Hi, I’ve started using Claude for some isolated bits of dev over the last few weeks to check it out. it’s been generally useful so I intend to keep on doing this moving forward.
I assume other people are now doing this and wondered how this has affected billing to clients if you’re doing bespoke work.
For example if I get Claude to do something that would probably take me 4-5 hours (some of that would be physically typing which the AI didn’t have to do) and it did it in an hour or so. On the surface that’s a great win for the client as he’s just saved themselves a load of money but this doesn’t take into account all my experience guiding Claude in the right direction, stopping it exploring dead ends, or correcting it when it did something that may cause problems in the future. So on top of the physical “this task took X amount of time”, there’s an experience component in there also.
How are people factoring that into their billing? Any changes or just proceeding as before?
Interested to head people’s thoughts and experience.
If I hired a contractor and I found out that they not only used AI, but billed me for 4-5 hours when they only spent 1 hour prompting Claude, I would not be very happy.
My last month of contracting before taking a full time job was Feb 2026. I was using Claude Code a bit by then, but not as much as I am now. No changes, I was just billing hourly, I didn’t pass the Claude costs onto any specific client.
So let me get this straight: you now bill your clients for fewer hours, and you now have to eat the cost of a Claude subscription out of your profits, and you think that’s a win?
If I can deliver more for less money, then that’s a win. That means I’m more likely to get more work and I can charge a higher rate for that work.
But I’m not contracting anymore, now I’m on the other side of the desk. If I hire a contractor, I don’t care how they get it done, I care about quality and cost. Use AI, get it done 2x as fast and charge 1.5x more per hour. Both sides win.
I would care about potential liabilities, so if a contractor uses AI, I’d want to know about that. (And I’d probably want the contractor to indemnify me against issues arising out of their use of AI.)
As for the billing, a more experienced developer doesn’t charge less because it takes them less time to complete a task than it takes a less experienced developer. They charge more because they do a better job and because they spent X years honing their craft.
It’s not smart to let the existence of AI devalue those skills. Even though anyone with AI can vibecode something that might be convincing enough to the client, your added value as a professional is that you know what is good for the client and what isn’t.
Even if you bill hourly, you shouldn’t really charge for the time something takes. Instead you charge for access to your experience. The hourly rate (or fixed project rate) is an approximation of the value you offer a client for a given project.
I think the IP stealing machine doesn’t need the source code for this anymore; the binary is enough for them.
But who’s going to need software in the future? Soon, AI will be hallucinating the functionality and experience directly, bypassing the need for software programs written for that purpose.
I’ve been struggling with this dilemma too. I have decided not to pass the costs directly because it is really my job to tell if AI is allowing me to do a task in less time and the same quality. Passing AI costs directly implies I don’t have to compromise in how much I use it, which is not only unprofessional, but predates my laziness and curiosity to experiment with it (the latter is not a bad thing but not something that my employer necessarily wants to fund). At the end of the day AI costs are similar to IDE costs, which I am not expecting my employers to cover. It is my choice how much I spend and I have to factor it indirectly in my rate. Adding a tool only makes sense when my employer perceives the same or better quality in my work, for the same fee we agreed on.
Just bear in mind heavy tasks are not free with AI models. You arn’t billed hourly by anthropic. For simple projects you may get away with using Claude but you should still be checking anything it is doing and bug testing. Large projects will require an actual account. LLMs are getting increasingly good at coding tasks. If you are a heavy developer heavily using AI it would make sense to not only bundle in the usage cost which accounts are $100 or $200 per month for heavy usage. Using agents to help with coding tasks you can rack up millions of tokens very quickly. I think looking at it as an hourly consideration wouldn’t be the correct fair billing model. There are your time inputs but also equipment/service costs related to the service you are providing. I think it would make sense to not only bill hourly but also based upon tools and services which would heavily track to your account cost or token costs. Of course you also need to realize that code made by Claude is not copyrighted so unless you actually create the code you may have to do substantial activities to honestly have owenership of the code or transfer ownership of the code via a work for hire context. Just my 2 cents. At the end of the day it amounts to your competitiveness, however I think AI use disclosures will be standardized and reverse analysis of source will be increasingly possible to fingerprint your code and know what AI programmed such as to remove any ability to control copyright to it in the future. So it may make sense to be constructive with your usages. Just offering an alternative position.
I agree with your position. Fully. AI by LLM is non-deterministic, you dont know where you end up. It is a tool - not magic.
My experience is you can use AI efficiently for something like 65-75% of the code, i.e. the simple (boring) parts. You then need to learn/check the generated code to know how to find you way round while testing/verifying/debugging. The last 25% are usually difficult to “AI” since as the AI-coder doesnt like details and likes to not only change what is needed but also make a few other changes along the way and if you are unlucky you may be in for a lot of hours chasing everything down. In my experience it is always faster to do those last changes by hand.
The world doesnt need more crappy code, use those hours to do better research and deliver more solid solutions no matter how good you are. Share a few saved hours with the customer as promotion if you need it. And always check if the model will keep a copy of the sources at their server to train future LLM-models. As an example, Google lets you use their models for free if they are allowed to keep your code for training. Customers will not appreciate that.
And if you are doing something like audio/music plugs I really think you should be using something more code-friendly than Claude, it is intended for office use. Try AI programming-editors like Zed or Cursor (both have free versions if you hire your own LLM-models) that can be setup to run the commend line tools of VS and XCode (use CMake!) so you can freely switch tools. Personally I use a mic and speech for the AI-prompts and my hands and keyboard for normal code .
.