Document automation looks like the safest way into AI: the data exists, the template exists, the task is mechanical. Which is exactly why it burns people. It stops being mechanical the moment a value appears in the document that exists in no source.
The main risk here is not spelling or formatting. It is a plausible term nobody agreed to: a deadline, a price, a volume or a warranty phrased by the model rather than taken from a system.
1. Every field has an owning system
The first thing to describe is not the template but a mapping table: which field comes from where. It sounds dull, and it is exactly what determines whether the document will be trustworthy.
It almost always turns out that the same value lives in two places and differs between them. Price in the CRM and in the price list, deadline in the contract and in the ticket, volume in the accounting system and in a manager's e-mail. While a person assembles the document by hand, they pick the right one from context. Automation cannot do that and should not try: the source of record is declared explicitly, and on a conflict the winner is not whichever was found first.
2. Terms are retrieved, not phrased
The line runs in one place: the language model is responsible for the wording of the text, never for the values inside it. Prices, deadlines, volumes and warranties are substituted from the source as they are. The model may write a paragraph explaining a term — it may not invent the term.
In practice that means the template consists of two kinds of fragment: substituted and generated. The boundary between them belongs in the template description, not in the implementation's discretion.
3. A gap halts the document
Mandatory fields are validated before assembly, not after. If a value is not found, the document is not produced — instead the owner gets a notice about what is missing.
The temptation to do otherwise is strong, and it comes in three guises:
Insert a dash
Keep last time's value
Write “to be confirmed”
All three are more dangerous than stopping, because the document reaches the client looking finished.
4. Confidence threshold and manual editing
Extracting data from an unstructured source — scans, e-mails, attachments — yields not just a value but a confidence in it. Below the threshold the block is flagged and routed to a person rather than substituted silently.
In our experience the threshold is better tuned to the cost of an error than to average recognition quality: a contract-sum field and a city-name field carry different costs when wrong, and a shared threshold is always wrong for one of them.
- Data source:
- Практика Aplora · 2026-08-24
- Calculation method:
- A generalisation across the team's document automation projects, not a measurement on a sample.
- Limitations:
- An estimate, not a fact: it depends on source document quality in the specific environment.
5. A person approves commercial terms
Sending a document with commercial terms to a client unattended is not optimisation — it is transferring risk to someone who decided nothing. Internal summaries and recurring reports on an agreed template can go out without a person. Anything containing a commitment cannot.
| Document type | Delivery mode |
|---|---|
| Internal management summary | Unattended |
| Recurring client report on an agreed template | Unattended; the specialist adds the findings |
| Proposal | Only after approval |
| Contract package, invoice, accounting document | Only after approval |
What to fix before automating
Some problems are not solved by automation but encoded by it. If the same figure is calculated differently in two reports, automation will reproduce both versions faster and at greater volume. A single definition is work that has to happen before, not after.
Figures with no unambiguous definition
Templates that do not say where a value comes from
Fields filled in “as agreed” with the agreement written down nowhere
Documents with nobody accountable for releasing them
A document that cannot be verified before it goes out is not worth generating automatically — however well it turns out.
Methodology notes, by e-mail
The same material we publish here: how to model the economics of an initiative, where rollouts break, and what to verify before work starts. Once a month at most, no market news and no sales e-mail.
