AI Document Analysis
When an employee asks AI about company documents, answer accuracy depends on three things. The needed policy clause should come first. Each answer needs a two-layer check: first the category or threshold, then any exception inside the rule. If the fragment has no answer, the rule may still exist elsewhere. Skipping any step means deciding by the wrong rule.
Why Can AI Give a Wrong Answer About Company Documents?
AI does not read a document the way a person does. Answer accuracy depends, in part, on where the needed clause sits in the text.
The study "Lost in the Middle" by Liu and colleagues (version 3, November 2023) tested two task types: answering questions across several documents and finding a value by its key in a structured data file (JSON). For several models studied, accuracy was higher when the relevant fragment appeared at the beginning or end of the input, and lower when it sat in the middle. The effect showed up differently for the two task types and was not uniform across models. The study described models from 2023 and did not measure modern tools.
We use this principle as a teaching exercise. The same input is assembled three ways: the needed clause at the top, in the middle, and at the end. Then the answers are compared on the same criteria. If the answer changes with position, that is a signal: input layout and completeness matter. Loading an entire document does not guarantee accuracy on its own.
The second source of errors is caveats. Corporate policies almost always contain conditions inside rules. The model picks up the general rule and misses the exception. The employee then makes a decision based on a simplified version of the rule.
How Do You Ask AI the Right Question About Company Documents?
A good question starts with a good input. The input is everything you place before the question - a cut excerpt of a policy or a set of reformulated clauses. Answer quality depends directly on input quality.
Three rules for a good input:
- Put the needed clause at the top. If the document is long, cut the section that answers your question and place it first.
- Name the source and your check date. For example: "GitLab Global Travel and Expense Policy, checked 04.10.2026." This matters when the policy changes.
- Add an instruction for the AI. Write: "Answer only by these clauses. If the question cannot be resolved from this fragment, say so - do not describe familiar practice."
Five Questions About an Expense Policy: Good Answer vs. Bad Answer
For this walkthrough we use GitLab's public policy - the Global Travel and Expense Policy. Educational English retelling of GitLab Global Travel and Expense Policy, checked 04.10.2026; language, abbreviation, and composed questions altered; CC BY-SA 4.0. The license allows distributing this retelling with attribution to the author and source under the same conditions. This is teaching material, not a legal interpretation of the full GitLab policy.

Teaching input with five clauses:
Expense policy clauses (GitLab, checked 04.10.2026):
- A single purchase up to $5,000 inclusive can be processed through Navan (an expense service); above $5,000 the procurement process in Zip (a procurement system) applies.
- An expense older than 90 days is not reimbursed without manager approval.
- An expense description is mandatory, even if the field in the interface is marked optional.
- Software subscriptions are not reimbursable.
- Software acquisition goes through the procurement process.
Answer only by these clauses. If the question cannot be resolved from this fragment, say so - do not describe familiar practice.
Questions:
- Equipment at $5,200: which channel to use?
- An expense bought 100 days ago: will it be reimbursed?
- A software subscription at $200 per month: reimbursable?
- The description field is marked optional: can it be left empty?
- Payment for an employee course: what does this fragment say?
Compare two answers. Both are composed as teaching samples - not taken from a real model.
Bad answer:
Equipment at $5,200 goes through Navan. The expense bought 100 days ago cannot be reimbursed - the deadline has passed. The software subscription at $200 can be reimbursed, the amount is small. The description can be skipped, the field is optional. The course will be reimbursed through the training budget - that is usually how it works.
Good answer:
Equipment at $5,200: the $5,000 threshold is exceeded, channel is Zip. Expense bought 100 days ago: manager approval is required; the fragment does not confirm that approval is sufficient for payment. Software subscription: not reimbursable. Description: mandatory - the optional mark in the interface does not cancel the policy clause. Course payment: the fragment does not address this, cannot answer.
The bad answer contains five errors. Let us walk through them to see exactly where each one breaks down.
How to Check an AI Answer: The Two-Layer Method
After each point of the answer, ask two questions in sequence.
Layer 1 - general threshold or category. Did the AI match the question to the right rule? For a dollar amount: did it compare against the threshold? For an expense type: did it identify the category correctly?
Layer 2 - specific condition inside. Does this rule contain a caveat, exception, or reference to approval? If so, does the answer reflect it?
