Manual control in sales is an illusion of security. AI is already capable of saving hours of work for legal and accounting teams when handling documents. Roman Kislitsin, Astarus, explains the key strengths of neural networks and provides a step-by-step plan for implementing an AI assistant.

Photo: Mikhail Grebenshchikov / RBC

AI does not solve every problem, but in certain tasks it performs better than even the most attentive employee. It does not get tired, lose focus, or become distracted, and it can review a thousand documents with the same level of accuracy.
In commercial proposals, AI can quickly detect incorrect pricing, structural inconsistencies, and missing elements. Based on my experience, incorrect pricing is one of the most common mistakes among SMBs. A manager is in a rush, opens an old template, forgets to update a number, and the company loses either money or the client.
AI is also capable of analyzing correspondence. It evaluates a manager’s communication style, compares it with internal standards, and highlights deviations. It can also help identify warm leads within inbound traffic by analyzing responses, client reaction speed, and the structure of the conversation. As a result, the sales team gains a clearer understanding of where genuinely interested prospects are and where effort should be focused.
How to implement AI in sales: below is a step-by-step plan.

Step 1. Choose one process

AI-based review works best in processes that involve a high degree of consistency and sufficient volume. Employees often make mistakes in routine tasks when repetition leads to reduced attention. That is why the first step is to create a list of processes.

Choose one process from the list using three criteria:

there is a clear quality control procedure;
the volume of routine work is high, and employees spend significant time on checks;
the cost of an error is high.

Next, define the control points and determine how to gather requirements for the review process.

Example

Our company has a contract approval process. In this case, the control point is the final review before signing. We can define the following criteria: first, the contract text must remain unchanged. Second, the contract terms must comply with company requirements — that is, pricing must be current, included correctly in the contract, and deadlines must be stated accurately. These control points can be defined in advance. We can then ask AI to review documents at the control stage to ensure they meet these criteria.

We implemented AI to review commercial proposals. In your case, it could be applications, contracts, or order processing.

Step 2. Use an existing procedure template

It’s ideal if you already have a regulation or checklist for the review process. If not, you will need to create one. Don’t worry, it doesn’t have to be a 50-page document. It is enough to describe what a correct output should look like. Not an ideal or theoretical version, but the one currently accepted within the company. This is necessary so that AI can compare actual outputs against expected standards.

Example checklist for contract review

1. Formal requirements:

all mandatory fields are completed (contract number, date, party details);
counterparty details match the database records;
authorized signatories are specified for both parties;
contract validity period is stated;
all required annexes are included;
stamps and signatures are placed in the required locations.

2. Substantive terms:

prices comply with the price list or agreed conditions;
delivery/work completion timelines and acceptance procedures are correctly specified;
payment terms comply with approved standards (advance payment, deferred payment, installment options);
dispute resolution procedure is defined;
penalties and liabilities of the parties are within approved limits.

3. Change control:

the contract has been compared against the legal department’s approved reference template;
no unilateral changes made by the counterparty are present.

Step 3. Connect AI to the corporate system

This is where technical support is required — either your in-house or external developers. In practice, it’s not as complicated as it may seem. Modern AI services integrate with most popular CRMs, email clients, and document management systems. If your processes are standard and the budget is limited, it’s best to use ready-made AI agents (for example, BitrixGPT). In most cases, they will cover the majority of tasks without unnecessary costs.

You can also use cloud-based LLMs such as Yandex GPT or GigaChat. Document processing can be implemented via Yandex AI Studio or by building a custom comparison algorithm using a programming language. If there is no involvement of personal data or commercial secrets, cloud-based Western solutions can also be used—such as Perplexity, ChatGPT, Claude Sonnet, Qwen, or DeepSeek — by integrating them via API.

If your industry is highly specialized, requires complex integrations, or has non-standard requirements, you should work with developers. Yes, it is more expensive, but the system will then operate exactly the way you need it to.

