What Birlasoft put on the record
Birlasoft is a mid-sized Indian services firm, part of the CK Birla group, selling software delivery and consulting into banking, manufacturing and life sciences. The report first surfaced via Multibagg AI, which quotes the company saying about a third of its code is written with agentic AI.
That is the whole of the public claim. The reporting available does not break it down by lines of code, by commits, by story points or by tasks where an agent did the bulk of the work. It does not say how much of the generated output survived review, and it does not describe who signs off on a merge.
None of that is unusual for an earnings-season statement. It is also why the figure is hard to use. "A third of our code" is a sentence a delivery head can say in a call and a procurement team cannot audit.
The number needs a denominator
Agentic coding means something specific: a model that plans a task, calls tools, edits files, runs tests and iterates across several steps, rather than completing a single line as you type. The output of one agent session can be four lines or four hundred, and the human share of the work varies just as widely.
A firm that counts "tasks where an agent was involved" will land on a much higher percentage than a firm counting "commits where no human edited the diff". Both can honestly report a third. Only one of them tells you anything about throughput or risk.
Token costs for a day of agentic coding
We do not have Birlasoft's token consumption. What we do have is the published API price list for the models a services firm would plausibly route agent traffic through in late 2026. To make the numbers comparable, take one stated assumption: 3 million input tokens and 150,000 output tokens per developer per working day. That ratio is typical of an agent loop that re-reads repository context, runs a test suite and retries after failures.
| Model | Input $/1M | Output $/1M | Per day (USD) | Per month, 20 days (USD) | Per month (EUR, approx.) |
|---|---|---|---|---|---|
| Claude Haiku 5.5 | 0.10 | 0.50 | 0.38 | 7.50 | 6.50 |
| DeepSeek-V4.1-Flash (1017) | 0.15 | 0.60 | 0.54 | 10.80 | 9.40 |
| GLM-5.3-Flash (high end) | 0.15 | 0.50 | 0.53 | 10.50 | 9.10 |
| Mistral Large 4 | 0.70 | 2.10 | 2.42 | 48.30 | 42.00 |
| Grok 4.7 | 2.00 | 6.00 | 6.90 | 138.00 | 120.00 |
| Gemini 4 Carbon | 2.00 | 10.00 | 7.50 | 150.00 | 130.00 |
| GPT-6 Rosalind | 15.00 | 75.00 | 56.25 | 1125.00 | 978.00 |
EUR figures use an assumed rate of 1.15 USD per euro, so they move with the market. The spread between the cheapest and the most expensive option is a factor of 150. That is the real decision a delivery team faces, not whether to use agents.
Caching and peak pricing move the ranking
Claude Haiku 5.5 sells cache reads at $0.05 per million tokens. If 80 percent of that daily 3 million input is cache, the day drops from $0.38 to roughly $0.26. Agent loops lean on caching harder than chat does, because the same file tree gets re-read on every step.
DeepSeek-V4.1-Flash doubles its price during peak hours, which turns $0.54 into $1.08 for any workload that cannot be scheduled overnight. Mistral Large 4 publishes free weights for eligible self-hosting, so an EU team can trade the invoice for GPU hours and keep the code inside its own perimeter.
The review hours cost more than the tokens
Even the most expensive row in that table comes to about €978 per developer per month. One hour of a senior engineer's time in Western Europe costs more than a full month of the cheapest row. If a third of the code arrives from an agent, the constraint is not the API bill. It is how many diffs a reviewer can read properly in a day.
That is where a services firm's margin actually lives under agentic delivery. Birlasoft's peers, from TCS and Infosys to France's Capgemini, bill largely by person-day. Generating code faster does not automatically generate more revenue when the invoice is tied to headcount. It generates more review work per billed hour unless the engagement model changes.
What a European client can ask for
Since 2 August 2026, Article 50 of the AI Act requires disclosure when users interact with an AI system and when they see synthetic content. Whether source code counts as "text" under that article is a question legal teams are still working through. The more practical pressure comes from the other direction: an EU client in a regulated sector carries its own AI Act obligations and pushes them down into the supplier contract.
In practice that means an audit trail. Which agent produced which change, which human reviewed it, and what test evidence exists. The Annex III high-risk deadline now sits at 2 December 2027 under the Digital Omnibus, so there is room to build that trail properly rather than retrofitting it.
Data residency is the second question. Sending an entire client repository to a US-hosted API is a GDPR conversation the moment the code touches personal data. Running open weights such as Mistral Large 4 on EU infrastructure removes that exposure. On our own AI Arena rig, the 16 GB of VRAM is fine for small local models and short agent tasks, but full-precision frontier weights are a server-GPU job.
What to ask before quoting a percentage
- What counts as agent-written: lines, commits, or tasks with any agent involvement?
- What share of generated diffs is edited by a human before merge?
- Which models run where, and does any repository context leave the EU?
- Who is liable when an agent's patch passes tests and still breaks production?
Birlasoft's one-third is a useful signal that agentic coding has moved past pilot stage inside a large services organisation. It is not yet a benchmark. A number with a review policy attached would be.
What makes an AI coding tool "agentic" rather than a code assistant?
An assistant suggests text inside your editor and stops. An agent takes a task, plans it, calls tools such as a shell or a test runner, edits several files and repeats until a condition is met. That loop is what drives the input-token volume in the table above, because the model re-reads context on every step.
Can a single workstation GPU handle agentic coding locally?
For small quantised models and short sessions, yes. Long agent loops with large repositories need far more memory and bandwidth than a 16 GB consumer card provides. Most teams run the orchestration locally and send the heavy inference to a server or an API.
Does the AI Act require me to label AI-generated source code?
Article 50 covers AI systems that generate text and synthetic media, and it has applied since August 2026. Whether source code falls inside that wording is not settled. For enterprise delivery the safe approach is to document which agent produced which change, so the question can be answered from records rather than memory.
