“We need to clean up our data first, and only then think about AI” is a phrase we hear at every third assessment. It sounds reasonable, which is exactly why it survives so long. In reality, what usually sits behind it is a project that runs a year to a year and a half and costs tens of millions of tenge: a data warehouse, an integration bus, data quality standards. After a marathon like that, only a handful ever reach the automation itself: the budget and the enthusiasm run out first.
You don’t need any of that for an agent prototype. You need something else, and you almost certainly have enough of that something already.
Where the myth of “perfect data” came from
It’s an inheritance from classic analytics. Reports and dashboards really do need a complete, verified database: one missing column and the numbers lie. An agent doesn’t need that. An agent learns from examples: here’s the input of a process (a call, a scan, a request), and here’s the correct output (a filled-in record, a posted document, a reply to the customer). Thirty to fifty such pairs give you enough material to configure a scenario and honestly measure quality.
The difference is fundamental. Order in the data is a property of the whole database; examples for a scenario are a small, specific sample. The first takes a year to assemble. The second takes half a day.
How much data do you need for a prototype?
It depends on the scenario. Some reference points from our practice:
| Scenario | What you hand over | How much |
|---|---|---|
| Sales call review | Call recordings + a deal export from CRM | 30–50 calls, deals for the quarter |
| Customer replies in chats | Conversation history + a knowledge base or guidelines | 100–200 dialogues |
| Source documents in accounting | Scans of acts, invoices, waybills, including handwritten ones | 20–30 sets |
| Medical documentation | Anonymized visit records, chart templates, the ICD-10 reference from the health information system | 30–40 cases |
| Minutes and action items | Meeting recordings + samples of finished minutes | 5–10 meetings |
The general principle: examples of the input plus examples of the correct result. Almost any format works: audio, PDF, photos, spreadsheets, exports. No neat structure required.
Where all of this already sits. Telephony keeps call recordings for months by default. Scans of source documents settle in email and network folders. The health information system remembers visits, the document management system keeps document versions, the CRM holds deal history. In two years of assessments we haven’t met a company that couldn’t find material for at least one scenario; the real question is usually not “do we have the data” but “who’s going to go and export it.”
What to do if the data is “dirty”?
Nothing special: hand it over as is. Dirty data isn’t a blocker, it’s a diagnosis of the process, and often it becomes the agent’s very first task. An example from a recent assessment: in a distributor’s export, 68% of deals had no amount, because the reps weren’t filling in the field. The company saw this as a reason to postpone automation. It turned out the opposite: in the prototype, the agent pulled the amounts from the call recordings and completed the records. That very dirt turned into a measurable effect in the first week.
If your database has duplicates, gaps and free-for-all naming, that’s an argument for automation, not against it. Cleaning by hand what the agent will maintain afterward is pointless: within a quarter the database gets dirty again, because the underlying causes haven’t gone anywhere.
What’s actually missing most often
Not data. Examples of correct decisions. In most companies “what’s correct” isn’t written down anywhere: it lives in the head of a strong employee. The chief accountant spots a suspicious invoice by eye; the senior operator senses when it’s time to move a subscriber to retention. That’s why, in the prototype week, we plan for two or three short calls with the process owner: their answers are worth more than any export.
A sufficient starting set = 30–50 input examples + 10–20 samples of the correct result + one person who can answer the question “what’s correct.”
How to hand over data without falling out with security
The rules we work by: data doesn’t leave the company’s perimeter, the prototype lives in an isolated environment, only the setup team has access, a confidentiality agreement is signed before we start, and after the demo the data is deleted with a formal record. For medicine, anonymization is added: the prototype doesn’t need patients’ names or ID numbers, the clinical part of the case is enough. Every step the agent takes is recorded in an action log, so the security team always has something to check. We covered the full security model in a separate article.
What it looks like in real life
Tuesday, 2:00 p.m., the accounting department of a distributor in Almaty. The chief accountant is skeptical: “half of the warehouse waybills are handwritten, your AI won’t be able to digest that.” Together with our engineer, in 40 minutes they put together a sample: 28 sets of acts, invoices and waybills, nine of them handwritten. An IT specialist was needed for 20 minutes to open access to the network folder. Four days later the prototype processed 26 of the 28 sets, and for two it honestly answered “not sure, a human is needed,” and it was precisely those two that had real discrepancies in the amounts. The chief accountant’s skepticism ended with a question: when can we connect the accounting scenario to the live database?
Why “order first, then AI” is an expensive mistake
Because order doesn’t hold without a consumer. You clean the database, and within a quarter it’s dirty again: people still don’t have time to fill in the fields. The agent changes the mechanics themselves: data gets filled in at the moment of the event, when a call ends, when a document arrives, when an event triggers the process. Order becomes a side effect of automation, not its precondition. It’s more profitable to start with a prototype on what you have: one week and you see the quality on your own examples, free and with no obligations.
Frequently asked questions
Do we need to label the data before handing it over?
No. Labeling in the classic sense, thousands of manually labeled examples, is not required. It’s enough to show where the inputs are and what the correct result looks like; the rest is our team’s work during the setup week.
Our legacy system has no API. How do we hand over an export?
For a prototype, a manual export to a spreadsheet or PDF is enough, and your specialist can do it in a couple of hours. In production, the agent will be able to work with such a system directly through its interface, just like an operator, with every step recorded in the log.
Can we hand over personal data?
For a prototype we ask you to anonymize it: names, ID numbers and contacts are usually not needed for the scenario. In the production perimeter, the agent works with personal data inside the company’s infrastructure, under the same access rules that apply to employees.