How to train an AI on your business data - a complete guide

What “training an AI” actually means for a business owner

The phrase “train an AI on your business” sounds like something engineers do - weeks of work, gigabytes of data, a model being shaped from scratch. That's not what's happening here. When you train Seekadu (or any modern customer-service AI) on your business, you're not building a model. You're feeding an existing language model a knowledge base it can look things up in, plus a set of instructions on how to sound and when to stop talking.

There are three pieces, and only three. One: a knowledge base - your menu, services, hours, policies, FAQs - that the AI searches every time a customer asks something. Two: a system prompt - a short paragraph that tells the AI what business it works for, what voice to use, and what it must never do. Three: handoff rules - explicit triggers that route a conversation to a human.

That's it. The model itself doesn't change. The same underlying language model that serves a Lisbon bistro also serves a Sofia dental practice - but each one feels completely different to its customers, because the knowledge base, voice, and rules are different. The technique behind this is called retrieval-augmented grounding: the AI retrieves the relevant slice of your business data at the exact moment it needs to answer, then composes a reply using only that slice. The benefit: it answers correctly even with 300 menu items or 200 dental procedures. The catch: garbage in, garbage out - if your knowledge base is vague or missing, the AI will sound vague or, worse, make things up.

Owners who get this right treat training as a one-day project plus a weekly 15-minute review. Owners who get it wrong either dump everything in at once with no structure (and get a confused AI) or skip the voice and handoff pieces entirely (and end up with a chatbot their regulars hate). This guide walks the right path.

The 5 categories of business data your AI needs

Almost every local business has the same five categories of data an AI customer-service layer needs. The names differ - a restaurant has a menu where a salon has a service list - but the structure is the same.

1. Services or products, with prices

The single biggest category. For a restaurant: every dish, with its price, allergen flags, and any “sold out by 21:00” constraints. For a salon: every service, with its duration, price, and the stylist tiers (junior vs senior). For a dental practice: every procedure, with price range, duration, whether it's covered by insurance, and whether it needs a prior consult. For a fitness studio: every class, with the cap, duration, and instructor. If a customer can pay for it or book it, the AI needs to know about it. AI booking depends on this category being correct.

2. Hours, location, and capacity data

Opening hours by day. Holiday closures. Last-seating time for restaurants. Last-appointment time for clinics. Class schedules and caps for fitness. Address, parking info, public transport. Any “closed for staff training on the first Monday of the month” exceptions. The AI uses this to refuse impossible slots before it tries to book them - which is half of what stops it from making things up.

3. House policies

Cancellation rules. Deposit rules. Refund rules. Late-arrival rules. Group-size rules. Pet policy. Children's policy. Plus the small ones every business has but never writes down: the corkage fee, the “no walk-ins on Saturday” rule, the 48-hour notice for groups over six. Each of these becomes a short paragraph in the knowledge base - and the AI quotes them verbatim when a customer asks.

4. FAQs - the questions you answer 40 times a week

Every local business has 10-25 questions that come in every week. “Do you take walk-ins?” “Do you have parking?” “Is there a vegetarian option?” “Can I bring my dog?” “Do you accept insurance X?” List them. Write the answer once. The AI answers them 40 times this week and 40 times next week.

5. Brand voice samples

The category most owners skip - and the one that makes the AI stop sounding like a chatbot. Three to five sample replies in your own voice, covering the most common situations: confirming a booking, declining a slot, answering a price question, handling a complaint. The AI mirrors the cadence, formality, and warmth. Without these samples, you get a corporate-chatbot voice by default - and regulars will notice.

What to upload - and what to type by hand

A useful rule for the first hour of setup: anything that already exists as a file gets uploaded; anything that lives in your head gets typed in. The split is what most owners get wrong - they try to upload things that aren't written anywhere, or they re-type things they could have uploaded in 30 seconds.

Upload

Menus (PDF, photo, Google Doc - anything readable). Service catalogues from your booking software. Hours from your Google Business profile. Price lists. FAQ pages you've already written. Any printed signage with policies (a photo of the corkage sign by the bar counts). Any standard-operating-procedure docs. For SaaS teams: your product docs, API reference, changelog, plan-tier matrix - the AI runs RAG against them with full URL citation. Full playbook at the dedicated AI chatbot for SaaS companies money page, with onboarding cadences handled by SaaS messaging automation and CRM enrichment at CRM for SaaS.

