Conversation watcherour own inbox
0messages watched
Of the conversations we dropped in 12 months, the watcher flagged about a third. It reads the same records an agent would. our own system · Aug 2026
Business systems · Custom AI agents
Custom AI agents built around your business.
An agent that reads your calendar, price list and customer records, answers the questions that repeat, drafts what needs a person and hands over the rest. It says it is an assistant, it routes custom quotes, payments and complaints to a person, and it belongs to you. From €500 for a basic assistant.
What it is made of
A conversation, three tools, and one exit to a person
Watch a message travel. Hover a node to hold it.
Agent · illustrative conversationreading
says it is an assistant · drafts by default · money and complaints always to a persona slot question: calendar checked, reply drafted
What you get
An assistant using approved sources
Approved hours, prices, services and connected records form the intended knowledge sources. We test source use and missing-information handovers; the model can still make mistakes, so sensitive answers require review.
Rules in writing
What it may answer, what needs review and which actions are excluded: payments, complaint resolution, health and legal advice. We define permissions, approval controls and a named owner for escalations.
Your accounts, your logs
The model, the messaging platform and the hosting are on your accounts. Logs record the agreed events and handovers, with access, personal-data handling and retention defined for the deployment.
Who this is for
Businesses where the same twenty questions arrive every day
Clinics and salons
Approved administrative information, published prices and appointment availability. Clinical questions and recommendations require a qualified person; the assistant is not a substitute for that review.
Shops and e-commerce
Availability, delivery, returns, order status from the store data. Complaints go to a person with the order history.
Services with a price list
Workshops, cleaning, repairs: what it costs, when you can come, what to prepare. A quote stays with a person.
Anyone with a messenger inbox
Supported Instagram, WhatsApp or website-chat connections, with account access checked first, agreed response rules and summaries for human handovers.
Proof
Run for ourselves; no client deployment published yet
We say this plainly because the alternative is a fake case. What we run on our own studio, with sources.
Drafts, not sendsrule
0messages sent by our systems on their own
Every outbound message from our own systems is a draft a person approves. The same default ships in every agent we build. our own rule · since Aug 2026
Client portalhosted by us
0requests on 25 threads, 0 failures
The cabinet an agent would write to, under load. load test · Aug 2026
What could go wrong
And what is built in against it
It answers wrong
We restrict the intended sources, validate important outputs and review realistic test cases. Models can still produce wrong answers; the controls, source material and escalation rules are improved when failures are found.
It pretends to be a person
The interface identifies the assistant and provides a clear route to a person. We test disclosure behaviour and avoid presenting generated responses as a human conversation.
It talks about money or health
Those actions are excluded from the approved autonomous scope and require a person. Access restrictions and approval controls support the handover; we test them rather than relying on a prompt alone.
The bill surprises you
Provider costs are estimated from expected usage. We configure alerts and available limits, document any gaps and show how to pause the workflow; a budget alert alone does not necessarily stop billing.
- Says it is an assistant. In the first message and whenever asked.
- Recording notice. Any recording requires an agreed lawful basis, appropriate notice and a retention policy reviewed for the deployment.
- Handover to a human. Emergencies, complaints, money and health go to a named person with a summary.
- Drafts by default. Sending on its own is switched on only per process, in writing.
- How we handle data, drafts and AI disclosure →
Price
From €500 for a basic assistant.
Build scope fixed in writing after a short call. Provider charges are estimated separately and depend on usage.
Define the agent by the work it is allowed to do
An AI agent can combine a language model with business information and tools, but the useful scope is usually narrower than the label suggests. The important question is which task it should complete and under what conditions. An assistant that helps qualify an enquiry has different responsibilities from one that changes a booking or updates a customer record. We define those responsibilities before choosing the model or interface.
Answering from approved information
A first version might answer questions from an approved service description, collect missing details and prepare a summary for a person. The information needs an owner and a clear update process. If the price list changes, the assistant should use the current version. A persuasive answer based on an old document is still a wrong answer.
Taking a permitted action
Taking an action requires a further boundary. Reading available appointment times is different from booking one, and drafting a reply is different from sending it. We specify which actions can run automatically, which need approval and which are excluded. This gives the implementation a testable contract instead of an open instruction to help the customer however it can.
Choose a useful first scenario
The best first scenario has a clear benefit, accessible information and a limited cost of error. It should occur often enough to evaluate but not require the system to solve every unusual situation. A narrow assistant can be expanded after the team understands its behaviour; an overly broad first version makes it difficult to know why it succeeded or failed.
A controlled internal assistant
An internal assistant can help staff find an approved policy, summarise a conversation or draft an answer. A person remains responsible for using the result. This makes it possible to observe errors and refine the knowledge sources before putting the assistant directly in front of customers. Internal use still requires access controls and appropriate handling of business information.
A customer-facing assistant
A customer-facing assistant needs a clear introduction, understandable limitations and a dependable route to a person. It should avoid asking the customer to repeat information already supplied. The handover needs an owner, a summary and an expected next step. Displaying a message that a human will help is not useful unless the business actually receives and follows up on it.
