Blog · Published · Updated
AI Agent for Code Review: Evidence
Short answer
Automated code review is a first layer of checks on a pull request, run by an AI agent, for example in Claude Code, before a person looks at it. The agent takes style, dead code, missing tests and departures from conventions. Architecture and business logic stay with the reviewer. At 60 pull requests a month the model costs about €4.36, and the reviewer time given back is worth €419.
- net for implementation for a team of up to 10 people, paid once
- from €1,150
- from start to work of the entire team
- 2 weeks
- after this time, we measure the effect on the same tasks
- 30 days
- conversation with an engineer and a written plan within 2 business days
- €0
We belong to the OpenAI and Anthropic partner programmes and learn through their official courses. What this means ↓
Author: Syntalith · 3 min read
Statistics
AI in software development: statistics 2026
€4.36
a month for the model when Claude Sonnet 5.5 reviews 60 pull requests. 12 reviewer minutes on each are worth €419. Scenario.
40%
According to Menlo Ventures (2025), Anthropic holds 40% of large companies' spending on language model APIs and OpenAI 27%, based on a survey of about 500 US decision makers.
Source: Menlo Ventures, State of Generative AI in the Enterprise (2025) ↗
15.1%
According to Eurostat (2025), 15.1% of people aged 16 to 74 in the EU used generative AI tools for work.
32.7%
According to Eurostat (2025), 32.7% of people aged 16 to 74 in the EU used generative AI tools in the previous 3 months.
14–56%
According to Anthropic (2025), studies with a control group found that AI cut task time by 14% to 56%.
Source: Anthropic, Estimating AI productivity gains (2025) ↗
Figures checked: 22 October 2026
What an AI agent really catches in a pull request
It helps most where the rule is clear and checking by hand is tedious. An agent can report a wrong comment with full confidence, so its output is material for a person to judge.
- Typos and inconsistent names of variables and functions
- Dead code and unused imports
- Missing or shallow handling of errors and edge cases
- Missing or superficial tests of new logic
- Departures from the repository's conventions
- Obvious vulnerabilities: a secret in the code, no validation of input data
What stays with the reviewer
| Review layer | Who leads | Examples |
|---|---|---|
| Mechanical | AI agent | Style, typos, dead code, missing tests |
| Repository conventions | Agent and person | Consistency with patterns, naming |
| Decisions | Person | Architecture, API design, domain logic, security |
A linter and static analysis check deterministic rules and invent nothing. The agent works one layer higher: it will notice that a test checks only the happy path. That is why they work best together.
How automated code review works in practice
- 01Opening the PRThe agent starts automatically and leaves comments: to fix or to consider.
- 02AuthorFixes the obvious things and dismisses false alarms without explaining.
- 03ReviewerGets a PR without the noise and focuses on architecture and logic.
- 04MergeApproved by the person responsible for the outcome.
The agent judges the pattern, and a person judges the intent. On unusual code it is wrong more often and sounds just as sure. Without written rules, the repository gets conventions from other projects.
What automated code review costs
- Reviewer time on the mechanical layer today
- 15 min per PR
- After the rollout, reviewing the agent's comments
- 3 min per PR
- Time given back
- 60 × 12 min = 12 h (€419)
- Model: 30,000 + 2,000 tokens per PR, Claude Sonnet 5.5
- €4.36
Left
€415 a month
The times and the token counts are the scenario's assumptions. We count at €34.90 an hour, the EU average hourly labour cost for 2025 according to Eurostat; a developer's hour usually costs more, so the result is careful. Model at Anthropic's rates, converted at the ECB rate of 1.1205 USD per euro on 22 October 2026.
The model is the smallest item in the bill. People's time is the largest, so the agent makes sense where the team reviews a lot of code. With a few pull requests a month the result will be small, and the cost of setting up and writing down the conventions stays the same.
When automated review is not worth rolling out
Automated review makes no sense without written conventions, because the agent has nothing to follow and produces noise. Also skip it when the team treats the agent's comments as an oracle, or with one small repository and a few pull requests a month, where review by hand is enough.
An example of noise: the agent asks for a test of a function that on purpose has no logic worth testing, or reports missing error handling for an error handled one layer up. These are no disasters, but they show why the author must have the right to dismiss a comment without explaining.
Data and the contract
Code is company data. Check which plan the model runs on, whether the vendor trains models on the company's code and which contract covers the repository. Keep API keys and secrets out of the context that goes to the model. On the business plans of Anthropic and OpenAI, business data does not train models by default, but you check the retention and connection settings separately.
How to introduce automated review
The team has no written conventions
first write down the repository rules
€0
The team wants to work with Claude Code every day
a Claude rollout for the team
from €1,150
The team wants to learn agentic work on its own repository
AI-Native course
priced after a call
Before you start, answer four questions: do you have written conventions, which review layers repeat the most, who approves the merge, and which plan and contract the model runs on. We describe a Claude Code rollout in a team in the article Claude Code rollout for teams.
„I already use AI.”
Your people too. Just no rules.
From our conversations with companies
5 na 10 employees paste sensitive data into private AI accounts: contracts, invoices, customer data.
In ESET survey (2026)
1 na 10 admits this in the survey.
Customer data pasted into a private AI account without an data processing agreement is a violation of the GDPR.
Don't worry, we won't tell anyone :)
What do we see in companies that "already use"
- Only ChatGPT or Claude accounts, without plugins, skills, integrations and MCP servers with company tools.
