For IT firms that deliver software for clients
Agents propose. Code confirms. A person decides.
AI agents take your tickets from spec to pull request, inside the repository and CI you already use. Fixed checks with no AI in them re-run every test before a person sees the work.
Scan. I measure your delivery line and score your codebase. You get a written go or no-go for a Pilot.
A 30-minute call. I open the real run logs on it.
- 120tickets entered the line
- 114passed every automated gate
- 103merged to main
Measured Run log, 21 September to 22 October 2026. One operator, one product. 80 merges by a person, 23 by the factory.
Why now
Writing code got cheap. Checking it did not.
more pull requests merged in teams with high AI adoption. Review time grows 91%.
Faros AI, AI Productivity Paradox Report, 2025of technology professionals report little or no trust in the code AI generates.
DORA, State of AI-assisted Software Development, 2025upfront annual cost reductions that IT service providers now commonly commit to in many service areas.
Gartner analyst DD Mishra, in Outlook Business, September 2025The tools made typing faster. The slow part is now proving the code is right, and clients have started to ask for the saving. The line is built the other way round: code checks what code can check, and a person sees a short card with the evidence and decides.
- Agent makes or checks
- Code confirms, with no AI in it
- Person decides
The line
Nine stations from idea to production. Each one shows who owns it.
No station trusts the one before it. Eight run today on the reference line. Release is designed; its first real run is the launch planned for 1 November 2026.
Station 5 of 9
Gate
Code confirms
The agent's claim that tests are green does not count. Only the orchestrator's run counts.
- Output
- A report with one row per gate.
- From the logs
- 13 checks per run (median). 162 mutations proven across 55 runs. 118 of 124 passing runs needed one attempt.
More on the line: the waves, the five roles and the six guardrails →
Illustrated replay of the nine runs of 5 October 2026. Ticket, tries, machine time and gate result come from the run log; the other stations show their job. Release is designed, not yet run.
Evidence
What the run logs say
21 September to 22 October 2026. One operator, one product build. Logs read again on 22 October 2026.
Green is not the finish line
112 tickets went to a separate AI review session after passing every automated gate. In 73 of them the review took the merge decision away from the machine and gave it to a person.
39 clean73 flagged for a person
UnconfirmedWhat the AI reviewer reported. No person has confirmed these findings yet, so they are not counted as defects.
- 55%
- of merged lines are tests and proofs
- 200
- lines added per merged ticket (median), in 4 files
- 3 min 50 s
- machine time per passing run (median)
- 162
- mutations proven across 55 runs
MeasuredFrom the line's run log, ticket index and run reports.
The story behind these numbers: “The model is the cheap part” →
| time (UTC) | ticket | result | tries | time |
|---|
“I could not get the flow tests to pass: the test server runs without the feature flag, and I cannot change that within the ticket's scope. [...] Without one of the two options, the Definition of Done is green only with skipped tests, and the flow is not proven.”
Your process
The line goes into the way you already deliver.
The first tickets on the line are the ones a check can confirm. Decisions about scope, production and sign-off stay with your people.
First on the line
- Migration and upgrade work with a clear acceptance test
- Component and template work that a visual check can confirm
- Regression tests for behaviour that must not change
Stays with people
- Architecture and scope decisions with the client
- Anything that touches production data
- Acceptance and sign-off
What the Pilot measures
- Lead time from ticket to merge
- Rework after review
- Defects found in client acceptance
Capacity you gain stays with you as margin.
A measured gain can fund the reduction clients have started to ask for, so the discount does not come out of your margin.
Install and run the line for your clients' teams under your own brand. The audit trail per ticket is what their CIO and their auditor ask for.
I do not promise a gain. The Scan sets your baseline and the Pilot measures the result on your backlog.
Offer
Three steps. You decide where to stop.
A fixed price is agreed in writing before each step, and each step ends with a written exit. The Scan price below is a starting price. No VAT is charged (not a VAT payer). Travel is not included.