Seekadu reads PDFs, images, spreadsheets, and pasted text. The AI parses the structure (dish name → price → allergens, or service → duration → price), extracts what it needs, and lets you confirm. You'll spot a missing allergen or two; you fix them inline and move on. The mechanical work - typing 60 menu items into a form - is the kind of work nobody should do anymore.

Type by hand

Voice samples. Escalation rules. The cancellation policy you've said out loud a thousand times but never written down. The things you'd tell a new employee in their first 20 minutes - “we never confirm allergies over chat, the chef always walks the diner through it”, “late arrivals over 15 minutes lose the slot on Saturdays”, “we never refund a deposit, we reschedule”. These don't live in a file because you've never had to write them. Take 20 minutes and write them now.

The honest test: if a new hire could walk into your business on Monday and figure something out from a document, upload that document. If they'd only learn it by being told, type it in.

How to capture and document your house policies

Most owners have policies; very few have them written down. The AI can't enforce a rule it doesn't know about - and customers who need an answer in 30 seconds won't wait for a manager to be paged. Twenty minutes of writing here saves you from 40 awkward conversations a month.

The corkage example

A restaurant's corkage policy is the classic case. You charge €15 per bottle, €30 for a magnum, and you waive it for regulars on their birthday. Three sentences. The AI can quote them verbatim, calculate the fee, and waive it when the customer mentions a birthday - but only if you've written it. The unwritten policy (“manager's call”) means the AI either refuses to answer or, worse, invents a different fee.

The 24-hour cancellation rule

“Free cancellation up to 24 hours before the appointment. Inside 24 hours, the deposit is forfeited unless the slot is re-filled.” Two sentences. Now the AI can confirm a 22-hour cancellation, charge the customer in line with policy, and re-offer the slot - without paging you. Write the policy in plain English; the AI handles the rest.

The deposit policy

“Groups of 8+ require a €25-per-cover deposit, charged at booking. The deposit is credited against the bill. Cancellations inside 48 hours forfeit the deposit.” Three sentences. The AI takes the deposit, confirms the booking, and applies the rule without asking. Without this paragraph, the AI either avoids taking deposits at all (and you lose the protection) or invents a number (and you have an angry customer).

The pet policy

“Well-behaved dogs welcome on the terrace; no dogs inside. Service animals welcome everywhere.” This is a 30-second policy that handles a question that comes in every week - and the kind that, when answered wrong, ends in a one-star review.

The pattern is the same for every policy: write it like you'd say it to a new hire, including the exceptions. The AI follows what's written, not what's assumed.

Voice training - the 20 minutes that make or break the AI

If you skip one section of this guide, don't skip this one. Voice is the difference between an AI your regulars don't notice and a chatbot they complain about. The good news: it's 20 minutes of work, and the leverage is enormous.

Write 3-5 sample replies in your own tone

Pick the situations that happen most often: confirming a booking, declining a sold-out slot, answering a price question, handling a late cancellation, replying to a compliment. Write the reply yourself, in the voice you'd use on WhatsApp at 8pm. Don't be polished - be real. The AI mirrors what you write, including the small things: do you use exclamation marks? “Hey” or “Hi”? First name or no name? Emoji or no emoji?

The corporate-chatbot anti-pattern

The default voice - the one you get if you skip this step - is corporate chatbot. “Thank you for contacting us! We appreciate your enquiry and will respond shortly.” Nobody writes like this. No restaurant owner, no salon manager, no dental receptionist talks like this in real life. When customers see this voice, they know instantly they're talking to a robot - and they bail.

Contrast: “Saturday 8 is fully booked I'm afraid - would 7:30 or 9:15 work?” That's a real reply. It assumes the customer is busy, gets to the point, offers alternatives, sounds human. Same information, completely different feel. The AI can write either - it depends on the samples you give it.

Voice rules in the system prompt

Beyond the samples, write 4-6 explicit rules into the system prompt: “Never start a reply with 'Thank you for your enquiry'.” “Use first names if the customer used a first name.” “Keep replies under 60 words unless the customer asked for detail.” “Match the customer's formality.” These rules apply to every reply, every time - and they catch the edge cases the samples miss.

Setting handoff rules - when the AI must call a human

Handoff rules are the rails. They protect you from two failure modes: the AI confidently answering something it shouldn't, and the AI dodging a human-only question. Three tiers, written explicitly.