Here are the bad answer's errors mapped to this method:
| Question | Error | Type |
|---|---|---|
| Equipment $5,200 | Amount not compared to the $5,000 threshold; routed to Navan instead of Zip | Layer 1 |
| Expense, 100 days | Manager approval condition lost; unconditional refusal given | Layer 2 |
| Software subscription $200 | Decision made by amount, not by expense category | Layer 1 |
| Optional field | Interface label taken as the rule instead of the policy clause | Interface vs. document |
| Course payment | Invented answer "through training budget" not present in the fragment | Fabricated answer |
Equipment $5,200. The employee stopped at "expensive, therefore Navan." Layer 1 error: the amount was not compared to the $5,000 threshold. Check the numerical comparison first - the channel depends on it.
Expense 100 days. The general rule is visible: 90-day limit. But inside it there is a caveat - an expense older than 90 days can still be reimbursed with manager approval. The bad answer lost that caveat and delivered an unconditional refusal. This is a Layer 2 error: the category and threshold were identified correctly, but the special condition inside the rule was missed. Also: the fragment confirms that approval is necessary but does not guarantee that approval is sufficient for payment.
Software subscription $200. The bad answer reversed the checking order: it used the amount instead of the category. The rule about software subscriptions has nothing to do with price. Even if the subscription cost $1, it would still not be reimbursable - the category "software subscription" is excluded. Category of expense first, then amount.
Optional field. Form fields describe the interface, not the rules. The optional mark in the form does not cancel clause 3 of the policy. The document and the interface say different things here, and the document takes priority.
Course payment. The bad answer invented "usually reimbursed through the training budget." That is not in the fragment. An answer is built only on what is in the input; what is absent is an unresolved question, not an invitation to guess.
The good answer passed both layers for every point: each was matched to a specific rule, not to a dollar amount or familiar practice.
Sometimes a condition is built from two clauses at once. Here is a second teaching example using the same input clauses:
Same policy clauses (GitLab, checked 04.10.2026), clauses 1, 2, 4, 5.
Questions:
- A peripheral device at $4,800, bought 95 days ago: which channel, and is there a condition?
- A conference ticket at $2,000, bought 30 days ago: does this fragment give a right to reimbursement?
Composed answer:
The device at $4,800 falls within the general threshold of clause 1, so Navan is a permissible channel. This is not confirmation that the category is reimbursable. Bought 95 days ago, clause 2 requires manager approval. Approval and guaranteed payment are different things; the fragment does not confirm the latter.
Conference ticket: the question of the right to reimbursement under the attached clauses remains unresolved. General thresholds and the 90-day limit do not confirm a right to reimbursement for this specific category. Next step - find the travel or training policy, identify its owner, and direct the question there.
Check against four criteria: $4,800 is below the $5,000 threshold, Navan is permissible as a channel; 95 days exceeds 90, approval is required; the device category does not match software or subscriptions; the fragment has no clause about conferences.
Key point: both rules apply at once here. An employee who sees "$4,800, therefore Navan" and stops misses clause 2.

What to Do With Each Type of Error
When an answer turns out to be wrong, first identify the type of error. Each type has its own next step.
Practical error. The employee composed an incomplete input, did not compare the amount to the threshold, or missed a caveat. Rewrite the question and check the answer again with the two layers. Adding "compare amounts to thresholds, check each rule for caveats, say so if a clause is absent" to the instruction is helpful but does not close all errors at once. Rewriting the instruction is a proposed practice, not an experimental finding from the Liu study. So after the new answer, the employee checks it again with the same two layers.
Document error. The policy itself contains a contradiction, an outdated clause, or two documents say different things. The document owner fixes this. The employee goes to them with a specific quote and the clause number.
Systemic error. The AI consistently misses what is in the loaded text, or loads the wrong document. This goes to IT. Check the input and wording first. Only if they are correct should you look for a systemic cause.
For the first example - the five questions - the employee's action list looks like this:
- Equipment $5,200: submit through Zip; Navan does not apply.
- Expense bought 100 days ago: request manager approval, knowing that the fragment does not guarantee payment.
- Software subscription: do not include.
- Description: fill it in despite the optional mark.
- Course payment: the fragment says nothing; check with the document owner.
Absent from the fragment and absent from the policy are two different things. If the needed section is not in the input, open the full policy and look there. If there is no such section at all, write a targeted question to the policy owner - for example: "The expense policy has no section on course payment. Is there a separate training policy?" Find that section, move it into the input, and ask again. Do not draw conclusions on your own.
How to Train a Team to Check AI Answers
The session structure follows a familiar cycle: task, question, answer, check, next action. Start with a public policy, as in the examples above. Then the employee takes documents from their own function and checks answers with the two layers on their own.
