Skip to main content Scroll Top

AI in Software Development: Where It Speeds the Work and Where Review Still Decides

Updated November 2026 · Written and maintained by the Progression Agency strategy team

AI in software development now runs from a suggestion that finishes a line to an agent that reads a repository, edits files, runs the tests and opens a pull request on its own. The speed is real, and so is the new work it creates: someone still has to read every change, check every package it adds and decide what ships. This page follows one change from ticket to production and marks where AI helps, where an agent can act alone and where a person signs off. Progression’s own developers use these tools on client builds, and the team is based in New York.

On this page · 14 sections
  1. A pull request that passed every test
  2. What AI in software development covers now
  3. What is agentic software development?
  4. AI assisted software development, step by step
  5. How the work shifts in each phase of the life cycle
  6. Reviewing AI-written code
  7. Which packages should an assistant be allowed to add?
  8. Security for code written with AI
  9. Who owns code written with AI?
  10. Keeping code, keys and data where they belong
  11. AI tools for software development, sorted by job
  12. How to use AI in software development on an existing codebase
  13. How Progression’s developers use AI
  14. What AI-assisted builds cost

Want AI inside your product instead? Building AI features that your users touch is a different job, covered on our AI app development page.

Examples marked “Illustrative example.” are hypothetical, and no real codebase, company or person is shown.

A pull request that passed every test

Picture a hypothetical team that hands a coding agent a small task: add a CSV export to the reports screen. Twenty minutes later a pull request arrives. Every test is green, the export works, and the diff touches fourteen files.

Two drawn boxes. The first shows a pull request for a CSV export where every test passes. The second lists what a reviewer found: an unrequested package, an API key in a test fixture, and a database role widened to write access.Two drawn boxes. The first shows a pull request for a CSV export where every test passes. The second lists what a reviewer found: an unrequested package, an API key in a test fixture, and a database role widened to write access.
Illustrative example. A hypothetical pull request that passed its tests and still needed three fixes.

The reviewer who read all fourteen files found three things no test was written to catch: a date library added for one line of formatting, an API key pasted into a test fixture, and a database role quietly widened to write access so a query would run. None of them broke the build. All of them would have shipped.

  • Tests prove the code does what the tests ask, nothing more.
  • An agent optimizes for finishing the task it was given.
  • A person reads for what the task never mentioned.

What AI in software development covers now

Three working modes sit under the same label, and they hand over very different amounts of work. GitHub’s documentation for Copilot in the editor lists all three: suggestions as you type, chat about the codebase, and an agent mode that finds relevant code, edits files, runs commands and validates its own work while you review and redirect it.

Three bands. Suggestions finish a line or block as you type. Chat explains code and plans changes. Agents find code, edit files, run commands, check their work and open pull requests.Three bands. Suggestions finish a line or block as you type. Chat explains code and plans changes. Agents find code, edit files, run commands, check their work and open pull requests.
Suggestions, chat and agents, ordered by how much work is handed over.
Five ways AI shows up in a development team’s day
ModeWhat the AI doesWhat the developer doesTypical risk
SuggestionsCompletes the line or block being typedAccepts, edits or rejects each onePlausible code that is subtly wrong
ChatExplains code, compares approaches, drafts snippetsAsks, judges and pastes what fitsConfident answers about code it has not read
Agent in the editorFinds code, edits files, runs commands, checks its workWatches, redirects, reviews the changesWide edits across many files
Cloud agentWorks a task in a remote environment and opens a pull requestWrites the task, reviews the pull requestChanges nobody watched being made
AI code reviewComments on a pull requestWeighs the comments, still reads the diffMissed issues read as approval

The same tools now reach past the editor. On the code host, Copilot can plan and make changes, review pull requests and automate repository tasks, and you can request a review from it on a pull request like any other reviewer.

What is agentic software development?

It is development where an AI agent carries a task through several steps on its own: reading the code, deciding what to change, editing files, running commands and checking the result, then handing back a finished change for a person to review. Anthropic describes Claude Code in the same terms, as an agentic coding tool that reads your codebase, edits files, runs commands and integrates with your development tools.

