How to prepare your content for an AI assistant

What an AI assistant actually needs#

An AI assistant needs content that is accurate, reachable and consistent with itself. It does not need your internal structure, and it does not need anything retyped into a new format.

Think of it as a new employee's first day. You do not hand them a database. You hand them the website, the price list, a few documents and a short explanation of how things work here. An AI assistant works the same way, it just reads everything in one day.

One difference matters: a new employee asks when something is unclear, and an AI assistant does not ask. What a person resolves in passing has to be resolved up front for an assistant.

Three things you must supply

• Accurate content. What it says has to be what applies today, not last year.
• One version of the truth. If a price appears in three places, all three must agree.
• Written-down rules. Cancellation terms, deadlines and exceptions have to exist on paper.

The third point is the only one that usually means writing something new, and it is one short document. The first two are tidying what already exists.

What reads well and what needs care#

Modern AI models read images too, so a scanned document is no longer the obstacle it was a few years ago. Two caveats still apply: reading an image is slower and more expensive per query than reading text, and dense data reduces reliability.

The quick check works like this. Open the document and try to select a few words with your mouse. If they highlight, it is text, and everything runs faster and cheaper. If they do not, it is an image, which the system will still read, at higher cost and with less reliability on dense tables.

Content can also be unreachable because of where it sits, not what format it is in. Anything behind a login, inside an internal system or in someone's inbox stays invisible until access is granted deliberately. Before you start, it is worth listing the places your content actually lives.

Where structure does matter

Longer lists are the exception to the rule that structure is optional. Price lists, service catalogues and equipment specifications need to be at least minimally tidy, because otherwise the boundary between items is unclear.

Take a price list with the row "Gastroscopy, endoscopy and pH test". Not even a person can tell whether that is one bundled service or three separate ones. The assistant will guess, and guessing is exactly what you are trying to avoid. It is enough for each item to have its own row and its own price, with no special table or format.

Four steps of preparation#

In this order, because each step makes the next one easier. None of it requires technical knowledge.

1. 1Collect your twenty most common questions. Go through email and messages and write down what people actually ask. Do not guess, copy them from real correspondence.
2. 2Find where each answer lives. Next to every question, note the document or page. Questions with no source are your real to-do list.
3. 3Resolve the contradictions. Where two sources disagree, decide which one is right, then fix or remove the other.
4. 4Write down the rules. Terms, deadlines, exceptions and who may approve what. One document, plain sentences, no formal language.

Step two is the one that usually opens eyes. Owners regularly discover that a third of their most common questions have no answer anywhere except in one person's head.

That list doubles as your benchmark, because once the system is running it has to answer those same twenty questions correctly. It is the simplest success test you can run.

What the rules document contains

This is the only part most companies have to write from scratch. It does not need to be long, but it does need to be specific.

• Deadlines. How late an appointment can be moved or cancelled, and what applies after that.
• Prices and exceptions. What is fixed, and what depends on negotiation and must go to a person.
• Authority. Who approves discounts, refunds and special requests.
• Limits. What the system must not offer an opinion or an estimate on.
• Handoff. Which cases always go to your team.

Write it in plain sentences, as if explaining to a new employee. Formal contract language works against you here, because it is harder to apply unambiguously. For most companies the document fits on one or two pages, and its main value is that it forces decisions that were previously avoided.

The most common mistakes#

These come up in almost every preparation, whatever the industry.

• An old price list still on the website. Nobody removed it because nobody visits it, but the system reads it as current.
• Rules only one person knows. When that person takes leave, neither the people nor the system have the answer.
• Three versions of the same document. The terms on your site, in the PDF and in the quote you email out are rarely identical.
• A price list with unclear boundaries between items. Several services listed in one row leave room for misreading.
• Content written for search engines rather than people. Pages full of keywords and short on specifics have nothing to give.

The usual scenario runs like this. A company sends over a website, two PDFs and a price list. First round of testing reveals the price list on the site and the one in the PDF have disagreed for months. Nobody noticed, because customers mostly ask by phone. The system did not create that problem; someone just saw it in one place for the first time.

None of these are technical problems. They were all confusing your customers before, and someone has now counted them systematically.

How much work this is, and who does it#

In our experience, preparation takes between half a day and two days for most small and mid-sized companies. It runs longer only when content has gone years without an update, or when sources disagree in several places at once.

The split of work is usually this. You decide what is accurate and what applies, because nobody outside your company can know that. Collecting, reading and connecting the sources is our job.

Decide up front who has the final say. Preparation usually stalls not because of the workload, but because nobody is authorised to declare which of two versions of the terms is current. Name that person before you start and a two-day job stays a two-day job.

Maintenance is the bigger problem

In our experience, the harder problem is not the initial preparation but what happens afterwards, when nobody maintains the content. With several clients the prices changed and the source the assistant reads from did not. The system gave the old price for months, exactly as the document stated it.

So decide who is responsible for pushing a price or terms change through to every place that detail appears. Without that, the same mess returns within months, this time with an assistant relaying it neatly to your customers.

The good news is that the work does not have to finish before you start. An assistant reading your website can run immediately, while rules and edge cases get added over the first few weeks. The list of unanswered questions shows you what to write down next, a point covered in the guide on stopping AI from making things up. If you are planning actions too, you also define the rules for what may be booked or rescheduled. That is covered in the guide on chatbot vs AI agent. What this looks like with a lot of content is visible on our page for AI in tourism.