decision, usually made later and for other reasons.

Not where legacy software has no API. It is now usually the last mile of a workflow rather than the whole solution.

Not with the commercial APIs we use, whose terms exclude customer data from training. We confirm this per provider in the data processing agreement.

As often as your source changes. We typically sync from a shared drive, Notion or your CMS on a schedule or on each edit.

Yes. Modern models read images directly, including phone photos of receipts, though sharper scans give higher accuracy.

Current engines do, including ő and ű, on reasonable-quality scans. Handwriting remains the weak spot.

Options in order of preference: a database connection, file import and export, or RPA that drives the interface. All three work; they cost more to maintain than an API.

When signed and verified, yes. We validate the signature on every incoming webhook and reject anything unsigned.

We write it so it is honest and useful: it confirms receipt, asks one or two relevant questions and names the person who will follow up. Most prospects prefer that to silence.

No, but they can be made rare and harmless: rare through grounding and scope, harmless by never letting the model act on uncertain information without a check.

For document processing, 2 to 8% after the first month. For support triage, 15 to 30% depending on how narrow the scope is. The share falls as rules are tuned.

It can be kept in the EU with the right provider settings, and where it cannot, standard contractual clauses apply. We set this per project and record it in the DPA.

Something a person would write: 'Hi, I'm with a customer and couldn't pick up. How can I help?' The AI then reads the reply, asks for the address and urgency, and offers two calendar slots.

For the first reply, yes, because it is automated. The human follow-up can come later; what matters is that the prospect hears back immediately.

Cal.com or Calendly for booking, Twilio or WhatsApp Business for messages, n8n or Make for the logic, and your existing calendar.

Yes, through the WhatsApp Business Cloud API with a verified business profile, which also lets you send order confirmations and tracking links.

The report's pattern is day 2, 7, 14 and a 'break-up' note on day 25, stopping the moment the prospect replies or books. 80% of deals close between the 5th and 12th touch, and most humans stop after one.

The invoice is held, the finance owner gets a one-line alert with the discrepancy, and approves or rejects with one click.

Yes. A technical user is created in the Online Számla portal and its credentials are used by the automation; no special licence is required.

The agent books a callback in the first free slot, confirms it by SMS, and logs the brief in the CRM.

Whisper and Deepgram handle Hungarian technical vocabulary and background noise well; the AI confirms extracted quantities back to the technician before the quote is sent.

PandaDoc, DocuSign, and for Hungarian legal weight the AVDH service; all have APIs the automation can drive.

For existing customers under legitimate interest, generally yes, with a clear opt-out. Any STOP reply blocks further messages automatically.

Asking everyone for feedback and offering the public link is fine; selectively blocking unhappy customers from Google is against Google's policy. The funnel asks everyone and simply adds a private channel for problems.

Yes; sheets are synced into a small warehouse first so questions run fast and consistently.

Handled with a dedicated technical account and an authenticator or SMS code the automation can read, agreed with the portal owner.

It is trained on your own past posts and speaks in your register; the human approval step catches anything off. Keyword stuffing and generic phrasing are explicitly avoided.

Wherever the owner already looks: WhatsApp, Slack, Teams or email. Phone-friendly, no dashboard to open.

Yes. Monitoring, retries and alerts are part of running the system, not an add-on.