Case studies
Ziroly

A smaller team, shipping more every month

Ziroly is a Polish e-invoicing and accounting app for sole traders and accounting offices, connected to the national KSeF system, with a backend, a web app and a mobile app. It is our own product, and the first team we took through AI enablement. Every number on this page comes from its git history and deploy pipelines.

7 to 4
people on the team, monthly average
64 to 93
merged pull requests a month
9 to 23
pull requests per person a month, about 2.7x
2 to 46
backend production deploys

Monthly averages. Before AI: October to December 2025. With AI: January to October 2026, every full month since the switch.

Before

July 2025 to mid-January 2026

  • Up to 10 people contributed, each owning one area: backend, web or mobile.
  • About 12 merged pull requests a week across the product repos.
  • The backend had reached production only twice. The web app had no production pipeline at all.
  • The bug backlog lived in Trello, outside the repos where the work happened.
  • No agent instructions or security docs in the repos, and 11 security pull requests in six months.

After

Mid-January to November 2026

  • 2 to 3 people run the whole product: backend, web app, mobile app and landing page.
  • About 93 merged pull requests a month on average, against 64 before, with a team about half the size.
  • The backend deploys to production about once a week: 46 times instead of 2.
  • The web app has shipped 48 production releases since its pipeline went live.
  • Smaller pull requests: the median dropped from 267 to 193 changed lines.

The enablement

What we changed

No training days and no slide decks. The changes went into the repos, the backlog and the pipelines, the same way we run the 8-week Sprint for clients.

  1. 01

    AI agents in daily work

    From mid-January 2026 AI coding agents became part of daily work, with a person reviewing and merging each pull request.

  2. 02

    Context the agents can use

    A dedicated dev-agent repo holds the instructions, conventions and workflows the agents follow. It grew through 86 pull requests.

  3. 03

    The backlog moved into the repo

    In February 2026 the Trello bug list became tasks the agents could pick up inside the repo. February was the busiest month, with 187 merged pull requests.

  4. 04

    Pipelines to production

    Production deploy pipelines for the backend and, from May 2026, the web app, so merged work actually ships.

  5. 05

    Security as a habit

    A security-docs repo with 79 pull requests, and dedicated security weeks from October 2026.

Month by month

How the numbers moved during the enablement

Merged pull requests per month in the product repos, with the number of people who contributed that month underneath. Output went up while the team got smaller.

19
37
76
40
76
105
187
57
103
101
104
124
30
29
92

Aug

2025

Sep

 

Oct

 

Nov

 

Dec

 

Jan

2026

Feb

 

Mar

 

Apr

 

May

 

Jun

 

Jul

 

Aug

 

Sep

 

Oct

 

3
5
6
6
10
9
7
3
3
5
3
3
3
2
2
Before AI With AInpeople on the team that month

February 2026 was the backlog month, when the old bug list became agent tasks. August and September 2026 were quieter, with 2 to 3 people. We show them anyway: a before-and-after you can trust has to include the slow months too.

The whole period, side by side

MetricBefore AIWith AI
Merged pull requests294921
Merged pull requests per week12.021.6
Commits per week26.732.2
People committing in an average week4.22.8
Pull requests per active person per week2.87.7
Median pull request size (lines)267193
Feature pull requests72343
Security pull requests1162

Before AI: 29 July 2025 to 15 January 2026 (24 weeks). With AI: 16 January to 9 November 2026 (43 weeks).

Production releases

  • Backend

    2 to 46

    successful production deploys

  • Web app

    0 to 48

    production pipeline added in May 2026

  • Mobile app

    45 to 349

    merges to main, each one a production release

How we counted

The data comes straight from GitHub: merged pull requests by people (bots excluded), non-merge commits on each default branch, and successful runs of the production deploy workflows. It covers the five product repos: backend, backend framework, web, mobile and landing page.

A person counts as on the team in a month if they made a commit or merged a pull request that month. People joined and left at different times, so we compare by team size each month.

Pull requests are an activity measure, not a quality measure. Next we are adding change failure rate and review wait time, the same metrics we baseline for clients in the AI Engineering Snapshot.

Want these numbers for your team?

The one-week Snapshot measures your delivery from git, CI and Jira, and shows where AI would pay off first. The fee is credited to the Sprint if you go ahead.

Book a Snapshot call