60-second escalation - interrupt whatever you're doing: severe allergies (“anaphylactic”, “EpiPen”), complaints (“rude”, “raw”, “refund”), anything mentioning illness from a previous visit, press enquiries, legal threats, and any message containing the words “manager” or “owner”. The AI sends the full thread to your phone within 60 seconds and stops the conversation on its end. Public reputation damage compounds fast; speed matters more than the wording of any reply.

30-minute escalation - get to it within half an hour: corporate enquiries, bookings over 8 covers (restaurants) or 5 services (salons), private-dining requests, group bookings, deposit disputes, refund requests inside policy, requests to change a booking on the day. The AI captures the details, confirms it has them, and tells the customer a manager will be in touch - then your phone pings.

No escalation - the AI handles it: standard bookings inside your routine size, routine allergen and dietary questions where the menu data answers cleanly, hours and location questions, FAQ-level questions, cancellations inside your policy window, gift voucher enquiries. These are the conversations the AI is for. Don't write handoff rules for them - you'll drown.

The principle: anything that affects revenue at scale, anything with a public-reputation risk, and anything with a medical implication goes to a human. Everything else, the AI handles - and a grounded AI handles it well.

Testing your AI before going live

Don't go live without 30 minutes of testing. The bugs you catch in testing are free; the ones a customer catches cost you a review. Run six test conversations - every business has the same six.

1. The standard booking

“Table for 4, Saturday 8pm” (restaurant). “Cut and colour, next Tuesday afternoon” (salon). “6pm yoga, Thursday” (fitness). The AI should capture the details, check availability, confirm. If it stumbles here, you've got a knowledge-base structure problem - fix the catalogue.

2. The edge-case allergen or sensitivity question

“My daughter is anaphylactic to peanuts - can she eat the pad thai?”. “I'm pregnant, can I do the deep tissue?”. “I have a heart condition, is the spin class safe?”. The AI should not answer with a yes or no. It should acknowledge the medical context, log it as a critical note, and route to a human. If the AI confidently answers, tighten the handoff rules immediately.

3. The complaint

“The lamb was raw last night.” “My colour came out wrong.” “Your trainer was 20 minutes late.” The AI should not apologise on your behalf or offer a refund. It should acknowledge, log the complaint, and route to a human within 60 seconds. Test that the routing actually fires.

4. The after-hours enquiry

Send a message at 23:30 from your personal phone. The AI should reply within the configured window - usually under a minute. If it doesn't, channel routing isn't connected. Fix it before going live.

5. The out-of-scope question

“What's the weather tomorrow?” “Do you know a good plumber?”. The AI should politely decline and stay on topic. If it tries to help, you've got a system-prompt problem - add an explicit “stay in scope” rule.

6. The group booking

“Looking for somewhere for 18 people for a 40th birthday in June.”. The AI should capture the date window, the size, the budget, the dietary requirements - then route to a human within 30 minutes. Test that all four details get captured and the routing fires.

Common training mistakes

1. Uploading too much

Owners new to AI training tend to dump everything - five years of old menus, archived event flyers, holiday party PDFs from 2022. The AI gets confused, retrieves stale information, quotes a price from two years ago. Upload only what's current. Archive the rest. The knowledge base is a working document, not a library.

2. No voice training

The single biggest leverage skip. Owners assume voice is “a nice-to-have” - it isn't. Without voice samples, your AI sounds like every other chatbot, and regulars notice instantly. Twenty minutes of writing sample replies has more impact on customer perception than any other hour of setup.

3. Vague policies

“We're flexible on cancellations” isn't a policy - it's a feeling. The AI can't act on it. Write the exact rule: “Free cancellation up to 24 hours before. Inside 24 hours, the deposit is forfeited.” Vague policies generate vague answers, and vague answers generate customer disputes.

4. No handoff list

Owners who skip the handoff rules end up with one of two problems. Either the AI escalates everything (and your phone is ringing constantly), or it escalates nothing (and you find out on Tuesday that a customer with a severe allergy got an ungrounded answer on Saturday). Both are setup failures, both are fixable in 15 minutes.

5. Training once and forgetting

The most expensive mistake long-term. The AI is correct on day one and slowly drifts: prices change, hours shift, new services launch, holiday closures get added, the menu gets a new section. If you never review, the AI quotes last quarter's price and a customer arrives angry. The fix is the weekly 15-minute maintenance loop in the next section.