  • The unit of work is a task, not a line.
  • The agent decides the order of steps.
  • It runs real commands, so its environment matters.
  • Its output is a diff or a pull request for a person to judge.

Sessions, sandboxes and permissions

GitHub separates three things worth separating in any setup: the client you talk to the agent through, the session that holds the conversation and its activity, and the environment where the agent actually executes, such as a cloud sandbox, a GitHub Actions runner or a remote custom environment. The environment decides what the agent can touch. Give it the repository and the test database, never production credentials, and make anything outside that a request a person approves.

AI assisted software development, step by step

A change still moves through the same gates it always did. What changes is who does the first draft at each gate and how closely a person checks it.

A six step flow: ticket, plan, code in a sandbox, automated checks, human review of the diff, and staged release.A six step flow: ticket, plan, code in a sandbox, automated checks, human review of the diff, and staged release.
Six steps a code change takes, with the human check at each one.
  1. Write the ticket with what done looks like, including what must not change.
  2. Let chat or an agent propose a plan, and approve it before code exists.
  3. Run the agent in a sandbox with the tests and nothing more.
  4. Let automated checks run on every push: tests, static analysis, dependency scans.
  5. Read the full diff, line by line, before anything merges.
  6. Release in stages and watch the numbers that matter.

In AI assisted software development the cheapest place to catch a mistake moves earlier: a plan that is wrong costs a minute to reject, while a fourteen file diff built on that plan costs an hour to unpick.

How the work shifts in each phase of the life cycle

The phases themselves do not change; our guide to the software development life cycle sets them out. The table shows how AI in the software development life cycle redistributes the work within each one.

What assistants, agents and people each do, phase by phase
PhaseWhat assistants help withWhat agents can take onWhat a person keeps
PlanningSummarize tickets and past decisionsDraft task breakdownsPriorities and scope
RequirementsTurn notes into user storiesSpot gaps against existing codeWhat the business actually needs
DesignCompare approaches and tradeoffsSketch interfaces and data modelsArchitecture and security design
DevelopmentComplete code, explain unfamiliar partsImplement scoped tasks end to endApproval of every merged line
TestingSuggest test cases and edge casesWrite and run tests, fix failuresWhether the tests prove the right thing
DeploymentExplain pipeline errorsDraft pipeline and config changesRelease decisions and access to production
MaintenanceRead logs and stack tracesPropose fixes and dependency updatesWhich fixes ship, and when

Which phase gains the most?

Development and testing, because the work there is concrete and checkable: code compiles or it does not, and a test passes or fails. Planning and design gain less, since their outputs are judgments that an assistant can inform but not settle.

Reviewing AI-written code

NIST’s Secure Software Development Framework asks teams to decide when to use code review, where a person looks directly at the code, and code analysis, where tools find issues either fully automated or alongside a person. Code from an assistant needs both, aimed differently.

Two columns comparing code review by a person with code analysis by tools: what each catches, speed, and where results are recorded.Two columns comparing code review by a person with code analysis by tools: what each catches, speed, and where results are recorded.
Code review and code analysis compared.
  • Peer review of every change, with earlier analysis and test results in front of the reviewer.
  • Expert reviewers looking for backdoors and other malicious content in sensitive areas.
  • Static analysis run automatically against the team’s coding standards.
  • Every issue found recorded and triaged in the team’s tracker, not left in a chat.

Why a passing test suite is not a review

Tests check the behavior someone thought to test. An agent that writes both the code and its tests can make them agree with each other and still miss what the task meant. The SSDF also suggests adding a test for each previously reported vulnerability so the same error cannot come back, a habit that matters more when code arrives faster.

Should AI review AI-written code?

As a first pass, yes. An AI reviewer comments quickly on every pull request and catches common slips. Treat its comments as one more input, and keep a person’s approval as the condition for merging.

Which packages should an assistant be allowed to add?

Assistants reach for libraries freely, and every new package is code your team now ships and maintains. The SSDF’s guidance on third-party components applies unchanged: acquire well-secured components, check them for publicly known vulnerabilities, and make sure each one is still maintained and has not reached end of life.

