AI-assisted work is faster. Here is how to bill clients for it honestly
If you are working out how to bill clients for AI-assisted work, the honest rule for hourly billing is short: bill the time you actually spent, including prompting, reviewing, verifying and fixing what the tool produced, and never the time the AI saved you. If that rule starts costing you money, change the price, not the count.
The usual advice is “switch to value pricing”, and it stops there. That is half an answer. Plenty of freelancers bill hourly because the client insists, and they need to know what goes on Monday’s timesheet.
What counts as billable time when you use AI?
Everything you do to get a usable result is work. The model generating text is not.
That distinction settles most arguments about AI billable hours. The hours on the timesheet are yours:
- Framing the task. Reading the brief, gathering context, writing the prompt.
- Reading the output. All of it, the way you would read a junior’s work before it went out under your name.
- Verifying. Running the tests, checking the claims, confirming the library function actually exists.
- Fixing and integrating. Rewriting what is wrong, fitting the result into the codebase, and the second attempt when the first one missed.
- The dead ends. A prompt that went nowhere is a search that went nowhere. Both were spent on the client’s problem.
Verification is the part people undercount, and it is where the skill lives. A model that writes a plausible migration in forty seconds has not saved you the forty minutes it takes to be certain it will not drop a column in production. Those minutes are what the client pays for.
What stays off the timesheet:
- Time the tool saved. If a job used to take six hours and now takes two, the invoice says two.
- Learning the tool. Getting fluent in an assistant you use for every client is professional upkeep, unless a client asks for a specific tool you have never touched. Then agree in advance whether the ramp-up is billable.
- Waiting you spent elsewhere. Start an agent on client A’s repository and spend the wait on client B, and that stretch belongs to B. Watching the run, ready to intervene, is A’s time. One minute never goes on two invoices.
Fragments and rounding
AI-assisted work arrives in small pieces: three minutes writing a prompt, six reading a diff, two rerunning a test. Round each fragment up to a fifteen-minute minimum, and ten three-minute interactions (thirty minutes of real work) become two and a half billed hours. Nobody meant to pad the invoice. The rounding did it. Rounding is a pricing decision, and fragmented work makes it a bigger one: round the day’s total per client, not each entry.
The other half is capture. A per-task timer will not survive a workflow that jumps between editor, terminal, browser and chat window every few minutes. You need a daily review while the day is fresh, or something that records it for you. More on recording that kind of day in our notes on time tracking for developers.
Can you bill for the time AI saved you?
No. The clearest reasoning comes from a profession with written fee rules and a long history of billing disputes.
On 29 July 2024 the American Bar Association’s ethics committee issued Formal Opinion 512, on lawyers’ use of generative AI. ABA opinions are guidance, not law, and this one was written for lawyers, so it binds no freelancer. Its section on fees still transfers almost unchanged:
- Hourly billers bill actual time. Lawyers who charge by the hour “must bill for their actual time”, even when a tool made them faster.
- The work around the tool is billable. The opinion’s example: fifteen minutes feeding information into an AI tool to draft a pleading can be billed, plus the time reviewing the draft for accuracy and completeness.
- Learning a tool is not billable when the lawyer will use it regularly for clients. If a client specifically asks for an unfamiliar tool, the ramp-up can be billed, provided it is agreed first.
- Fixed fees are not a loophole. If the tool makes the work much quicker, charging the same flat fee as before may be unreasonable.
- General tools are overhead. An AI grammar checker built into a word processor is overhead, like rent. A per-use service run for one client’s matter can be billed as an expense at actual cost, with no mark-up unless disclosed in advance.
Nobody will discipline a freelancer under this. Take it seriously anyway, because it is simply what an hourly rate means. An hourly invoice is a statement about what happened. Billing the hours you would have spent turns it into a statement about what did not.
Should freelancers charge less if they use AI?
Per deliverable, often yes. Per hour, no.
If you bill hourly, the invoice already shrinks when the job takes less time. That is not a discount. It is the rate doing what it says.
The sharper question is whether your rate should rise. It probably should: directing a model well, and catching what it gets wrong, takes experience, and an hour of your time now produces more than it did.
So raise it, openly, at renewal or for new clients. What you cannot do is keep the old rate and quietly log the old hours. The first is a pricing decision the client can accept or decline. The second misdescribes the work.
The opposite trap: a client who suspects AI made the job trivial pushes for a lower rate. Do not negotiate against a suspicion. Show them the record: if review, testing and correction took half the time, that is visible, and it is your argument.
