Your First 14 Days With an AI Receptionist

Day three after go-live. The owner of a painting and decorating firm sits at the laptop in the evening and listens to a test call in the dashboard. The assistant picks up, introduces itself as a digital assistant, asks what the call is about and, when the caller asks for a quote, offers a callback straight away. A good conversation. He still notes three details to sharpen: the opening hours in the prompt, the closing line in its own field, and a rule for company names. Twenty minutes later all three are in. On day 14 the summaries arrive exactly the way he wants them. A typical scenario, and the normal course of events with an AI receptionist: the first 14 days of fine-tuning rarely take more than 30 minutes a week and follow a clear pattern. This guide covers seven levers, a ten-call test plan and a 14-day plan. With rednea you edit the prompt and the knowledge base yourself in the dashboard, whenever you like.

In short: Three rules cover most of the fine-tuning of an AI receptionist in the first 14 days. On the phone, the prompt counts, so opening hours live there. Rules of behaviour go in the prompt, facts about your business in the knowledge base. And every piece of information has exactly one home. Make ten test calls to a plan, listen to the audio daily in week one, then fine-tune once a week. With rednea you edit prompt and knowledge base yourself in the dashboard, and by day 14 the assistant runs the way you want it to.

Why is the first 14 days the right time to fine-tune your AI receptionist?

An AI receptionist is as good as its prompt and its knowledge base. Both are filled at setup with what you know on day zero. The first real calls then show which facts are still missing and which phrasings you want to sharpen. Enter those points in the first two weeks and you have an assistant that runs steadily for months and only needs a touch when you launch a new service or close for the holidays.

How the assistant handles a call from greeting to summary is described in how an AI receptionist answers business calls. This guide is about the layer after setup: the levers we adjust most often when we support customers, and how you operate them yourself.

The decorating firm from the opening shows how quickly it goes (typical scenario). The owner reworked the prompt himself on day two: shorter answers, an immediate "send a photo by e-mail or request a callback" offer for quote enquiries, and a separate paragraph for job applications and emergencies. Since then the calls have been short, and the first real customer call arrived with a complete callback request in the summary: name, number, reason, preferred time. The three details from his test call are levers 1, 3 and 5 below.

Seven levers for the first two weeks

Seven settings account for most of the fine-tuning in the first two weeks. All seven sit in the dashboard, and most take a few minutes to set. The same sequence applies to each one: listen to the audio first, change exactly one thing, then call the affected scenario again.

1. Opening hours: on the phone, the prompt counts

Opening hours live in three places: in the company settings, in the knowledge base and in the prompt. What the assistant says out loud is what the prompt says. So the "Opening hours" paragraph in the prompt is your source of truth: update it with every change, including before holidays, and align the company settings and the knowledge base with it. A sentence such as "We are available Monday to Friday from 7am to 4.30pm" reads the same in all three places. Callers then hear the same hours in every call as they see on your website.

2. Rules of behaviour in the prompt, facts in the knowledge base

The prompt is read in full on every call; the knowledge base is only consulted when needed. So conversation rules belong in the prompt as short paragraphs, and facts about your business in the knowledge base. A new behaviour section, for example "Job applications", sits in the prompt right next to "Sales calls", not in an uploaded PDF. If a document contains rules, move them into the prompt and leave the facts in the document. The caller then hears the same tone from the first sentence to the last, and every rule applies the same way in every call.

3. Keep company names and surnames apart

Many callers introduce themselves with a first name and a company: "Tom from Keller Joinery" (illustrative example). One sentence in the prompt makes sure the assistant never files the company as a surname: "If someone gives their first name followed by a company, never use the company name as a surname. Address the person without a surname or ask for it." Every summary then lists the company as a company and the name as a name, and your team calls back the right person.

4. Spell back and confirm

Unusual company names and e-mail addresses are the most demanding moment in any call for any speech recognition. Two rules in the prompt keep every summary clean. First: "Spell back company names and e-mail addresses when you are unsure, and have them confirmed before you end the call." Second: "Read e-mail addresses out with 'at' and 'dot'." The caller hears the name spelled back, confirms with one word, and the spelling in the summary is right. A translation agency added these two sentences after its first test call (typical scenario).

5. One closing line, one field

The closing line has its own field in the dashboard. It belongs in that field, and then it comes exactly once at the end of every call. The prompt itself contains no goodbye. For the caller that means a clear, short close, for example "Thank you for calling, we will get back to you today." A physiotherapy clinic set that sentence for every call with a single entry (typical scenario).

6. The greeting: how callers stay on the line

Company name first, disclosure second, then an open question, and the whole thing in under ten seconds. An example: "Example Decorators Ltd, this is the digital assistant speaking. How can I help you?" The caller knows after three seconds where they have landed and after five what they can do. The disclosure also meets Article 50 of the EU AI Act, which requires callers to be told they are talking to an AI; rednea's assistant introduces itself as a digital assistant on every call, as the guide to EU AI Act disclosure on AI phone calls explains. Offer the callback early, not only at the end, and even callers in a hurry leave their number.

