How should you invoice AI-assisted work?

A team uses an AI coding agent to complete a migration workstream in 18 human hours instead of the 50 hours originally expected. The work meets the agreed scope. A senior developer reviewed the output, corrected several edge cases, and remains responsible for what ships.

What should appear on the invoice?

Should the agency bill 18 hours because that is the time people actually spent? Charge the fixed fee the client already accepted? Add the cost of the model? Apply a discounted rate to AI-assisted time? Charge for the value of the completed migration rather than the production method?

There is no universal answer. The right model depends on the agreement, the predictability of the work, the client’s expectations, and which risks each party has agreed to carry.

What has changed is the need for evidence. When human effort, AI contribution, and delivered output are recorded separately, a firm can choose a pricing model deliberately. Without that record, the same firm is forced to guess, hide the role of AI, or defend an invoice after the fact.

Start with how the work is priced

AI does not change the commercial terms by itself.

If the client pays for actual hours, bill the human time actually worked rather than an imagined number of “equivalent hours.” If the client accepted a fixed fee for a defined result, efficient delivery does not turn the engagement into hourly billing. If compute is treated as overhead, adding it as a surprise pass-through will damage trust even when the amount is small.

Before comparing models, separate three records:

1. Delivery record: What was produced, under which project and workstream, and who accepted responsibility?
2. Effort record: How much human time was actually spent?
3. AI contribution record: What the AI system contributed, what it cost, and how the output was reviewed?

These facts are related, but they are not the same unit. A sound invoice policy decides which facts determine price and which remain explanatory context.

Six ways to price AI-assisted work

1. Bill actual human hours

The simplest approach under time and materials is to invoice the human time actually worked. AI tooling remains a cost of delivery, like development software or cloud infrastructure.

How it works

• Record preparation, prompting, review, correction, integration, and client communication.
• Invoice those hours at the agreed role or blended rate.
• Keep agent activity and compute cost in the internal project record unless the agreement requires disclosure.

Works well when

• The contract explicitly requires actual time.
• The client values flexibility and accepts that scope may change.
• AI efficiency is one of several ways the firm improves delivery.
• The firm is comfortable retaining less revenue when the same work needs fewer human hours.

Main risk

The firm can be punished for efficiency. Better tools and workflows reduce the number of billable hours even when the delivered value stays the same or improves. This can make an AI-forward team economically worse off under a model that rewards time consumed.

What to disclose

Be accurate about material AI involvement if the client, policy, or risk profile requires it. Do not add fictional AI hours to restore the original estimate.

2. Charge a fixed fee

A fixed fee prices an agreed scope or deliverable rather than the hours eventually required.

How it works

• The parties agree on scope, acceptance criteria, timing, and price.
• The firm manages the mix of human effort and AI contribution internally.
• AI-enabled efficiency can improve margin if quality and scope remain controlled.

Works well when

• The work can be scoped with reasonable confidence.
• Acceptance criteria are clear.
• The firm has enough delivery history to price risk.
• The client wants cost certainty.

Main risk

The firm owns estimation and delivery risk. AI may accelerate the first draft while increasing review, correction, or integration work. A fixed fee based on optimistic assumptions can turn apparent efficiency into margin loss.

What to disclose

Explain material AI use according to the agreed disclosure posture. The price does not need to be decomposed into human and AI units unless the agreement says it should be.

3. Price the value or outcome

Value-based pricing connects the fee to the economic importance of the result rather than the resources consumed.

How it works

• The firm and client define the outcome and its value.
• Price reflects a share of that value, the risk assumed, or the importance of the result.
• Production method becomes secondary as long as quality, responsibility, and constraints are satisfied.

Works well when

• The outcome can be defined and valued credibly.
• The provider has strong differentiation and trust.
• The client can connect the work to revenue, cost, risk, or strategic importance.
• Both parties are willing to discuss value openly.

Main risk

Many professional deliverables do not have a clean measurable outcome. Clients often buy a website, research report, migration, or campaign rather than a provable amount of business value. Calling a price “value-based” does not create evidence of value.

What to disclose

Focus on the delivered result, responsibility, and quality controls. AI attribution can reinforce trust, but token counts and agent duration should not become a substitute for the value case.

4. Apply an AI-assisted rate or multiplier

Some firms and clients may agree that hours with material AI acceleration are billed at a different factor. For example, an AI-assisted hour might be billed at 0.5 times the standard hourly rate, while human review and specialist judgment remain at the full rate.

How it works

• Define what qualifies as an AI-assisted hour.
• Agree on the multiplier before the work begins.
• Record actual human time and material AI involvement.
• Apply the factor only to the eligible entries.