When should you stop billing AI-assisted work by the hour?
When AI compresses the time on a well-defined deliverable, hourly billing turns against you. Your invoice falls every time your tools improve. You are penalised for getting better.
The decision rule: bill hourly where the scope is unknown, price the deliverable where it is known.
| Work | What AI tends to do to the time | How to bill it |
|---|---|---|
| Debugging, investigation, incidents | Helps at the edges; the unknowns dominate | Hourly |
| Advisory, architecture, code review | Little change; the judgement is the product | Hourly or day rate |
| Well-specified features, integrations, migrations | Compresses a lot | Fixed price per deliverable |
| Documentation, test suites, content to a clear spec | Compresses a lot | Fixed price per unit or batch |
A worked example
The assumptions, so you can swap in your own: you bill $90 an hour. A client wants CSV import with validation and error reporting added to their app. Without AI, this took you about 12 hours. With an assistant it takes about 5: one to scope and prompt, two reviewing and correcting the generated code, two testing edge cases and integrating.
- Hourly, before AI: 12 hours, $1,080.
- Hourly, with AI: 5 hours, $450. Same feature, less than half the fee, for being faster.
- Fixed price, with AI: you quote $800 for the feature as specified. The client pays $280 less than the old hourly job would have cost. You earn the equivalent of $160 an hour.
Both sides beat the old arrangement, and nobody billed an hour that did not happen. The quote sits below the old hourly total on purpose: that is the ABA’s flat-fee point applied. Share the gain rather than pocketing all of it.
Two conditions keep this honest. First, never justify a fixed price with hours you will not spend. Telling a client “this is about twelve hours of work” when it takes five is misleading at any price. Second, carry the risk you are charging for. If the generated code turns out subtly wrong and the job runs to ten hours, the price holds. That risk, plus a written scope and a change process, earns the higher effective rate.
How do you charge clients for AI tools and subscriptions?
Charging clients for AI use is really two questions: the time, covered above, and the tool. For the tool, there are three honest options.
| Option | Use it for | How it appears |
|---|---|---|
| Built into your rate | Subscriptions used across every client: chat assistants, code completion | Nowhere. It is overhead, like your laptop |
| Expense line at cost | Usage-based costs for one client: API calls on their project | An itemised line at actual cost, receipt available |
| Client pays directly | Heavy usage, or a client who wants control of the data | The client’s own account or API key |
Four rules keep it clean:
- Do not carve up a flat subscription as expenses. If you would pay for it with no clients at all, it is overhead. Recover it through your rate.
- Pass costs through at cost unless agreed otherwise. A mark-up the client never agreed to is a dispute waiting to happen. Want a margin? Write it into the contract.
- Get approval before heavy usage. Agree a monthly threshold above which you ask first.
- Keep usage attributable. Use a separate API key or project per client where your provider allows it, so the expense line is a copy of a bill, not an estimate.
If you are VAT registered in the UK, HMRC’s guidance says costs your business incurs itself in supplying your service are not disbursements, so recharging them means charging VAT on them. An AI tool you used to do the work was supplied to you, not the client, even when it gets its own line. If a recharge is borderline, check it with your accountant.
Do you have to disclose AI use to clients?
Sometimes the contract has already decided. Where it has not, disclose anyway, in writing, at the start.
Three situations take the decision out of your hands:
- The contract or the client’s policy covers it. Some client agreements ban AI tools, restrict them or require notice. Read the one you signed.
- Client data would go into the tool. That needs permission, not just notice. More below.
- The client asks. Answer plainly. Being caught downplaying it costs far more than any honest answer.
Everywhere else, disclosure is a choice, and a cheap one. A few lines in the statement of work (you use AI tools, you are responsible for everything you deliver, and this is how it affects billing) prevent the worst version of this conversation: the client who finds out later and wonders what else they were not told. Opinion 512 agrees: even where disclosure is not required, the engagement agreement is the natural place for it.
Client data is the hard line
Disclosure is about trust. Confidentiality is about obligations you have already signed. Before a client’s code, documents or customer records go into an AI tool, check three things:
- Your NDA or contract. A typical confidentiality clause restricts sharing client information with third parties. An AI provider is a third party.
- Personal data. If you handle personal data on the client’s behalf, you are probably their processor under the GDPR, and Article 28(2) bars a processor from engaging another processor without the controller’s prior written authorisation.
- The tool’s data terms. Whether inputs are used for training, how long they are kept, and who can see them. Terms can differ between tiers of one product, so check the tier you pay for.