One option at the start of the call: some configurations ask "I can see this number is registered to X. Is that you?" That comes from a setting, not from the prompt, and can be switched off per configuration if you prefer shorter calls; a short note to rednea is enough. How callbacks and enquiries are captured is covered in how an AI receptionist qualifies leads.

7. Choose the assistant's voice and language deliberately

rednea speaks English, German, French, Italian and Spanish and switches automatically when the caller changes language, whatever you set as the default. What you choose is the sound of the answer: the default language, a clear, neutral voice and a steady pace, all set in the dashboard. A physiotherapy clinic with callers from several countries set English as the default and left the rest to the automatic switch (typical scenario). Voice and pace matter most for older callers; the guide to AI receptionists and older callers covers those settings in detail.

What goes in the prompt and what goes in the knowledge base?

The prompt governs how the assistant behaves. The knowledge base supplies what it knows about your business. This separation is the foundation for levers 1, 2 and 5: the prompt is read in full on every call, the knowledge base only when needed. Every piece of information has exactly one home, and the answer is then the same in every call.

A tip several customers use: copy the prompt into an AI chat tool and have one section rewritten, for example "shorter, friendlier, more formal". Paste only that section back, not the whole result. The rest stays stable, and the next test call shows how the change sounds.

How do you test properly?

Ten test calls over two days cover the situations that come up in everyday business and show which of the seven levers you still want to adjust. Call yourself or ask someone who has not seen the prompt. Then listen to the audio in the dashboard, not only the summary: the summary shows what the assistant understood, the audio shows how the call sounded to the caller.

Note one sentence per call. After ten calls you have three to five adjustments. Make them in one session and repeat only the affected scenarios.

The 14-day plan

The plan spreads the work over three phases and takes about three hours in total. In week one you check in daily, after that once a week. By day 14 the summaries arrive exactly as you want them, and the assistant handles the standard cases on its own.

Days 1 to 2: test and align

• Ten test calls following the table above, listen to the audio
• Align the opening hours in the prompt with the website and the company settings
• Closing line in its own field only, not in the prompt
• Add the rule for company names and spelling back

Days 3 to 7: watch real calls

• Read every summary the same day, listen to the audio whenever something is unclear
• Check the greeting in the audio: company name first, under ten seconds, callback offered early
• Check the knowledge base: facts only? Move any rules into the prompt
• Add new questions from the calls to the knowledge base as facts

Days 8 to 14: stabilise

• One fine-tuning session a week, no more than three changes per session
• After every change, call the affected scenarios again
• Confirm the default language and voice, and check the language switch on one call
• Tell your team which callbacks from the summaries they take on

After that, one fixed slot a week is enough. Two weeks of real calls also show how many AI minutes your business needs. The wider frame for choosing, setting up and running an assistant is in what an AI phone assistant is.

Key Takeaways

• On the phone, the prompt counts. Opening hours and availability live there; align the company settings and the knowledge base with it. One source of truth per piece of information.
• Rules in the prompt, facts in the knowledge base. Every rule then applies the same way in every call, and the answers stay consistent.
• One closing line, one field. In its own field it comes exactly once at the end of every call.
• Spell back and confirm before the goodbye: company names and e-mail addresses then appear correctly in every summary.
• A greeting under ten seconds: company name, disclosure, one open question, an early callback offer. Callers stay on the line and leave their number.
• Fine-tuning your AI receptionist takes ten test calls, one week of daily listening and 30 minutes a week after that. With rednea you do it yourself in the dashboard.

Frequently Asked Questions

How long does it take to fine-tune an AI receptionist?

Allow about three hours over 14 days: one hour for ten test calls and their review, around ten minutes a day in week one for summaries and audio, and one 30-minute session in week two. After that, one fixed slot a week is enough to add new enquiries to the prompt or the knowledge base.

Can I edit the prompt safely on my own?

Yes, as long as you work in small steps. Change no more than three places per session and call the affected scenarios afterwards. If you prefer the previous version, undo the last change. Before a larger rework, keep a copy of the prompt in a text document. And if you are unsure, rednea helps.

Where do I enter the opening hours so they are right on the phone?

In the prompt. On the phone, what the prompt says counts. Find the paragraph with the opening hours, update it with every change, and align the company settings and the knowledge base with it so all three places say the same thing. Check the paragraph before holidays too, and the assistant gives the right availability from the first day off.

Should I check the summary or the audio?

Both in the first two weeks, and always the audio when something is unclear. The summary shows what the assistant understood; the audio shows how the call sounded to the caller. How the greeting comes across, whether names are confirmed cleanly and how the close lands, you only hear in the audio. How the assistant handles the harder calls is covered in can AI handle complex customer enquiries?

Sources

• Information Commissioner's Office (ICO), guidance on transparency and recording calls under the UK GDPR: ico.org.uk
• Regulation (EU) 2024/1689 (EU AI Act), Article 50, transparency obligations for AI systems interacting with people: eur-lex.europa.eu
• Harvard Business Review, Oldroyd, McElheran and Elkington, "The Short Life of Online Sales Leads" (2011), on response speed and lead qualification: hbr.org
• Federation of Small Businesses (FSB), resources on customer service processes in small firms: fsb.org.uk