Track and cap AI spend across clients: a spreadsheet template for agencies

This page is for an agency owner or the person who looks after the money. The problem is simple: AI credits are pooled, several people use the same account, and an AI agent can spend more than you expect on one task. If you cannot say what a client's work cost, you cannot price the next job or notice the margin shrinking.

I give you a ledger you can copy, a way to cost the attempts that failed, and a few billing models. I also say plainly what Higgsfield, an online platform for AI images and video, can and cannot track, because for client attribution it is not the answer. A spreadsheet is.

This page may contain affiliate links from which Blur Studio earns commissions.

The problem: credits disappear and invoices do not match

Three things make attribution hard:

• Pooled credits. A credit is the unit most AI tools charge in. Team plans pool credits across all seats, so the bill shows one total, not one per client.
• Failed attempts. A clip that took five tries to get right cost five tries. If you only bill the keepers, you eat the other four.
• Agents. When an AI agent (a program that plans and carries out a task in steps) does the work, its own thinking uses credits too. Higgsfield's own published build log of a website shows the agent reporting 380 credits for a task while the usage panel showed 1,479.32. The agent does not count its own reasoning. So take the number from the platform counter, never from the agent's memory.

Start with a spreadsheet: the free option that works

Higgsfield does not tag spend by client (more on that below), and each tool you add would have its own counter. A plain spreadsheet costs nothing and works across all of them. The rest of this page is how to set one up.

What to log for every job

One row per generation job (one batch of tries for one deliverable). Here is a layout you can copy.

Column	What goes in it
Date	When the job ran
Client	Use one fixed spelling per client
Project or job ID	Your own reference, such as ACME-0312
Deliverable	For example "15 second product clip"
Tool and model	The platform and the model picked
Settings	Resolution, seconds, number of images
Cost	Credits or dollars, copied from the usage panel
Source of credits	Monthly plan, pack, promo or API balance (they cost different amounts)
Attempts	How many tries it took
Approved attempts	How many tries a person approved
Used in final	Yes or no
Operator	Who ran it
Notes	Anything odd, such as a refund after a failed run

Give the file five tabs: Log (the table above), Client budgets (monthly cap, spent, remaining, alert flag), Rate card (credit to dollar), Monthly summary (a pivot table, which adds up the log by client) and Invoice lines.

Turn credits into dollars once a month

Use the plan you actually pay for, not a list price in your head. The effective rate is plan price divided by credits included. Using the Higgsfield prices on its pricing page on 3 November 2026 (check the live page, they change):

Plan	Price a month	Credits	Dollars per credit
Starter	$19	270	about $0.070
Plus, monthly billing	$59	1,200	about $0.049
Plus, annual billing	$47	1,200	about $0.039
Ultra, monthly billing	$129	3,000	about $0.043
Ultra, annual billing	$99	3,000	about $0.033

Credits from a top-up pack, auto-refill or a promotion cost a different amount and expire. Higgsfield's help center says pack, auto-refill and boost credits last 90 days, while subscription credits reset each month (every 30 days on annual plans). Keep the "source of credits" column so you do not under-cost a job paid for with expensive pack credits. For how plans work, see the Higgsfield pricing explainer.

Cost per approved asset: a worked example

All numbers here are made up to show the method. Suppose a client wants one approved 10 second clip. Your team spends 40 credits on 5 attempts, and one attempt is approved. The cost per approved clip is 40 credits, not 8. At a made-up $0.05 per credit, that is $2.00 of tool cost for the clip.

The formula is:

Cost per approved asset = total spend on the client that month divided by approved assets that month.

Bill and price from that number. It is the figure that protects your margin. If you want to cut the number itself, see tips and settings that save credits and how to spend credits wisely. Those pages are about spending less. This page is about knowing who spent it.

Set a cap per client, and what to do at 80 and 100 percent

Pick a monthly cap for each client from the cost per approved asset and the number of assets in the retainer. Put it on the Client budgets tab with a simple formula: spent divided by cap.

• At 80 percent: a message goes to the account lead. Ask if the remaining work is on plan.
• At 100 percent: generation stops for that client until the lead agrees to an overage or a change order.

The platform will not enforce this for you, so the rule has to be a habit: nobody starts a new job for a client without checking its row first.

If you can, give each client its own key or account

A spreadsheet is a fix after the fact. The stronger fix is to make the tool itself do the sorting. A gateway company, CometAPI, writes that agencies can issue one API key (a password a program uses to call an AI service) per client, so each key becomes one client's bucket and the bill is read off the dashboard instead of rebuilt from logs. It also notes that revoking one client's key does not touch the others. That is a vendor describing its own product, so treat it as one way to do it, not proof. The same idea works with separate logins or separate balances where a tool allows it. Where a tool only has one shared login, as Higgsfield's web plans do, you are back to the spreadsheet.

Find the spend you cannot see yet

Before you trust any total, list every place AI money leaves your business. The finance software company CloudZero suggests starting with an inventory of every source: provider accounts, coding assistants, add-ons inside other tools, and subscriptions people expense on their own cards. Its advice for the first pass is to get a complete list and a rough total, then improve it. For a small agency that means one hour with the card statements. Match each line to a client or to "overhead". A rule from the same source is worth borrowing: having 80 percent of spend mapped to a named owner beats arguing over 100 percent.

