The short answer: sooner and cheaper only if your vendor has moved the time it saved on writing code into reviewing and testing it. A Harvard working paper by Fiona Chen and James Stratton, updated on 4 August 2026 and reported widely from 10 October, tracked 718 companies and about 700,000 workers from January 2021 to March 2026. After a company adopted AI coding agents, its engineers produced 30% more code, but the number of tasks and features that reached production did not change measurably, because every extra line still had to pass a human reviewer. A vendor that quotes a lower price or a shorter timeline on the strength of AI tools is making a claim about its review and testing capacity, not about the tools. Ask to see that capacity in the quote.
What the Evidence Says, and When It Was Published
| Date | Study | What it found |
|---|---|---|
| 24 September 2025 | DORA, State of AI-assisted Software Development (Google Cloud), nearly 5,000 respondents | 90% use AI at work and over 80% believe it raised their productivity, yet 30% have little or no trust in the code it writes. AI adoption went with higher delivery throughput and lower delivery stability. AI amplifies how a team already works. |
| 24 February 2026 | METR, developer productivity update | Its 2025 trial had found experienced developers 19% slower with AI tools. New data pointed to speedups of about 18% for returning developers and 4% for new ones, but both ranges include zero and METR calls the results unreliable, because 30% to 50% of developers withheld tasks they did not want to do without AI. |
| 27 April 2026 | GitHub, Copilot billing | From 1 June 2026 every Copilot plan is billed on token usage through AI Credits, because agent use raised compute costs beyond what the old request-based pricing could sustain. Copilot Business at USD 19 per user a month includes USD 19 of credits; code completions stay unmetered. |
| June 2026 | Checkmarx, 2027 Outlook: The Future of Application Security in the Era of AI, 2,350 respondents in 14 countries | Respondents put 49% of their production code as AI-generated. Companies where 81% to 100% of code is AI-generated often ship known-vulnerable code at 3.4 times the rate of those at 1% to 20% (47% against 14%). 93% had been breached through their own applications. |
| 10 June 2026 | DORA, ROI of AI-assisted Software Development (Google Cloud) | Most organisations meet a J-curve: a temporary productivity dip caused by the learning curve, a verification tax on the extra code, and testing and approval pipelines that must be scaled to the new volume. |
| 4 August 2026 | Chen and Stratton (Harvard), Artificial Intelligence in the Firm: Bottlenecks in Software Production, current version | Agents raise code output 30% at the firm level; finished software and employment do not move significantly; code review absorbs the gain. |
The Harvard paper is the one that matters for a buyer, because it measures firms rather than individual developers, and output rather than activity. Its data comes from Jellyfish, an engineering analytics platform, and joins GitHub activity, Jira issues, calendars, HR records and usage of GitHub Copilot, Cursor and Claude Code. Adoption of agents in the sample rose from almost none in October 2024 to over 95% of firms by January 2026, so the sample describes the market a buyer is now purchasing from. It is a working paper, not yet peer reviewed, and its data ends in March 2026.
Why More Code Did Not Become More Software
Writing code is a third of the job. The authors estimate, and checked with a survey of 100 engineers, that engineers spend 30% to 40% of their time writing code; the rest goes to scoping, review, testing and deployment. When agents sped up the writing, the work piled up at the next step.
| Measure, per firm, after agent adoption | Before | Change |
|---|---|---|
| Lines of code added per worker-month | 5,017 | +1,494 (30%) |
| Days from a pull request being opened to being merged | 7.03 | +3.45 (49%) |
| Share of pull requests sent back with changes requested | 13% | 25%, nearly double |
| Review comments per pull request | 1.66 | +0.58 (35%) |
| Share of workers doing code review | 29% | 33% (+14%) |
| Jira issues resolved per worker-month | 3.67 | No significant change; an increase above 12% is ruled out |
Pull requests did not get bigger, so the extra comments and rejections reflect code that needed more correction, not more code per request. And AI review tools did not clear the queue: by March 2026 nearly 80% of the firms used AI in code review, yet AI wrote only 23.3% of review comments and touched 10.8% of pull requests. Human reviewers stayed central. The paper also finds no significant effect on employment: a precise zero on overall headcount and no measurable cut to engineering teams. The firms did not get smaller; they moved people to review.
What This Means for the Price and Timeline of Your Build
A custom application is quoted in engineering hours, and the hours are now distributed differently. Fewer go to writing code. More go to reviewing it, testing it and fixing what the review finds, and the Checkmarx figures show what happens when a company skips that step: the heaviest users of AI-generated code ship known vulnerabilities at more than three times the rate of the lightest. In 2026 a credible quote looks like this:
- Coding hours lower than two years ago for the same scope. The 30% figure is a firm-level average that includes teams that barely use the tools; a vendor should be able to say what it sees on its own projects.
- Review and testing shown as paid, planned work, with a named reviewer who is not the engineer or agent that wrote the code, automated tests in the pipeline, and a security scan before release. If these lines are missing, the saving was taken and the risk was left with you.
- Acceptance defined by working software, not by delivery of code. Short cycles with a demo at the end of each, a staging site you can use, and payment milestones tied to features that pass your acceptance criteria.
- The vendor’s AI tooling as a cost of doing business, not a line on your invoice. GitHub’s metered billing puts Copilot Business at USD 19 per user a month with usage beyond the included credits charged on top; the tools are a small cost next to the review hours they create.
- A timeline that allows for the J-curve. DORA’s finding is that teams adopting agents slow down before they speed up. A vendor that adopted them in 2025 has passed that dip; one adopting them on your project has not.
Seven Questions to Ask Before You Sign
- Which AI coding tools do your engineers use, and on which steps? A specific answer (agents for first drafts and tests, people for design, review and release) is a good sign. “We use AI” is not an answer.
- Who reviews the code, and how many hours of review are in this quote? Compare the review hours with the coding hours.
- What share of your pull requests is sent back for changes? A vendor that tracks this knows its own bottleneck. The study’s average after agent adoption was one in four.
- What runs automatically before a release? Unit and integration tests, dependency and vulnerability scans, and a staging deployment, with the results visible to you.
- Where do our code and data go? Agents send code to a model provider. Ask which provider, under what data terms, and whether your production data is ever used in prompts.
- If a feature fails acceptance, who pays for the rework? In a fixed-scope contract the vendor does; in a time-and-materials contract the review hours should already be in the estimate.
- Who owns the code, the repositories and the AI tool accounts at handover? You should, in your own name, with the tool subscriptions transferable or cancellable.
Where SoftwareBiz fits. Our custom software development team scopes a web application, portal or internal tool in a discovery phase that decides whether to build, buy or extend, then delivers in short development cycles with a demo at the end of each, automated testing and a staging site you can review before a release, with a monthly support plan afterwards. Where a journey matters to customers or staff, our UI/UX design team tests structure and flow in a clickable prototype before development, which removes the most expensive kind of change request: the one that arrives after the code is written. For an application another vendor built, we begin with a review of the code and infrastructure and document what we find before planning changes. Source code, hosting and platform accounts are set up in your name or transferred at handover.
What to Do This Quarter
If you are comparing proposals, put the review and testing hours of each side by side and ask the seven questions in writing. If you already have a vendor, ask for the share of its pull requests sent back for changes and the time from a pull request to a merge; a vendor who can answer within a week is managing the bottleneck the study describes. The same discipline applies to the agents your own staff are starting to run: our article on deploying AI agents you can still control covers the identity, permission and approval controls that keep an agent’s speed from becoming your risk.