Don’t start with a chatbot. First, write down where in the company you keep customers, orders, prices and contracts. Then consolidate them so that each of these types of information has one main place. Only on top of data like that does it make sense to build AI, because a chatbot working with scattered information answers fluently and confidently, but not reliably.

How long the journey will take depends mainly on three things. How many places you currently keep the same information in. How many important things only specific people know and are not written down anywhere. And who the AI is supposed to answer: your own people, who will spot a mistake, or customers, who won’t.

A chatbot working with scattered data answers confidently, but wrongly

A typical situation: a customer asks about a price. The price list says one thing, a spring e-mail from the sales rep says another. And the sales rep remembers also promising them free delivery this year.

A colleague can find their way through this because they know whom to ask. AI doesn’t.

A language model can’t tell which of the three versions is valid. It picks the one it finds and presents it convincingly. That is worse than no answer at all, because nobody expects a mistake delivered in such a confident tone.

This is not a technical detail. According to Gartner, as reported by Freevacy, by the end of 2026 companies will abandon roughly three in five AI projects that lack AI-ready data. According to the same source, almost two-thirds of organizations don’t trust their own data management practices for AI.

For a smaller company, the conclusion is simple. Time invested in getting your data in order pays off with every tool you adopt, not just AI. Time invested in a chatbot built on top of a mess is often wasted.

Write down where your customers, orders, prices and contracts are today

Don’t start with technology, start with an inventory. Take the four types of information the company can’t function without. For each one, write down every place where it currently lives.

Include the awkward ones too, such as a sales rep’s personal mailbox or a notebook in the warehouse.

Information Where it tends to be scattered What to ask during the inventory
Customers contacts in the accounting software, a sales rep’s spreadsheet, phone contacts, e-mails Where is the contact person and where are the agreed terms?
Jobs and orders e-mails, spreadsheets, paper printouts, accounting software Where can I find out the status of a specific job?
Prices price list in a spreadsheet, individual discounts in e-mails, sales reps’ memory Where is the price a specific customer actually pays today?
Contracts shared folder, ring binder, e-mail attachment Where do I find the valid version and the date it expires?

One page is enough. Involve the people who work with the information every day. You yourself probably don’t know most of these places.

Readiness test: Pick one regular customer at random. Find their contact person, their latest order, the price they pay and their valid contract. Count how many places you had to open and how many people you had to ask. Wherever it was more than one place, start there.

Give each type of information one main place and one owner

The goal is not one big system for everything. The goal is for each type of information to have one place that people trust. The other places are either copies that pull from it, or they disappear.

For each type of information, decide:

  1. Where the main version lives. One place, not “mostly here, sometimes there”.
  2. Who corrects it. One person or one role, not whoever happens to notice something.
  3. When it gets recorded. For example, a new price is entered in the main place before the sales rep sends it to the customer.
  4. What happens to the old places. A read-only archive, or deletion.

If someone manually re-keys data between programs, make a note of it. Manual re-keying is where discrepancies arise. It is often also a good candidate for your first automation.

Readiness test: Ask two people from different departments where they would find the current price for a specific customer. If they point to the same place and find the same figure, the step is done.

Related solution: B2B ordering portal for wholesalers and manufacturers

Also write down what today lives only in people’s heads and in e-mails

The most valuable information is often the unwritten kind. That one customer always wants the invoice to include their purchase order number. That another customer should never be called before ten. That regular customers get a discount on a certain type of goods.

As long as this lives only in people’s heads and in sent mail, AI can’t see it. And the company loses it when someone leaves or falls ill.

How to go about it:

  • Go through the most common exceptions with everyone who deals with customers.
  • Add a place for notes and agreements to every customer record. Write things there, not in e-mails.
  • Transfer any agreement made by e-mail or phone to the main place the same day.
  • For contracts, extract the key details: with whom, from when, until when, notice period and special terms.

Readiness test: Imagine a sales rep falls ill tomorrow for two weeks. Can a colleague process an order from one of their customers with the correct price and terms using only the main place? If they have to call the person who is off sick, your data isn’t ready.

Clean up the data before you let AI near it

Consolidating into one place usually exposes the mess. The same customer appears three times under different names. Contacts lead to people who no longer work for the customer. Contracts have no end date.

The clean-up doesn’t have to be perfect. It has to be regular.

  • Merge duplicates and agree on how companies and products are named.
  • Flag records you are not sure are still valid and have them verified.
  • Decide how often the owner checks the data, for example once a month.
  • For personal data, find out what rules apply to you regarding its retention and sharing with external tools, including AI tools. If in doubt, consult a local specialist.

Readiness test: Pull ten random customer records and ten price records. Check whether each one is up to date. If you find more than a few errors, AI would repeat them to your people and your customers.

Choose a narrow first AI task for your own people

Only now does AI come into play. The first task shouldn’t be a customer-facing chatbot, but an assistant for your own people. They will spot a mistake and fix it before it gets out.

Good first tasks share three features: a narrow question, a single data source and an output that is easy to check. For example:

  • searching valid contracts: when they expire, what the notice period is, what was agreed about delivery,
  • transferring an order from an e-mail into the system, with a person just confirming it,
  • drafting a reply to a recurring customer question, with a link to the record it is based on.

A customer-facing chatbot only makes sense once internal tasks have been running for a while without major errors and you know where AI makes mistakes.

Checklist before your first AI project

  • Customers, orders, prices and contracts each have one main place.
  • Each main place has an owner.
  • Agreements made by e-mail and phone are recorded in the main place.
  • The data has been cleaned up and spot-checked on a sample.
  • For every answer, the AI shows which record it is based on.
  • You know who checks answers before they go out.

One simple rule applies. If a person can’t answer briefly from the main place without calling a colleague, AI can’t either.

Related solution: Custom software vs off-the-shelf: when each one pays off

Sources

Frequently asked questions

Do I need to have everything in one big system before I start with AI?

No. It is enough for each type of information to have one place people trust and that can be read from. Even a well-kept spreadsheet can do the job, as long as nobody copies it into other versions. A shared system starts to make sense once there are too many places and too much manual re-keying.

Can’t AI clean up the messy data itself?

Partly, yes. It can find likely duplicates, extract dates and deadlines from contracts or suggest standardized names. But deciding which of two prices is valid has to be done by a person who knows the agreement with the customer.

Where do I start if absolutely everything is scattered?

With the information people ask about most often and where a mistake hurts most. In wholesale, that tends to be customer-specific prices; in an accounting firm, client documents and deadlines. One type of information taken all the way teaches the company more than five half-finished ones.

How do I tell whether an AI project is failing because of the data rather than the tool?

Take the wrong answers and trace where they came from. If the AI quoted an old price list or an incomplete record, the problem is the data and a different tool won’t fix it. If the data was correct and the answer still wasn’t, it makes sense to adjust the tool or the instructions.

Should people stop handling things with customers by e-mail?

They don’t have to. E-mail stays for communication; it just stops serving as storage. One simple rule helps: an agreement only counts once it is recorded in the main place.