  • Confirm the package exists under the exact name suggested, from the expected publisher.
  • Ask whether the standard library or an existing dependency already does the job.
  • Check open vulnerabilities and the date of the last release.
  • Read the license before the package reaches the main branch.
  • Pin versions, and let automated scans watch them afterward.

Security for code written with AI

Two public references cover most of the ground. NIST published SP 800-218A in July 2024, an SSDF profile that adds practices for generative AI and dual-use foundation models across the life cycle, written for producers of AI models, producers of systems that use them, and organizations that acquire those systems. It is used together with the original SSDF.

A drawn card with four rows naming PO, PS, PW and RV with a short description of each.A drawn card with four rows naming PO, PS, PW and RV with a short description of each.
A drawn card listing the SSDF’s four sets of practices: prepare the organization, protect the software, produce well-secured software and respond to vulnerabilities.

The OWASP Top 10 for Large Language Model Applications identifies the most critical security vulnerabilities in applications built on large language models, and the project has grown into the OWASP GenAI Security Project, which also covers agentic AI systems. Generative AI for software development sits in both worlds: it is a tool that writes code, and an AI system with its own weaknesses.

  • Treat text an agent reads from issues, web pages and files as untrusted input, not instructions.
  • Keep secrets out of prompts, context files and repositories an agent can read.
  • Limit the commands and network access an agent has without asking.
  • Log what each agent session changed and which account it ran under.

Software to build this year?Describe the product and the deadline. You get a free proposal with the scope, the team and the price.

Get a free proposalTell us about a project you’d like help with.

Who owns code written with AI?

The U.S. Copyright Office answered the core question in Part 2 of its AI report, published January 29, 2025. Outputs of generative AI can be protected by copyright only where a human author has determined sufficient expressive elements.

  • A human-authored work that is perceptible in the output can be protected.
  • Creative arrangements or modifications of the output by a person can be protected.
  • The mere provision of prompts is not enough.
  • Using AI to assist, or including AI material in a larger human work, does not bar protection of the human contribution.

The Office’s copyright and AI page tracks the rest of the series, including Part 3 on training AI models, released in pre-publication form on May 9, 2025. For a software owner the practical step is a record: who designed the system, which parts people wrote or substantially reworked, and how AI output was selected and changed. Contracts with any development team should also say plainly who owns the code and how AI tools may be used on it.

A timeline of six dates from February 2022 to October 2026: the SSDF, the Copyright Office initiative, NIST's generative AI profile, the Office's Part 2 and Part 3 reports, and GitHub's cloud experience for paid plans.A timeline of six dates from February 2022 to October 2026: the SSDF, the Copyright Office initiative, NIST's generative AI profile, the Office's Part 2 and Part 3 reports, and GitHub's cloud experience for paid plans.
Six dated publications and changes that bear on code written with AI.

Keeping code, keys and data where they belong

The SSDF asks that all code be stored on the principle of least privilege, so only authorized people, tools and services can reach it, and that every change be tracked to an individual account. Agents count as tools under that rule.

  • Give each agent its own identity, so its commits are attributable.
  • Scope repository access to the projects the task needs.
  • Check where each tool sends code and whether the vendor keeps it.
  • Store keys in a secrets manager, never in fixtures or configs.
  • Review the access an agent holds as often as you review a person’s.

AI tools for software development, sorted by job

Product names change faster than the jobs they do. Sorting AI software development tools by where they run and what they touch makes the adoption questions clearer.