-
Scan
from €4,900
- Interviews with the delivery lead, two engineers and the owner of client risk
- Read-only review of one repository, its CI and its ticket flow, on GitHub, GitLab or Azure DevOps
- A readiness score for the codebase: six checks a script runs, from a one-command start to lint and type checks in CI
- One of your real tickets rewritten to the line's standard and walked through the gates on paper
- Policy check: which code may go to which model, under which contract
- A baseline, including the tech lead's minutes per merge today
ExitA written go or no-go for the Pilot. Below the readiness threshold, the go comes with the readiness work first. The report is yours either way.
-
Pilot
- One team of up to six engineers, one repository, your CI, your model accounts
- Five agents of the base line on your stack: ticket drafting, dev agent, review agent, status agent and morning brief
- 15 to 20 real tickets taken through the line
- A human-gate matrix agreed with you: what may merge alone, what never does
- Three targets written at the end of the Scan and measured on your backlog
ExitA measured result on your backlog. The line stays with your team for internal use.
-
Implementation
- Two to four more teams, with a line lead trained in each
- Governance pack: AI use policy, human-gate matrix, audit trail per ticket
- Input for your EU AI Act and GDPR reviews: what the system logs, who decides, what is never automated
- Decision log and memory structure for the delivery organisation, and a management view of line health
- Wave 1 on your line once its checkpoint is met: designed today, priced in writing before it starts
ExitYour line leads run it without me.
What you own
Your tickets, gates, runbooks and the configured line. The kit stays mine, under licence. Lessons, not code: I improve the kit with what I learn, never with your code, prompts or data.
Limits
What you should know before the call
- The proof so far is one product and one operator. The line is installed on a new project built from scratch (0 to 1). Testing on larger codebases is under way.
- The review counts on this page are what the AI reviewer reported. A person has not confirmed them.
- The release station has a written checklist and a rollback procedure. Its first real run is the launch planned for 1 November 2026.
- The reference line runs on a Next.js and PostgreSQL codebase on Windows, with CI on GitHub Actions. Your stack needs its own gates; the Pilot builds them. The GitLab CI and Azure Pipelines templates have not run yet.
- Wave 1 and the later waves are designs. Each is priced in writing before it starts.
- Model and tool accounts stay yours. Your code does not pass through mine.
- What I need from you: read access to one repository for the Scan. For the Pilot, a team lead with merge rights and a test environment.
- You can stop after any step. The Scan report is yours even if the answer is no-go.
About
Andrej Ileršič
IT and business consultant · Rakek, Slovenia
Also written Andrej Ilersic.
More than ten years in system engineering, DevOps and observability. At Comtrade Gaming I owned the observability strategy and took a live KPI monitoring product from concept to production, with monitoring for more than 2,000 virtual servers. Independent since 2023.
I built the line to ship e-šolstvo, an education platform planned to launch on 1 November 2026.
I build the control system. I am a systems and DevOps engineer, not your developer.
info@e-solstvo.si
Questions
Questions you will ask
Who reviews what the agent writes?
Your people keep the merge. On the reference line the review sent 73 of 112 green tickets to a person. I build the control system; I am a systems and DevOps engineer, not your developer.
Our code cannot go to a cloud model.
The Scan starts with that policy check: which code may go to which model, under which contract. Some repositories stay off the line.
What does it cost to run?
In my logs, about $2.70 of model usage per merged ticket at API list prices, reviews included: $276 for 163 runs and 121 reviews. That is the coding tool's estimate, not a bill. Your stack will differ.
What if the tool vendor changes its price or limits?
The ticket standard and the gates do not depend on the vendor. The agent sits behind one settings file, and a smoke test runs after every tool upgrade.
Why not build it ourselves?
You can. Each of the line's six rules cost an incident first. You buy the shortcut.
Book a demo
Pick one team and one backlog.
- I reply within one working day with times for a call.
- A 30-minute call on the real run logs.
- You decide on a Scan: you name a team, a repository and the person who owns delivery risk.
You talk to Andrej Ileršič, who built the line.