Step 4. Configure validation rules

Usually, three to five rules are enough for AI to start detecting obvious deviations. The simpler, the better. For commercial proposals (CPs), for example, this could include checking prices against the price list, verifying delivery timelines, validating discount calculations, or ensuring that all company details are correctly filled in.

Example prompt for reviewing a commercial proposal

You are a sales quality control assistant. You are provided with a commercial proposal (CP) and the company price list.

Check the CP according to the following criteria.

1. Price compliance with the price list:

compare each item in the CP with the price list;
indicate any items where the price differs.

2. Calculation accuracy:

check the correctness of the total amount;
check the correct application of discounts for each item;
verify that all intermediate totals are calculated correctly.

3. Mandatory details:
presence of CP date, CP number, client name, delivery/work execution timelines, payment terms, CP validity period, and contact information.

If no issues are found, state that the CP complies with the requirements. List any detected errors separately, specifying the item and the nature of the issue. Price discrepancies and calculation errors must be marked as critical.

How to use this prompt:

1. Copy the prompt into an AI service.
2. Attach the CP and price list files.
3. Receive the review results.

Step 5. Allow time for team testing

Give employees one week to work with the system in test mode. It is important not to tailor or adjust anything specifically for the test. The AI should observe real user actions, real mistakes, and real workload.

At the end of the testing period, the manager receives statistics showing how many errors were detected, how critical they are, and which issues occur most frequently.

Step 6. Refine or scale

If the system is working well, add new rules and expand it to other processes. If there are many false positives, refine the logic. If functionality is insufficient, integrate additional modules.

In which adjacent areas AI can be useful

There are several additional areas where AI demonstrates high effectiveness, largely because errors there tend to be repetitive and predictable.

Customer service

Customer support is always challenging. People are human, sometimes they improvise, sometimes they forget, and sometimes they are simply tired by the end of the day, so scripts and procedures are not always followed consistently. AI acts as a safeguard. It analyzes support agents’ responses in real time and highlights issues such as:

responses not aligned with the approved script;
agents making promises the company cannot fulfill;
communication tone being too formal or, conversely, too casual;
missing clarifying questions that should have been asked.

Another useful function is automatic ticket categorization. AI reads the customer request and immediately determines its category: technical issue, financial request, complaint, or suggestion. It also assesses whether the request belongs to first-line support, second-line support, or requires escalation to a supervisor. This helps route tickets to the right specialist faster.

Document management

This is an area where mistakes can be costly. Miss a clause in a contract — you may end up in litigation. Enter incorrect details — payments go to the wrong account. Forget an appendix — the counterparty may refuse to sign.

AI can detect missing clauses, data inconsistencies, and dynamic changes during the approval process, as well as track unilateral modifications in contracts — something that often goes unnoticed in manual review.

For example, a counterparty may send a contract for signing where one clause has been “accidentally” changed — removing late-payment penalties or adding favorable conditions for themselves. A human reviewer may miss this, especially in long documents. AI compares versions and immediately highlights differences, saving legal teams hours of work. It also helps accountants avoid errors when processing primary documents. In manufacturing companies, AI checks labels before printing to prevent errors that could lead to costly batch rework.

Operational processes

In operational workflows, errors tend to accumulate like a snowball effect. A poorly created request ends up in the wrong queue. If priority is not set, an urgent order may be processed too late. AI identifies such cases and flags them for managers or administrators.

AI can also be configured to verify compliance with business processes:

whether all required fields are filled when a task is created;
whether process steps are followed in the correct sequence;
whether status updates are made on time;
whether required approvals are in place.

So, if a company is facing an increase in errors during scaling, it is worth starting with something simple. Choose one process, set up a basic validation check, and review the results. There is no need to build a complex system right away. And yes — AI will not replace employees, but it can effectively compensate for the human factor.

Get in touch

info@astarus.com

Get in touch