Six kinds of AI developer tool and what to check first
Tool typeWhere it runsGood forCheck before adopting
Editor assistantInside the code editorSuggestions, chat about open files, agent modeWhat code leaves the laptop, and who keeps it
Terminal agentThe developer’s command lineMulti-file tasks, running and fixing testsWhich commands it may run without asking
Cloud coding agentA remote sandbox or CI runnerBackground tasks that end in a pull requestWhich secrets and networks the environment can reach
Pull request reviewerThe code hostFirst pass comments on every changeThat a person still approves the merge
Chat assistantA browser or desktop appExplaining, drafting, rubber duck debuggingPasting code under the company’s data rules
Test generatorEditor or CIFilling coverage gapsTests that assert behavior, not just run lines
4: Sets of SSDF practices. Prepare, protect, produce, respond.
2024: NIST's generative AI profile. SP 800-218A, final July 26.
Jan 29, 2025: Copyright Office Part 2. Copyrightability of AI outputs.
10,000+: Comments to the Office. On its 2023 notice of inquiry.
600+: OWASP GenAI contributors. From more than 18 countries.
3: Ways to work with AI. Suggestions, chat and agents.

Tool choices for app teams, from editors to cloud backends, are covered on our app development tools page, and no-code builders on our low-code app development page.

AI in mobile app development

Mobile adds checks an assistant cannot run on a laptop alone: real devices, store review rules and platform permissions. AI in mobile app development speeds the screens and data models; device testing and the store submission still need a person who knows both platforms. See our iOS app development and Android app development pages.

How to use AI in software development on an existing codebase

Older code is where assistants help most and where agents can do the most damage, because the reasons behind odd choices are rarely written down. If you want engineers who already work this way, our page on how to hire AI developers sets out the skills to test and the ways to engage them.

  1. Start with reading: ask chat to explain modules before anyone changes them.
  2. Write tests around the behavior you must keep, with AI help, before refactoring.
  3. Hand agents small, bounded tasks with a clear finish line.
  4. Add instruction files that tell agents the conventions and the no-go areas.
  5. Widen the scope only after a run of clean reviews.
A drawn checklist of seven checks. Four ticked: a clear task, a sandbox without production keys, tests and analysis passed, a developer read the diff. Three crossed: new packages approved, no secrets, permissions explained.A drawn checklist of seven checks. Four ticked: a clear task, a sandbox without production keys, tests and analysis passed, a developer read the diff. Three crossed: new packages approved, no secrets, permissions explained.
Illustrative example. A hypothetical agent change passes four of seven checks and waits.

Teams moving an old system forward often pair this with DevOps work on the pipeline, so every agent change meets the same automated checks as a person’s.

How Progression’s developers use AI

Progression’s own developers use AI coding tools on client builds, inside the same delivery process as hand-written work. Generated code goes through the same review, security checks and testing as code a person typed, and nothing merges before that.

  • Two-week sprints with automated tests and code review.
  • Releases through CI/CD pipelines with staged rollouts.
  • AI-written changes reviewed like any other change before merge.
  • Clients own the code they pay for.

The wider build work is described on our software development company, MVP development and AI agent development pages.

What AI-assisted builds cost

AI in software development speeds parts of the work, but scope, integrations and review still set the price. The bands below run from the typical low to the published high in the US market.

Software builds: published US price bands (market ranges only)
BuildTypical low to published highNotes
US development team, hourly$100 to $200 an hourAgency or contract bill rate
App MVP$15,000 to $80,000Single or cross-platform, core flows
Integration or internal tool$30,000 to $100,000One integration or one workflow
SaaS MVP$30,000 to $150,000First paid version of a product
Mid complexity app$60,000 to $150,000Payments, integrations, admin panel
Portal or platform$75,000 to $400,000Multi-role business systems
Horizontal range bars for an app MVP, an integration or internal tool, a SaaS MVP, a mid complexity app and a portal or platform.Horizontal range bars for an app MVP, an integration or internal tool, a SaaS MVP, a mid complexity app and a portal or platform.
Published US price bands (market ranges only) for five kinds of software build, typical low to published high.

Progression’s price is set in the written proposal, after scope, and is subject to change. Our app development cost guide breaks the bands down further.

Does AI make custom software cheaper?

It can shorten the drafting, and it adds review, security and testing work that cannot be skipped. The saving shows up most on well-defined tasks in a well-tested codebase, and least on new architecture or unclear requirements.

What this page leaves out: the Progression Method, proprietary and reserved for clients. Reach out to learn more

Pricing depends on scope and is confirmed in a written proposal. Published prices are subject to change.

