Blog · Published · Updated
AI for developers: choosing a working mode
Short answer
When you choose AI for developers, decide which part of the work you hand to the tool: a suggestion while writing a function, a conversation about existing code, or an independent change across many files. Each mode needs a different kind of check, and a good result in one mode does not settle the choice for the whole repository.
- 1:1 course or for a team of programmers
- Let’s talk
- work on the participants' repository
- Claude Code, Codex
- talking to an engineer about the scope
- €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
Developers and AI: statistics 2026
0.5 mo.
payback of a €1,150 training day for 8 people saving 2 hours a week each: €2,418 of time recovered against €143 of licences a month. Scenario.
55.8%
According to an experiment by Peng et al. (2023), developers with GitHub Copilot finished a task 55.8% faster than the control group.
Source: Peng et al., The Impact of AI on Developer Productivity (2023) ↗
19%
According to METR (2025), 16 experienced open-source developers took 19% longer with early-2025 AI tools, although they expected a 24% speed-up.
Source: METR, Early-2025 AI and experienced open-source developers (2025) ↗
14–56%
According to Anthropic (2025), AI cut task time by 14 to 56% in studies with a control group.
Source: Anthropic, Estimating AI productivity gains (2025) ↗
65.7%
According to Eurostat (2025), 65.7% of EU companies in computer programming, consultancy and information services used AI.
Source: Eurostat, AI by NACE activity (isoc_eb_ain2) (2025) ↗
Figures checked: 22 October 2026
According to an experiment by Peng et al. (2023), developers with GitHub Copilot finished a task 55.8% faster (study). According to METR (2025), experienced developers in repositories they knew well took 19% longer with AI. The difference comes from the task and from the checks someone has to run.
A suggestion while writing a function
the developer reads every line
at once
A conversation about existing code
the developer checks the conclusion in the code
with each question
An agent changing many files
a review of the changes, tests and a limited scope
after every change
Match the mode to the checks your team can run
With suggestions, the developer picks each fragment of code as they go. In a conversation, they need to check whether the explanation matches the real files. An agent can take more actions, so the review also covers the commands it ran, the files it changed and how the result was verified.
Do not start the comparison with the biggest task. Choose a change whose requirements are known and whose result can be judged. For the trial, prepare an isolated copy of the project with no production credentials. Permissions should match the scope of the exercise.
One fix, different amounts of human work
In a fictional task, the sorting of a list has to be fixed while the existing response format stays the same. Tool A proposes a short change, and tool B rebuilds several modules. Both results may need their behaviour assessed, but the cost of review and the risk of regression will differ.
Check the required behaviour, compatibility with existing interfaces and the scope of the edit. A green test run is not enough if the test does not cover the changed contract. The developer should be able to explain the fix after reading the code, without relying only on the agent's report.
Compare the path to an accepted change
Record preparing the task, the tool's time, the review and the corrections. Rejected runs use time too. Do not compare one accepted attempt with a series of rejected ones without showing that difference.
Write down the product, the model, the settings and the starting point of the repository. The client that drives the model and the model itself are different parts. We describe a detailed protocol in the article on testing a local coding agent. Its historical results are no new test of current products.
What to settle before you buy for the team
| Area | Decision to write down |
|---|---|
| Access | Which repositories and commands the work covers |
| Acceptance | Who reviews the change and which behaviour they check |
| Cost | Usage limits and people's time after each run |
| Upkeep | Who updates the instructions and the set of trials |
If the team is only learning to judge changes, giving the agent more freedom can lengthen the review queue. Widen the scope after you have assessed real tasks. We present no ranking of tools here and no results of a new comparison.
Syntalith runs the AI-Native course for developers. We agree the programme and the price after a conversation about the repository and the way the team works. For the call, bring the type of change where AI helps you write but still leaves a hard review.
What it costs
- Time saved
- 8 people × 2 h a week × 4.33 = 69.3 h
- Value of that time
- €2,418
- Licences
- $160 (€143)
- Left after licences
- €2,275 a month
Payback of a €1,150 training day
0.5 months
The hours saved are the scenario's assumption, which we check by timing tasks before the training and after 30 days. €34.90 an hour is the EU average hourly labour cost for 2025 according to Eurostat. Claude Team seats at $20 per person a month, at the ECB rate of 1.1205 USD per euro on 22 October 2026.
- The time of the task before the training and after 30 days
- The corrections made by the person checking the result
- Tasks where a participant still needs help
- Licences and tools cleared for work
Where to start
We run the workshop as part of team training, from €1,150 net a day. For the free review, a description of one task the team does every week and the tools it uses is enough. You receive the written training programme within 2 working days.
Worked example · Scenario
How we count the payback: a development team
We run the same count for your team in the free review.
€4,188
a month for the work on 600 questions
EU cloud version payback: 0.8 mo.
Our assumptions are 12 minutes to find an answer in the documentation, and checking in 15% of the old time. €34.90 an hour is the EU average hourly labour cost for 2025 according to Eurostat, for companies with 10 or more employees. Model cost: Claude Sonnet 5.5, about 6,000 + 500 tokens per answer in the EU cloud version, at the ECB rate of 1.1205 USD per euro on 22 October 2026. We count the payback for an app with a model in the EU cloud (from €3,000 net). We price the model on your server after a call, and the cost of people's work stays the same.
See the full calculation →
