Find where your team sits, and what it takes to reach the next level.
An AI software delivery maturity model is a scale for how much of software delivery AI agents do and how much people do, from a few engineers experimenting with AI to agents that deliver software on their own. We use the five levels Cortex set out in The Five Levels of the Enterprise AI Software Factory, described here in our own words: Evangelists, Automations, human-driven and agent-assisted, human-assisted and agent-driven, and autonomous software delivery. Pento delivers client projects at level 4, on Steer by Pento.
Changes requested in review go back to Build.
A few engineers try AI on their own: a chat window, a personal copilot, a demo at the all-hands. How the team ships hasn't changed, and nobody measures whether it helps.
Pick one well-understood job and automate it end to end, with an owner.
AI runs in narrow jobs the team understands well: a CI bot that bumps dependencies, a bot that comments on pull requests, a copilot in a few editors. Each one is useful and separate, and people still write nearly all the code.
Give engineers a coding agent in their daily work, with shared conventions for how it's used and reviewed.
Engineers work with a coding agent in the IDE or terminal and stay at the keyboard: they prompt, watch, and accept or undo each step. Output per engineer goes up, but the agent works only while someone watches it, and review is still line by line.
Write goals with acceptance criteria before agents start, run each agent in its own sandbox, and gate merges by risk.
Agents take feature-sized goals with written acceptance criteria and work them in isolated sandboxes, with a budget and a time limit. People write the specs, steer running sessions, answer questions and approve what the merge policy reserves for them. Every change carries evidence. This is where Steer by Pento runs client work.
Most of the gain lives here: people spend their time on specs, judgment and approvals instead of typing.
Specs go in and production code comes out, with people designing the factory instead of approving its output. Cortex's own framing is that enterprises can't run this way today.
Evidence shows a change does what its spec says. Changes that rewrite data or reach outside systems are the ones you can't take back, so on client work Steer keeps a person in the loop for them: every merge policy asks before an external change merges, and the Gated policy asks on data changes too.
Work starts as a goal: feature-sized, with acceptance criteria and a checklist the orchestrator keeps. Agents build against the criteria, and the review checks each one.
Each session runs in an isolated sandbox with a budget and a hard time limit, and the sandbox is thrown away when the run ends. A person can steer a running session with a message.
The review runs on a different model than the coder and returns approve, approve with findings, or request changes, each finding tied to a file, a line and a criterion. A forward deployed engineer judges the result.
Each change is classed Low, Data or External risk, and the policy you choose decides which ones merge on evidence and which ask a person.
An evidence report proves a change does what its spec says: the tests ran, the screenshots match, the review found nothing. It can't tell you whether the spec was right. A forward deployed engineer judges every change for that.
Steer classes every change as Low, Data or External risk. Every merge policy asks a person before an external change merges, the Gated policy asks on data changes too, and a change the review model rates under 0.80 safe always asks.
Every orchestrator decision carries a probability, and below 0.70 Steer asks a person instead of acting. Questions come with a deadline and a default, so a person stays in the loop without the project stalling.
Pento forms a pod of forward deployed engineers for your work, for example four engineers or whatever machine learning, software, DevOps and MLOps talent you need. They work with Steer alongside, and the fee depends on the pod's size.
For integrations and smaller projects you can scope up front. After the free audit we write the scope with you as goals with acceptance criteria and fix the price. Evidence on every pull request shows when each goal is done.
Not covered here? Bring the question to a discovery call. An engineer reads every request.
Book a discovery callContact us
Start with the free audit. It ends with a written quote: a team of forward deployed engineers, or a fixed price for a first project with a clear scope. An engineer reads every request.