Two more habits help. First, give each client one number that answers a real decision, such as cost per approved asset. If a figure in your report does not change what you do, drop it. Second, set an alert for a stuck loop. Zapier's chief operating officer describes a limit that messages a person at 75 percent of their monthly allowance, mainly to catch runaway spend such as a stuck agent loop, and says most people never get the message. Your 80 percent rule does the same job.

What Higgsfield can and cannot track

For fairness, here is what Higgsfield's own help center says, read on 3 November 2026:

• Web plans: usage by member can be viewed in Manage Account. Credits pool across seats and assets are shared. Team plans (up to 9 seats) show no member spend controls in Higgsfield's plan comparison. Scale (up to 15 seats) can pause a member's spend. Enterprise has per-user caps, pause and usage reports. Workspace admins on any plan may access members' content. All of this is by member, not by client.
• API: the API is a separate product with a prepaid dollar balance, and its console at open.higgsfield.ai has an Analytics view of spend and request volume over time that exports to CSV, and a Billing view with balance, payment methods and invoices. You can sign up as an individual or as a company. The balance cannot go below zero (requests wait), which acts as a natural hard cap. Higgsfield's help center describes no per-key or per-client budget, so ask support before assuming one exists.
• Agent work: Higgsfield's Supercomputer agent shows a running cost as a task proceeds and deducts credits at standard rates per step. Complex tasks use more credits than a single generation. MCP use (letting an outside AI assistant such as Claude run Higgsfield) deducts plan credits and has no native spend cap, so your own weekly rule has to do the capping.

So Higgsfield contributes two things to a client ledger: member-level usage on the web, and a CSV export on the API. For client attribution, copy figures into your sheet, or use a naming convention in project and file names that lets you match jobs to clients later. One idea for the API is to fund a separate balance for each client, which makes the balance itself the cap. Whether that suits your setup is not documented, so check with Higgsfield first.

Billing models: at cost, with a fee, retainer allowance or absorb

There is no single norm. Digiday reported on 3 March 2026 that agencies differ: Merge passes metered costs case by case, Big Spaceship treats compute as a production cost, RPA absorbs it, and Lerma shows AI costs at cost as a line item. Ebiquity's chief executive said he expected the consultancy to be asked to audit an agency's token usage, as it audits media spending now (a token is a small unit of text that language models are billed in).

Model	How it works	Suits
Pass-through at cost	AI cost is a line item on the invoice	Clients with a finance team that wants detail
Pass-through plus a handling fee	Cost plus a clearly labelled fee for setup and review	Heavy AI work with real management time
Retainer with an allowance	A monthly credit allowance included, overage at a stated rate	Steady monthly work
Absorb	You carry the cost inside your fee	Small clients, or while you test

One consultancy's page (Blck Alpaca) calls a 30 to 50 percent margin on pass-through costs the industry standard. It sells agent services and is the only source I saw for that figure, while Digiday shows agencies at cost, absorbing and metering. Treat it as one opinion, not a rule.

The contract clause: caps, overage, revisions, price changes

Put this in your proposal or statement of work. Edit it with a lawyer for your country; this is information, not legal advice.

AI production costs. The monthly fee includes up to [N] approved AI-generated
assets. Additional approved assets are billed at [rate] each. One round of
revisions per asset is included; further rounds count as new assets. AI tool
costs are [included / passed through at cost / passed through plus [x]%].
If tool prices change by more than [x]%, we may adjust the rate with 30 days'
written notice. Usage reports are available on request.

Counting revisions matters most. A "small change" that needs a full regeneration is a new asset in cost terms, so say so before the first job.

A 15 minute weekly review

1. Export or copy the week's usage from each platform's counter.
2. Compare it with the Log tab. If they differ, find the missing rows.
3. Check each client's spent-to-cap figure. Flag anyone above 80 percent.
4. Check cost per approved asset by client. A rising number means more retries or more complex requests.
5. Decide which clients need a conversation about scope.

If an AI agent ran any job, reconcile against the counter, not against the agent's own report.

When a spreadsheet stops working

A sheet is enough for a handful of clients and a few thousand rows. It breaks down when many people log jobs by hand and rows go missing, or when you want one live view across several tools. At that point, a small internal tool or a shared system fed by the platforms' exports is worth building. If you build your own pipeline on an API, a gateway (a service that sits between your software and the AI vendor and logs every call) can tag calls by client. The Blck Alpaca page names Helicone and Portkey as examples that track cost per client by API key. The page does not say more, so check what they support. A tool that can cap credit per project, such as Creatomate (a service that makes videos from templates), shows what a native cap looks like; Higgsfield's web plans show none in its help center.

Who should skip this

• You work for one client only. Your plan bill is your client cost.
• You use AI for internal work only and never bill for it.
• Your clients pay a flat project price and you never expect to change the scope. Even then, I would keep the Log tab for your own margin.

Also check rights and consent for any real face or voice in generated work. Outputs from AI tools may not be unique, and you should keep real footage when a claim must be real.

How I'd approach this

I would start with the Log tab only, for one month, with no caps. After a month you know your real cost per approved asset by client, and the caps and the clause follow from the data. If you decide you need member-level controls such as pausing a person's spend, check the Scale plan on Higgsfield's pricing page and confirm the details with them, because Team has none. Several other pages here help on the cost side: delivering video volume at an agency and a client approval workflow.

When the sheet is no longer enough, Blur Studio builds the repeatable workflow around these tools: usage logging, a client approval portal and a shared asset library. See how I work with business owners.