The ABA is stricter than most freelancers expect: before putting client information into a tool that learns from its inputs, a lawyer needs informed consent, and engagement-letter boilerplate does not count. The practical translation: name the tools, say what goes into them, and get a yes in writing. If the client says no, work without the tool or with anonymised material.
What should an AI clause in a freelance contract say?
Four things: that you use AI tools, who answers for the output, how billing works, and what happens to client information. A starting point:
Use of AI tools. The Contractor may use AI-assisted tools, including code, writing and research assistants, in performing the Services. The Contractor reviews all output and is fully responsible for every deliverable, exactly as if it had been produced without such tools.
Billing. Where the Services are billed hourly, the Contractor bills only time actually spent, including time spent instructing, reviewing, testing and correcting AI-assisted work. The Contractor does not bill for time saved by such tools, or for time spent learning tools used in the ordinary course of business.
Tool costs. General-purpose software subscriptions are included in the Contractor’s rates. Usage-based costs incurred solely for the Client’s project are billed at actual cost, itemised on the invoice, and require the Client’s prior written approval above [amount] per month.
Confidential information. The Contractor will not enter the Client’s Confidential Information or any personal data into a third-party AI tool other than those listed in [Schedule X], used on terms under which inputs are not used to train the provider’s models. The Client may add to, restrict or withdraw approval of those tools by written notice.
Disclosure. On request, the Contractor will tell the Client which tools were used on any deliverable.
For a fixed-price engagement, replace the billing paragraph with one line: “Fees are fixed per deliverable as set out in the Statement of Work and do not depend on the tools used.” For a large engagement or sensitive data, have a lawyer in your jurisdiction review it. This is a template, not legal advice.
What should you do before your next invoice?
- Log the real time. Prompting, reviewing, testing and fixing, every fragment, rounded per day rather than per entry.
- Sort the work into two piles. Unknown scope stays hourly. Known deliverables that AI has compressed get a fixed price, set below what the old hours cost the client.
- Split your tool costs. Subscriptions go into the rate. Client-specific usage goes on its own line, at cost.
- Put the clause in your next statement of work. Before the work starts, not after the client asks.
- Itemise. Lines that name what was done answer questions before anyone asks, and clients who know the work can be fast will ask. More in what to do when a client questions your invoice.
All of it rests on an accurate record of the time you actually spent, the one thing the AI cannot produce for you. If you want that record without running a timer, Stunda records time in the apps you work in automatically. By hand, a five minute review at the end of each day catches most of the fragments and costs nothing.
Questions people ask
Should freelancers charge less if they use AI?
Per hour, no. If you bill hourly, a faster job already produces a smaller invoice, so there is nothing further to discount. What you can reasonably change is how you price, either a higher hourly rate announced openly at renewal, or a fixed price per deliverable that shares part of the time saving with the client while paying you more per hour worked.
Can I bill for time spent prompting AI?
Yes. Writing the prompt, gathering the context it needs, reading the output, testing it and correcting it are all work done on the client's problem, so they belong on the timesheet. What you cannot bill is the time the tool saved you, time spent learning a tool you use for every client, or minutes spent on another client while it ran.
Should I pass my AI subscription costs on to clients?
Usually not as a separate line. A subscription you use across every client is overhead, like your laptop or your internet connection, and it belongs in your rate. Usage-based costs incurred for one client, such as API calls on their project, can go on the invoice as an itemised expense at actual cost, ideally with an approval threshold agreed in advance.
Do I have to tell clients I use AI?
Check your contract first, because some clients require notice or restrict AI use outright. Beyond that, disclosure is not always mandatory, but it costs almost nothing and prevents a much worse conversation later. A short clause in the statement of work saying which tools you use, who is responsible for the output and how billing works covers most engagements.
Is it safe to paste client code or documents into an AI tool?
Only with permission and on the right terms. Confidentiality clauses usually restrict sharing with third parties, and an AI provider is one. If personal data is involved, data protection law may require the client's written authorisation first. Check whether the tool trains on inputs and how long it keeps them, and name approved tools in your contract.
Is value pricing better than hourly billing for AI-assisted work?
For well-defined deliverables, usually. When AI compresses the time on a job with a clear scope, hourly billing shrinks your income for an identical result. A fixed price lets the client pay less than before while you earn more per hour. For open-ended work such as debugging or advisory, hourly billing is still fairer, because nobody knows the scope yet.