Frequently asked questions

What does AI in software development do day to day?
It completes code as developers type, explains unfamiliar code, drafts tests, and, as an agent, carries scoped tasks from edit to pull request. People still plan the work, review every change and decide what ships.
How would you define AI assisted software development?
It is software built by developers who use AI tools for parts of the work, from suggestions and chat to agents, with a person reviewing and approving everything that merges.
How does a coding assistant differ from a coding agent?
An assistant answers or suggests while you work; an agent takes a task and works through several steps itself, editing files and running commands, then hands back a change for review.
Can an AI agent open a pull request by itself?
Yes. Cloud coding agents can work a task in a remote environment and open a pull request; a person then reviews it like any other.
Should AI review AI-written code?
It can give a fast first pass on every pull request. A person’s approval should still be the condition for merging.
What does NIST say about code review?
Its Secure Software Development Framework separates code review, where a person reads the code, from code analysis, where tools find issues, and asks teams to record and triage everything found.
How do you check a package an assistant suggests?
Confirm the exact name and publisher, look for known vulnerabilities, check it is still maintained, read its license, and ask whether an existing dependency already does the job.
Is there a security framework for generative AI development?
Yes. NIST’s SP 800-218A, final in July 2024, adds generative AI practices to the Secure Software Development Framework and is used together with it.
What is the OWASP Top 10 for LLM applications?
It is OWASP’s list of the most critical security vulnerabilities in applications built on large language models, now part of the OWASP GenAI Security Project.
Can code written with AI be copyrighted?
Parts of it can. The U.S. Copyright Office says generative AI output is protected only where a human author determined sufficient expressive elements, for example through creative arrangement or modification.
Does writing prompts make me the author of the code?
Prompts alone are not enough, according to the Copyright Office’s January 2025 report. Human authorship comes from the expressive choices a person makes in the work.
Does using AI stop the rest of my code being protected?
It does not, in the Copyright Office’s view. Using AI to assist, or including AI material in a larger human-written work, does not bar protection of the human contribution.
Is my code sent to the AI vendor?
Usually some of it is, as context for the model. Check each tool’s data terms for what is sent, whether it is kept, and whether it is used for training, before it touches client code.
Where should a coding agent run?
In an environment that holds the repository and test resources but not production credentials, such as a sandbox or a CI runner, with anything wider granted only on request.
Can ChatGPT create apps?
It can write a lot of the code for one, and builder tools can turn a description into working screens. A real product still needs architecture, data, security, testing and store or server deployment, which takes people who know what to check.
Does AI replace developers on a project?
It changes their work rather than removing it. Developers write fewer first drafts and spend more time planning tasks, reviewing diffs and deciding what ships.
Which phase gains most from AI in software development?
Development and testing, where the work is concrete and checkable. Planning and design gain less because their outputs are judgments.
How do we start using AI on an old codebase?
Begin with explanation and tests, then give agents small, bounded tasks with written conventions, and widen the scope after clean reviews.
How does Progression’s team use AI on client work?
Progression’s developers use AI coding tools inside the normal delivery process, and generated code goes through the same review, security checks and testing as hand-written code before it merges.
Does AI make custom software cheaper?
It can shorten drafting on well-defined tasks, while review, security and testing work remain. Scope and integrations still set most of the price.
What does an AI-assisted build cost?
Published US bands run from $15,000 to $80,000 for an app MVP and $30,000 to $150,000 for a SaaS MVP, with US development teams at $100 to $200 an hour. Progression sets its price in the written proposal.

Software to build this year?Describe the product and the deadline. You get a free proposal with the scope, the team and the price.

Get a free proposalTell us about a project you’d like help with.

Get a free marketing proposal

Tell us what you are trying to grow and we will come back with a plan, not a pitch deck. Same-day reply on weekdays.

Privacy Preferences
When you visit our website, it may store information through your browser from specific services, usually in form of cookies. Here you can change your privacy preferences. Please note that blocking some types of cookies may impact your experience on our website and the services we offer.
Contact Us