What the pension enrolment assessor costs
The model reads what staff write, and nothing else. Assessing every worker in every pay run is a comparison between a number and a threshold row, and it costs rounding error whether the payroll holds twenty people or two thousand. Ninety messages a month is an employer in an enrolment or re-enrolment month. Here is where each cent goes.
Key takeaways
- About $2 a month at 90 messages. Roughly $4 at 400 messages.
- One Bedrock read per message is the only line that scales. Everything else is rounding.
- Nothing is always-on, so a quiet month genuinely costs almost nothing.
- The bill tracks messages, not headcount. Assessing five hundred people every week costs cents more than assessing fifty.
- The duplicate test runs before the read, so resends are free.
- The three real risks: a retry loop, storage nobody expires, and a bigger model than the job needs.
The bill at three volumes
These are US East prices at the time of writing, at three volumes that bracket most small businesses. Find the bar closest to your own and read across.
Line by line
| Line | At 90 messages | How it scales |
|---|---|---|
| Bedrock read | $0.45 | Linear. One call per message, roughly 1,800 in and 200 out tokens. |
| SES | $0.06 | Linear. About 2 messages per message. |
| DynamoDB + S3 | $0.35 | Storage grows with what you retain, not with throughput. |
| Lambda + SQS | $0.12 | Linear, and effectively free at this scale. |
| CloudWatch | $0.16 | Flat, if you set retention. Unbounded if you do not. |
| Secrets Manager | $0.40 | Flat. One secret, $0.40 a month. |
| AWS Budgets | $0.46 | Flat. Two actions, so you find out before the bill does. |
Roughly one model call per message, against a small, fast model. A message about the pension is two or three sentences, and the job is to pick one or more of four kinds and check for a signature or a personal-submission statement. That is classification with a short extraction. A larger model buys nothing here except a stronger temptation to let it write the reply, which it must never do.
The three ways this bill surprises you
Every one of these has happened to somebody, and all three are cheap to prevent.
- Sending the payroll export to the model. It is a fixed-header file with a row per worker per pay period. Parse it; a model asked to categorise workers would be slower, dearer and occasionally wrong about a threshold by 75 pence.
- Re-reading a whole thread every time a worker replies. Each message is read once, on arrival. A worker’s history is a query against stored classifications, not a prompt with the conversation pasted into it.
- Keeping every raw payroll export forever. The assessments, the duties and their evidence must be kept. The weekly export behind them does not need to sit in standard storage once it has been assessed and reconciled to the pay run.
What it costs when nothing happens
This matters more than the headline number for a seasonal business. In a month with nothing to process the bill is the fixed band: Secrets Manager at forty cents, AWS Budgets at forty-six, and a few cents of storage. Call it a dollar. There is no instance to stop and nothing to remember to turn off.
- Management
- Analytics
- Front-end & mobile
Set these on day one
- A dead-letter queue on every SQS queue, with a maximum receive count of three.
- Thirty-day retention on every CloudWatch log group. There is no default that is safe.
- An S3 lifecycle rule on the object prefix, tiering at 90 days and expiring at your actual record-keeping horizon.
- Two AWS Budgets actions — one that emails at half your expected spend, one at double it. The second is how you find out about a loop in an hour instead of a month.
- Provisioned concurrency: none. Nothing here is latency-sensitive enough to justify paying for a warm function.
Next: the same system drawn for engineers — service names, resource identifiers, IAM scopes, table schemas and the model id.
All posts