After the second example, the trainer gives the employee a task with the same input clauses but new numbers and one unresolved question. The employee formulates the question themselves, receives an answer, and notes which clauses they checked and where doubts remain. Repeating a prepared answer teaches recognizing someone else's analysis. Independent checking teaches acting: finding the right clause, matching an amount to a threshold, noticing a caveat. These are different tasks.
For feedback, keep three things: the employee's question, the answer received, notes on the check. These show where the person verified against a clause and where they simply trusted.
Example: an employee indicated Navan as the channel for a device at $4,800 bought 95 days ago but did not mention approval. The manager notes: "Clause 2 was missed - recheck the 90-day condition." The employee's new answer includes a request for approval. Approval and guaranteed payment are different things; the next step is to clarify the submission procedure with the policy owner before obtaining approval.
The readiness criterion is practical: the employee asks a question on their own documents with a full input, checks the answer with the two layers, and correctly routes the error to the document owner or IT. There is no fixed number of sessions.
Before moving to closed documents, ask IT three questions: which service is approved for your data, which document categories are agreed, and what access it has. If there is no simple answer to these questions, it is too early to start with closed documents. Anonymizing a document does not replace this approval. A correct answer on GitLab's public policy proves nothing about AI search over your own internal regulations: someone else's tidy policy and your internal documents with duplicates and outdated versions create very different working conditions.
Questions Before Launching Training
After a demo on someone else's policy: can we immediately work with internal regulations?
No. The teaching example explains the checking method - it does not prove the employee is ready for real company documents. First agree with IT on which service is approved for your data and what storage conditions are agreed. Anonymizing a document does not replace that approval. Only after the IT agreement should you move to exercises on the function's own documents. A correct answer on a public input does not mean your closed regulations will give the same result.
Documents are scattered across folders and duplicated. Where to start?
With the same practice. The exercises include tasks to select the right document from several similar ones and to find the current version. An ideal knowledge base is not required before the start. What is required is the habit of naming the document and the check date in every question to the AI. The document owner confirms the current version, because the publication date and the date a rule takes effect can differ.
How much time and money does it take to train employees to work with AI and documents?
Volume and cost are discussed with each client separately, because the documents, team structure, and starting level differ everywhere. Readiness is checked not by the number of sessions, but by how the employee performs on their own documents: composing a full input, checking in two layers, and correctly routing an error to the document owner or IT. The plan contains no fixed number of sessions.
Who is responsible for an error in an AI answer about company documents?
It depends on the type of error. Practical - the employee's task: rewrite the input and check the answer again with the two layers. Document - the document owner's task: fix the contradiction or the outdated clause. Systemic - IT's task: investigate why the tool consistently misses the needed fragment. The employee's trained task is not to guarantee a correct model answer, but to notice the error and route it to the right specialist.
Skeptics and believers: is the checking cycle the same for both groups?
The method is the same for both groups. Two-layer checking requires only one thing: compare each point of the answer to a specific rule in the document. A skeptic checks because they do not trust. Someone who trusts also checks because the cost of an error is high - a wrong procurement channel or a denied reimbursement. The same effect for both groups has not been measured specifically, but the method applies to either. You can start with the GitLab public policy from this article.
What is a "caveat" and why is it so easy to miss?
A caveat is a condition inside a rule that limits or extends its scope. Example: the clause "an expense older than 90 days is not reimbursed" contains the caveat "without manager approval." If the AI reproduces the general rule but not the caveat, the employee receives an incomplete answer. That is precisely why the second layer of checking exists: after confirming that the category and threshold are identified correctly, check whether the rule contains a condition that changes the conclusion.
How do you know an answer is missing from the fragment, not from the policy itself?
These are two different situations, and confusing them is dangerous. If the input contains only part of the policy, a missing answer means only that the needed section was not included in the fragment. To confirm that no rule exists at all, open the full policy and check every section. If there is no such section, write a targeted question to the document owner: "The expense policy has no section on course payment. Is there a separate training policy?" Do not draw conclusions on your own.
What if the policy changed after the employee received the answer?
An old answer is not rewritten retroactively. An answer received on a given date remains the answer based on the policy version loaded at that time. After an update, ask the question again with the new version of the document. The check date in every question is not a formality - it is a label: in a dispute over interpretation or amount, the document owner decides, not the model's answer.
How to Apply the Method to Your Own Company Documents
Asking AI about company documents saves time only when the answer is checked. Putting the right clause first, checking each point with the two layers, and routing errors to the right specialist - these are skills learned on the real documents of your own function.
To discuss training on your company's documents, see our AI training program. Questions about getting started: hello@majento.ai or Telegram https://t.me/shimaoz.