By KYN AI Advisory Team — AI implementation specialists, Singapore
A finance agent has three costs per batch of 100 bills: the licence or build fee spread over volume, the model tokens used to read and classify, and the human review time. In most small teams, review time is the largest line and tokens are the smallest. Model usage on narrow tasks costs cents, not dollars. The number to manage is the share of bills that need a person, because each one costs minutes.
What are the three cost lines?
| Line | What drives it | Behaves like | |---|---|---| | Licence or build | Fee for the system, monthly retainer, hosting | Fixed per month, falls per bill as volume rises | | Model tokens | Pages per bill, steps per bill, model chosen | Variable, small, near linear in volume | | Review time | Share of bills flagged, minutes per flag, hourly cost of the reviewer | Variable, the largest, falls as rules improve |
A fourth line exists for a build: the one-off fee. Treat it like a licence and spread it over the months you expect to use it.
How do you cost licence or build per 100 bills?
Take the monthly fixed cost and divide by monthly volume, then scale to 100.
Formula: (monthly fixed cost ÷ bills per month) × 100.
Fixed cost includes whichever of these apply:
- a software subscription, if you buy a packaged tool;
- the monthly retainer, if you commission a custom system;
- cloud hosting for the database, the app and storage;
- the one-off build fee divided by the months of expected use.
Volume changes the answer more than anything else. The same fixed cost spread over 200 bills a month is five times the per-bill cost of 1,000 bills. That is why a custom build usually suits a team with steady volume and several workflows, while a packaged tool suits a team with one workflow and low volume. We do not publish a price list. Our pricing is a one-off build fee and a monthly retainer of at least twelve months, quoted after the review. See pricing.
For packaged tools, read the vendor's own pricing page and record the date you checked. Prices change, and per-document fees, seat fees and tier limits all vary. Compare on cost per 100 bills at your volume, not on the headline plan price.
How do you cost model tokens per 100 bills?
Tokens are the unit a language model charges by. A bill read by a model costs roughly the tokens in the document plus the tokens in the instructions plus the tokens in the answer. Three things move it: pages per bill, the number of model calls per bill, and which model you choose.
Measure it. Do not estimate from a vendor's rate card alone.
- Take 100 real bills from the last quarter, across your usual vendors.
- Run them through the agent end to end.
- Read the usage total from the model provider's billing page or API response.
- Divide by 100.
For a sense of scale, our own live systems report unit costs for narrow tasks: about S$0.05 per morning report to a department, about S$0.02 per outreach email with a checkable first line, about S$0.01 per reply sorted and about S$0.03 per group benefits comparison. These are KYN figures and depend on model and text length. A bill with a one-page PDF and a short extraction is a task of the same kind, so the unit cost is also in cents. We have not published a per-bill figure, because it varies with the bill layout and the model.
Reduce token cost with design, not with hope:
- Extract structured fields once and reuse them. Do not re-read the PDF at each step.
- Use rules for what rules can do. A duplicate check is a database query and uses no tokens.
- Use a smaller model for classification and a larger one only where extraction fails.
- Cache vendor templates so a known layout is parsed with fixed rules first.
How do you cost review time per 100 bills?
This is the line most business cases omit.
Formula: bills flagged × minutes per flag × cost per minute of the reviewer.
Take an illustrative case. These inputs are assumptions, not measurements:
- 100 bills in the batch.
- 15 of them flagged for review.
- 2 minutes per flag.
- A reviewer cost of S$0.60 a minute (S$36 an hour).
Review time is 15 × 2 = 30 minutes, which costs S$18. For comparison, entering 100 bills by hand at an assumed 4 minutes each is 400 minutes. At the same rate that is S$240. The gap comes from the human minutes saved, not from the model.
The sensitivity matters more than the base case. If the flag rate doubles to 30%, review time doubles. If false flags dominate, reviewers stop reading them. Track three figures weekly:
- the share of bills flagged;
- the share of flags that were real problems;
- the median minutes per flag.
Falling flag rates with a stable true-problem rate mean the rules are improving. A falling flag rate with a falling true-problem rate means the agent is missing things.
What does a full worked example look like?
Use your own numbers in this table. The values below are placeholders that show the shape.
| Line | Input (assumed) | Per 100 bills | |---|---|---| | Fixed cost | S$X per month ÷ 800 bills × 100 | S$X ÷ 8 | | Tokens | measured: total usage ÷ 100 | measured | | Review | 15 flags × 2 min × S$0.60 | S$18.00 | | Manual baseline | 100 bills × 4 min × S$0.60 | S$240.00 |
The only line you cannot fill from this article is the fixed cost, because it depends on what you buy. Fill it, then compute the break-even volume: the number of bills per month at which the agent's cost per bill falls below the manual cost per bill. Below that volume, a spreadsheet and a part-time clerk may be the right answer. We would say so.
What costs are easy to miss?
- Setup and rules tuning. The first two months have a higher flag rate while rules are tuned. Budget for it.
- Exceptions that never go away. Handwritten bills, scanned receipts and foreign language invoices need a person or a special path.
- Integration upkeep. APIs change. For example, Xero's API limits cap calls per minute per organisation, so a design that polls too often is throttled. Read the limit before you design the sync.
- Audit and retention. Logs cost little to store. Making them searchable costs engineering time. See audit trail for finance agents.
- The cost of a miss. A duplicate payment or a wrong vendor account costs more than a year of tokens. Weigh controls by what they prevent, not just by what they cost. See duplicate payment and vendor change checks.
Where KYN fits
KYN builds finance agents in a cloud account in your name, so hosting, model usage and logs are visible on your own bills. Rights in the custom code and configuration pass to you on full payment, and you can revoke KYN's role while everything keeps running. On our own books, bills arrive from Gmail into a register, receivables are chased at 3, 14 and 30 days and money movement waits for a named person. See finance systems.
The first 14 days are a review of your stack and interviews with your team. A full build ships in under 90 days.
Frequently asked questions
What is the cheapest part of a finance agent?
Model tokens, for narrow tasks like reading a bill. Measured on KYN systems, comparable tasks cost between S$0.01 and S$0.05 each. The fixed cost and the review time are larger.
How do I lower review time?
Cut false flags. Tune the thresholds per vendor, add a rule for each recurring false flag and show both records side by side in the review screen. Do not remove the flag on real risks such as a changed bank account.
When is a custom build cheaper than a packaged tool?
When volume is steady, you run several workflows on one data set, or the packaged tool charges per document or per seat in a way that grows faster than your volume. At low volume with one workflow, a packaged tool is usually cheaper.
Checked 30 October 2026.
Curious whether Marketing agents fits your business? Talk to KYN on WhatsApp — no forms, just a conversation.