Prepare the knowledge and the tool boundaries
Curate the source material
Collect the documents the assistant is allowed to use and remove contradictory or outdated versions. Organise information around the questions customers ask, including exclusions and conditions. A long folder of unreviewed files is not automatically a reliable knowledge base. We test whether the relevant answer can be found and whether the source supports the complete statement being made.
Limit available operations
Give the assistant only the tool access required by the scenario. A service enquiry assistant may need to read a calendar but not delete appointments or change billing settings. Credentials and permissions belong in the system configuration, not in a conversation prompt. Important constraints should be enforced by the surrounding application rather than relying only on the model to remember them.
Define exception handling
Define what happens when information is absent, a connection fails or the request is outside the approved scope. The assistant can ask a clarifying question, provide a limited answer or hand over. A failed lookup should not be replaced with a plausible guess. The user should also have a direct way to request a person without arguing with the automated system.
Test difficult conversations before launch
Ambiguity
Use realistic questions with incomplete details, spelling mistakes and several topics in one message. Check whether the assistant asks only the information necessary to continue. Include different languages if they are part of the scope. The test set should reflect actual customer behaviour rather than a demonstration script designed to make the assistant look competent.
Unsupported requests
Ask about a service that is not offered, a price absent from the approved list and a request to ignore its instructions. Inspect whether retrieved documents contain text that could be mistaken for commands. Prompt injection and misleading source content need technical and procedural controls; a sentence telling the model to be careful is not a complete defence.
Actions and recovery
For actions, test duplicates, timeouts and partial failures. A repeated message should not create two bookings or send two confirmations. Review the approval path and the record left for staff. Keep a manual fallback and an accessible way to pause the assistant. Recovery is part of the service design, not an optional concern after the first error.
Evaluate usefulness and running costs together
- Quality
- Useful quality measures depend on the task: supported answers, correctly collected fields, appropriate handovers and successful permitted actions. Review a sample of conversations with the person who understands the business. A fluent tone can hide factual mistakes, so evaluation should compare the answer with the source and expected behaviour rather than assess how natural it sounds.
- Work saved
- Measure the work the team still performs. If staff must rewrite every response or investigate frequent errors, apparent automation may not save time. Record review effort and the kinds of questions that repeatedly require a person. Those findings may suggest better documentation, a narrower scope or a simple rule instead of a more elaborate agent.
- Operating bill
- The running bill can include model usage, hosting, messaging services and connected tools. Estimate it using expected conversation volume and realistic message length, then monitor actual use. Costs can rise when the assistant makes repeated tool calls or processes large documents. Spending limits and alerts should match the business’s tolerance for unexpected usage.
Scope a pilot you can accept or stop
Bring examples, including awkward ones
Send several representative enquiries, the approved information needed to answer them and a description of what should happen next. Remove unnecessary personal details from examples. Name the person who will review behaviour and decide which actions are acceptable. That gives us enough context to propose a first scenario with clear boundaries.
A basic assistant starts at €500; integrations, knowledge preparation, additional channels and complex permissions affect the scope. The build fee and ongoing provider charges should be separate in the proposal. Voice, payments or sensitive workflows require their own assessment and should not be assumed to be included because the service is called an agent.
Agree the acceptance cases, the pilot audience and the criteria for expansion before launch. A first version may remain draft-only while the team checks its results. If it does not produce enough value, the business should be able to stop it and retain the documented workflow and materials. A pilot is useful when it makes that decision clearer.
Questions
Questions owners ask about AI agents
AI agent vs chatbot: what is the difference for a small business?
Chatbot describes a conversational interface; it may use rules or a language model. An agent typically adds tool use and a process for pursuing a task. The practical difference is what the system may read and do, how those actions are checked and when a person takes over. A simpler chatbot can be the better fit for a narrow question set.
How much does an AI agent cost per month to run?
The build is from €500 for a basic assistant on a site or a messenger; the running cost is not paid to us. It is the AI model, the messaging platform and the hosting, billed to your own accounts, with costs depending on usage and the providers selected. We estimate provider charges before you decide, document the assumptions and show how to pause the assistant. Existing subscriptions and charges already incurred remain payable.
Can an AI agent act as a receptionist and answer the phone?
Messaging channels can include website chat, Instagram, WhatsApp and email, subject to supported integrations and account permissions. For voice, we build it only after the recording notice, the handover rules and the data agreement are in place, and we do not put a voice agent first on any page until a client deployment has run long enough to be honest about it. Today the voice part is an audit and a plan, not a switch.
What happens when the agent does not know the answer?
We define and test fallback and escalation rules, including missing information and out-of-scope requests. Models can still make errors, so consequential actions need technical controls and appropriate review. The receiving person gets the available context, and logs help identify repeated failures.
How are customer disclosures and data requirements handled?
The deployment needs a review of the data processed, the purpose, access, retention, disclosures and any recording. We document the proposed behaviour and implement approved requirements; the business obtains appropriate legal review for its situation before launch. A generic assistant setup does not establish compliance for every use.
How do call recording, disclosure and handoff work in practice?
Disclosure: the first message says it is an assistant, and it repeats that if asked. Recording: a notice at the start of a call with the reason and the retention period, texts you approve. Handoff: a written list of what always goes to a person, with a named person and a fallback. All three are part of the delivery and of the trust page.
Send us your twenty most common questions.
We show which ones an assistant can answer from your facts, and what it costs to run.