Works well when

• The client wants a visible share of AI-enabled efficiency.
• The firm wants to remain within a familiar hourly structure.
• Attribution coverage is high enough to apply the rule consistently.
• The eligible activities can be defined without constant negotiation.

Main risk

The model can accidentally imply that AI-assisted professional judgment is worth less. It also creates classification pressure. Teams may debate whether an entry was “assisted enough” to receive a discount.

What to disclose

Show the agreed rule, eligible hours, applied factor, and human review. Do not advertise “savings” unless the figure comes directly from agreed invoice math.

5. Pass through AI compute cost

Compute pass-through treats model usage as a reimbursable project expense.

How it works

• Attribute provider cost to the correct client and project.
• Pass it through at cost or with an agreed markup.
• Show the amount separately from professional fees.

Works well when

• Compute cost is material.
• Usage can be attributed accurately.
• The client expects variable infrastructure expenses.
• The agreement defines eligible providers, models, and markup.

Main risk

Tokens are a poor proxy for delivered value. Passing through a few dollars of compute can add administrative complexity without meaningful revenue. It can also encourage the wrong comparison between model cost and professional fees.

What to disclose

Show actual usage cost and the agreed treatment. Avoid presenting tokens as labor or as proof that useful work occurred.

6. Use a hybrid model

Hybrid pricing combines elements based on the shape of the engagement.

Examples include:

• A fixed fee for the defined deliverable, with hourly billing for approved scope changes
• A retainer for access and capacity, with project fees for major outcomes
• Full rates for human review and advisory work, with an agreed multiplier for repeatable AI-assisted production
• A fixed professional fee with material compute passed through at cost

Works well when

• Different workstreams carry different levels of uncertainty.
• The client wants both predictability and flexibility.
• The firm can explain the model without turning every invoice into a negotiation.

Main risk

Complexity. A theoretically perfect pricing model can fail if delivery teams cannot classify work consistently or clients cannot understand the invoice.

One project under six models

Consider an illustrative migration workstream:

• Expected human effort before delivery: 50 hours
• Actual human effort: 18 hours
• Standard rate: $200 per hour
• AI compute cost: $120
• Agreed fixed fee option: $10,000
• Illustrative AI-assisted multiplier: 0.5 times the standard rate
• Human review and architecture time: 8 of the 18 hours
• AI-assisted production time: 10 of the 18 hours

Here is how the invoice basis changes:

These are not six ways to describe the same economic reality. They allocate risk and efficiency differently.

Under actual-hour billing, the client receives nearly all of the efficiency gain. Under a fixed fee, the firm keeps the upside but also bears the risk if the work takes longer than expected. Under the multiplier, the parties share some efficiency within the hourly model. Under compute pass-through, the client reimburses a measurable input that may be economically trivial compared with the professional result.

The table does not reveal which model is “fair.” Fairness comes from a clear agreement, honest records, and a price that both parties understand before the invoice arrives.

What belongs on the invoice?

An invoice should be clear enough for the client to understand without exposing every internal production detail.

For most firms, a strong AI-aware invoice contains three layers:

Commercial line item

Name the project, workstream, outcome, quantity, rate, or agreed fee in the same structure used for other work.

Plain-language context

If disclosure is appropriate, explain material AI involvement in one sentence. For example:

AI was used to generate an initial migration test suite. The engineering team reviewed coverage, corrected environment-specific assumptions, ran the final suite, and approved the delivered work.

Supporting detail

Provide more detail only when the contract or client needs it. This might include reviewed status, eligible AI-assisted hours, the agreed multiplier, or compute passed through at cost.

Avoid dumping model traces, prompt histories, or token tables into a client invoice. Technical exhaust is not automatically commercial clarity.

Four questions to agree before the work begins

AI billing disputes are easiest to prevent at proposal or kickoff stage.

Ask:

1. What determines the fee? Actual time, defined scope, value, usage cost, or a combination?
2. What AI use must be disclosed? Any use, material contribution, selected workstreams, or only exceptions?
3. How will review be recorded? Who is accountable, and what does reviewed or verified mean?
4. Which evidence can the client request? A disclosure statement, activity summary, tool and model list, usage cost, or a fuller audit record?

Agree the answers with the client before work begins, then carry them into the delivery process. A policy that exists only in finance will fail when the team logs the work.

Price the agreement. Attribute the work.

AI-assisted work does not require every firm to abandon hourly billing or adopt value-based pricing. It does require firms to stop treating human time as a complete description of delivery.

The pricing model determines how money changes hands. The attribution record explains how the work happened.

Keep those concepts separate, then connect them deliberately. Record actual human effort. Record material AI contribution. Name the responsible reviewer. Describe the outcome. Apply the pricing rule the client agreed to.

That produces an invoice a firm can defend without hiding its methods or discounting its judgment by default.