AI skeptics: AI isn't good enough right now

There's a lot of hype about what AI can do and there's a lot of incentives from frontier companies to sell you AI as the next best thing. Many people have tried AI for work and felt that it wasn't ready and have disregarded it. This would've been fine for any other technology but the state of AI changes significantly every three months and so its important to update your priors. If the last time you tried AI was in mid-2025, the landscape has significantly changed.

Software Projects by AI

AI will only accelerate existing behaviors.

One of the key benefits of AI is accelerating what you are doing. For a great engineering organization, AI will accelerate its output. If the organization is shitty, AI is just going to accelerate the diarrhea.

Before we talk about building an AI organization, let's briefly talk about what a good engineering organization is about. For this, we are going to use the years of research across thousands of organizations collected by Google under its DORA (DevOps Research and Assessment) project. In summary, DORA identifies five key metrics that show what great companies perform on:

The ROI of AI-assisted Software Development - DORA Google Cloud

Impact of Generative AI in Software Development - DORA Google Cloud

State of AI-assisted Software Development 2025

Throughput

Throughput measures how quickly changes move through your delivery system into production. DORA breaks throughput into three metrics:

1. Lead time for changes

The amount of time it takes for a change to go from committed to version control to deployed in production.

2. Deployment frequency

The number of deployments over a given period or the time between deployments.

3. Failed deployment recovery time

How quickly can you recover from a failed deployment or incident.

Stability

Stability measures how reliable your delivery process is and how quickly you can recover from failures. DORA breaks stability into two metrics:

1. Change failure rate

The percentage of deployments causing a failure in production that requires remediation (e.g., a hotfix, rollback, or patch).

2. Deployment rework rate

The ratio of deployments that are unplanned but happen as a result of an incident in production.

Here is how different performing organizations rank against these metrics.

Performance level	Deployment frequency	Change lead time	Change failure rate	Failed deployment recovery time	% of respondents
Elite	On demand	Less than one day	5%	Less than one hour	18%
High	Between once per day and once per week	Between one day and one week	10%	Less than one day	31%
Medium	Between once per week and once per month	Between one week and one month	15%	Between one day and one week	33%
Low	Between once per week and once per month	Between one week and one month	64%	Between one month and six months	17%

Rework rate distribution

Survey question: What percentage of deployments were unplanned fixes for user-facing bugs?

Rework rate	% at level	Top %
0%-2%	6.9%	6.9%
2%-4%	5.8%	12.8%
4%-8%	13.7%	26.5%
8%-16%	26.1%	52.6%
16%-32%	24.7%	77.3%
32%-64%	15.4%	92.7%
>64%	7.3%	100%

Where do you stand on the DORA metrics? Take a 1-min self-assessment here

Implications from DORA

If your organization is performing poorly on the DORA metrics, it is essential to explore diagnostics, solutions, and general practices that can enhance these metrics. While AI can assist significantly, it is crucial to identify fundamental issues, such as excessive technical debt, overly complex architecture, or insufficient testing prior to release, before determining how AI might address these challenges.

The AI Maturity Model for Development Teams

The problem is that everyone is using AI, but there is a spectrum. At the worst end of the spectrum: a developer can run code, see an error, paste it into ChatGPT, see what to fix, and then go back to fix it manually. At the highest end: the developer uses an agent to describe what has to be done and sees the end result. The impact can be 0.5x to 100x depending on where you land on the spectrum and how you design your organizations.

The rise of agentic coding. LLMs run in a loop and perform multiple tasks. They also get access to the CLI and file system giving them unprecedented power. Now, instead of developers asking and manually running commands, the AI can directly read outputs and react.

Practical recommendations:

1
Start small. One stage at a time.
Don't try to implement all stages at once, attempt one-at-a-time.
2
Add tests. Build coverage for confidence.
You want to build confidence for changes before you let an agent make them. Coverage is the prerequisite.
3
Reduce technical debt.
A lot of debt doesn't get prioritized and slows down engineering and AI. You have to work around debt with lots of exception clauses and .md files explaining why something isn't quite the way it should be.
4
Consolidate your repos.
Your code should ideally be in a single repo, or the agent needs access to all repos if a feature touches multiple.
5
Remove dead code.
Exclude previous versions from AI context. You want to limit the amount of searching and context building an AI has to do.
6
Upgrade frameworks and languages to latest versions.
e.g. from Python 2 to Python 3.14, from Django 2.0 to Django 6, etc.
7
Keep your README.md or AGENTS.md lean.
Document non-obvious things not discoverable in your repo. Reference "golden" items: a great function, module, API, etc.
8
Capture gotchas and tech debt in your docs.
Add them to your README or AGENTS.md, but work to eliminate them over time.
9
Use a secrets manager.
Protect against exfiltration attacks. Ideally your AI should not be able to access secrets directly.
10
Set up browser access for AI.
AI can now connect to Chrome's dev tools through CDP and read browser state. With computer use and browser plugins, the AI can test features on the browser itself. Make sure you set these up.
11
Moving up the ladder requires cross-functional buy-in.
Development alone can't improve the process. You'll need to work with product management, QA, product marketing, and DevOps. This requires buy-in from leadership that can affect change across the chain.
12
As coding time collapses, bottlenecks shift.
Bottlenecks switch to reviews, spec, and architecture. Eventually, with enough codifying of your criteria, these can be automated with infrequent human review.
13
Minimize cross-team communication overhead.
Teams move fast when they can take independent decisions. A full-stack developer works faster with AI than a two-person front-end/back-end split.
14
Be biased towards planning and spec'ing.
Reviewing outputs after the fact ('I'll see what I get') is slower and more expensive.
15
Tokens add up — know your pricing.
The latest frontier models are expensive. As a general rule, models used via Anthropic or OpenAI's own subscription give you 3-5x more tokens than via their APIs. If you use an IDE with an API key, it will not be as cost-effective.
16
Benchmark open models.
Models such as Kimi 2.6 are benchmarking incredibly well against Opus 4.6 and are great for many coding tasks. You should benchmark various models and compare the results.
17
Follow recommendations from model providers.
See Claude recommendations and Codex best practices for guidance on working effectively with AI coding assistants.

Measuring AI Impact

1If your company is using AI, leadership should have a sense that things are moving fast. There should be a material impact on the business's top and bottom lines. Without that in place, just using AI is an intermediary metric without focusing on the actual end result.

2Within engineering itself, you measure productivity in the traditional ways. Most engineering teams work in sprints and have a sprint velocity, with everything being equal (duration of the sprint and the team members in the sprint). There should be an increase in the story points being delivered. 2-3x is typical but fully AI-pilled companies can run 10-100x speed if you collapse product, design, testing into small teams or even a single person.

3If you're not using story points, you can use other proxy metrics. You should see an increase in the number of commits, PRs, LOCs, or tickets closed. Almost all ticket management software has an AI chat that you can ask to give you a sprint-by-sprint view of tickets/points/PRs closed.

Both of them are using AI, but the agent user is several times more productive. This is not easy to setup and requires infra investment and leadership commitment.

Organizational AI Maturity Ladder

1User driven chat: research jon@example.com, look for x, now look for y, now look for z
2User uses projects and re-usable skills: research jon@example.com
3Reactive: new lead arrives in CRM, runs through enrichement
4Proactive: agents plan, search and find leads and run them through qualification process

AI Maturity Guide for Development Teams