- Company programs are not connected to AI. People still click through systems and Excel, and enter chat data manually.
- Lack of company standards: everyone writes instructions in their own way, no one shares what works.
- No rules for personal data, GDPR and company secrets.
- No limits and no visibility: no one knows who uses AI, how much and for what.
62% Polish employees use AI at work · 27% companies have written rules for working with AI.
Source: ESET and DAGMA, Cyberportrait of Polish business 2026 (March 2026, 1,026 people) ↗
ChatGPT or Claude space review
Accounts, permissions, rules for working with data and 5 tasks set for the team. 1 day.
€450 net
Company accounts, rules of working with data and team training in 2 weeks, from €1,150 net.View implementation for the team →
„Is my data safe?”
Yes. You choose where they go: from your company account to the model on your server.
00Employee's private account
Customer and company data ends up in an account that the company cannot see or control. There is no data processing agreement, so it is a violation of the GDPR. When an employee leaves, the chat history goes with them.
01ChatGPT or Claude corporate account
The provider does not use your data to train models. A copy of the queries remains in his logs for up to 30 days.
02API with zero retention (ZDR)
The provider processes the request and does not save anything. It requires his consent, we prepare the application together.
03Processing in the EU
Queries are counted and stored in Europe: OpenAI in the EU region, Claude via clouds with the EU region.
04Model at your place
The open model runs on your server or in your cloud in the EU. The data does not go outside the company.
What we do with every implementation
- Confidentiality agreement and data processing agreement before we receive any data.
- You choose the cases for the prototype and we anonymize them upon request.
- The data stays in the EU and we delete it after the test.
- Corporate accounts with an administrator on your side.
When data cannot leave the company
The open model runs on your server or in your cloud in the EU. To choose from:
- White-tailed eagle
- Polish model, SpeakLeash and ACK Cyfronet AGH
- PLLUM
- Polish model, HIVE consortium commissioned by the Ministry of Digitization
- Qwen
- Alibaba
- GLM
- Zhipu AI
- DeepSeek
- DeepSeek
We select a model for the task and check it on your cases. We run it, tune it and maintain it.
Source: Bielik and PLLuM, ACK Cyfronet AGH ↗ · PLLuM, Ministry of Digitization ↗
Do you have a lot of your own data? We will fine-tune the small model.
- 1 Your examples
- 2 tuning
- 3 test on new cases
- 4 model at your place
When you have thousands of similar emails, documents or reports, we teach a small open model based on your examples. It works for you, responds quickly and costs less to maintain than a large model. We measure quality on issues that the model has not seen while learning.
Przemysław and Piotr did it in large programming and consulting companies. Artem did it before and does it today at Syntalith. It's about models on your own infrastructure, tuning and quality measurement.
You choose the protection level during the interview. We sign the data processing agreement before we see any data.Schedule a free analysis →
Price
Three packages. You know the price right away.
Team of up to 10 people
from €1,150
net, paid once · 2 weeks
Company, from 11 to 40 people
from €2,300
net, paid once · 3–4 weeks
Organization, over 40 people
from €4,600
net, paid once · 6–8 weeks
- The same packages once you have licenses. To start, an inspection of the space in 1 day is enough: €450 net.
- You buy Claude Team licenses directly from Anthropic, without our markup: USD 20 per person per month for annual payment or USD 25 for monthly payment, from 2 to 150 people.
From conversation to measurement
01 · 30 min
Free process review€0
An engineer picks 3 to 5 of your team's tasks with you. Written plan within 2 working days.
02 · 2 weeks
Rolloutfrom €1,150
Accounts and roles, rules for working with data, training on your team's tasks. You buy licences directly from the provider.
03 · after 30 days
Measurementincluded
We compare task times before and after. We name in writing the tasks worth handing to a system.
or call+48 888 78 48 78WhatsAppMon–Fri 9:00–17:00 CET
Worked example · Scenario
How we count the payback: an IT team in Claude or ChatGPT
We run the same count for your developers' tasks in the free analysis.
€3,627
a month for moving agreements into tickets in a team of 8
Payback: 1.2 mo.
3 to 4 hours a week per person for moving agreements from emails and meetings into tickets is our assumption. A 30% cut sits within the range of controlled studies (MIT: 40%, Harvard and BCG: 25%). €34.90 an hour is the EU average hourly labour cost for 2025 according to Eurostat, for companies with 10 or more employees. Licence: ChatGPT Business or Claude Team at 20 USD per person a month on annual billing, about €17.85 at the ECB rate of 1.1205 USD per euro on 22 October 2026. Rollout: from €1,150 for up to 10 people, from €2,300 for 11 to 40 people.
See the full calculation →Questions
What does an AI agent catch in a pull request?
Typos and inconsistent names, dead code and unused imports, missing handling of errors and edge cases, shallow tests, departures from the repository's conventions and obvious vulnerabilities, for example a secret written into the code.
Will the agent replace the reviewer?
No. The agent checks rules and patterns. Architecture, API design, business logic and code that is critical for money and security are judged by the person who approves the merge.
How much does automated code review cost?
At 60 pull requests a month, with 30,000 input tokens and 2,000 output tokens each on Claude Sonnet 5.5, the model costs about €4.36. A Claude rollout with Claude Code for a team of up to 10 people costs from €1,150 net at Syntalith.
When is it not worth rolling out?
When you have no written conventions, because the agent has nothing to follow, when the team treats its comments as an oracle, or with one small repository and a few pull requests a month.
