Checking one order for a missing delivery instruction might save a few minutes. Checking every order before it reaches the warehouse could prevent an afternoon of phone calls. The obstacle is often the effort required to perform the check consistently.
Falling AI prices make it worth revisiting these small jobs. A task that looked uneconomic when assessed individually may become useful when it runs cheaply across a steady stream of records. For a small business, that opens up a different kind of automation: routine attention to details that currently depend on somebody remembering to look.
The useful question is which neglected checks your business can now afford to perform, and whether their results justify the total cost.
What actually became cheaper?
Anthropic’s 7 October 2026 Haiku 5.5 announcement lists input and output token prices for prompts up to 100,000 tokens at one tenth of Haiku 4.5’s rates. The company also notes that the new tokenizer can use more tokens for the same work. Longer prompts have a different price reduction.
Those are model charges. They exclude connecting your systems, preparing information, checking answers, handling failures and maintaining the workflow. A tenfold reduction in one price component will have a smaller effect where people or integration account for most of the expense.
Treat the announcement as a reason to retest a business case. It does not establish that a particular workflow is ten times cheaper, equally accurate or suitable for unattended use.
Look for useful work below the attention threshold
A good candidate is a question your team understands, can answer from available records and would benefit from asking more consistently.
Consider a small distributor receiving orders by email. Staff can spot a missing postcode, but a message about an unusual unloading requirement may be buried several replies back. A proposed assistant could assemble the relevant order details and flag what needs clarification before the job is released.
Another candidate is supplier correspondence. A small model could classify incoming notices, extract stated changes to delivery terms and place potentially relevant changes in a review queue. A person would check the original notice before altering purchasing decisions.
An agency could apply the same approach to project updates: identify requests that appear to introduce a new deliverable, then attach the source message for the project manager. Each example is a proposed workflow, not a claim that a model will understand every exception correctly.
Choose one decision with a visible result
Keep the first brief specific. For the distributor, the task might be to check whether each order includes the fields needed for dispatch and identify conflicting information.
Define the expected output before choosing a tool:
- The order identifier and links to the records checked.
- The missing or conflicting detail, in plain language.
- The evidence that caused the flag.
- A suggested question for a colleague to review.
- A clear indication when the available information is insufficient.
The assistant should distinguish a missing detail from an inaccessible document. If it cannot read the attachment, it has not completed the check. That distinction prevents an apparently clean result from concealing a broken connection.
Use ordinary validation rules for structured fields where they work. AI is more useful for interpreting free text and connecting relevant context than for checking whether a required spreadsheet cell is empty.
Calculate the cost of a completed check
Here is an illustrative monthly calculation. Suppose an order-checking workflow costs €100 in model usage and €300 in review and operating effort. Reducing the model component to €10 lowers the total from €400 to €310. That is a 22.5% reduction, before any changes in quality, setup cost or workload.
Now suppose the cheaper model also produces more ambiguous flags. If reviewing those flags adds €120 of staff time, the total becomes €430. The cheaper model has produced a more expensive process.
Use your own figures, including a sensible allocation of initial implementation costs. Count useful checks completed, rather than API calls made. Record whether somebody acted on each important finding and whether that action avoided rework or delay.
Time recovered is capacity. It becomes a cash saving only when an actual cost falls, and it becomes additional revenue only when the business converts that capacity into paid work.
Test the awkward cases before increasing volume
Select a small set of past orders that your team can assess independently. Include normal orders, incomplete messages, a later correction, conflicting addresses and an attachment the system cannot open.
Have a colleague establish the expected result. Then compare the assistant’s output against it. Count both false alarms and missed problems; reviewing only the flagged records will hide failures among records passed as complete.
Keep some examples out of the initial configuration work so the final check uses unfamiliar material. If you change the model, instructions or source connections, rerun the same evaluation before relying on the new version.
A low usage price makes experimentation accessible. It also makes it easy to run a weak process thousands of times before anyone notices.
Give the workflow an operating limit
Start with a daily batch and a cap on records processed. Keep the pilot read-only, with results delivered to a named colleague. That person should know how to pause the process and how to check whether a missing report means there was nothing to flag or the run failed.
Set a spending limit in the available billing or infrastructure controls. A sentence in a prompt is not a spending control. Also define what happens when the source system is unavailable, a document is duplicated or a run stops halfway through.
After a short trial, review the useful findings, the missed cases, the time spent supervising and the full operating cost. Expand only if that evidence supports it. Some tasks will remain better served by a fixed rule or an occasional human check.
Put cheaper AI to work on one recurring problem
XYZ by FORMATION helps small teams choose useful automation, connect existing systems and test the result against real work. We can help separate model cost from the cost of running a dependable process.
Our AI workflow assessment starts with the work your team does today. Bring one check that is regularly skipped because nobody has time for it. We can assess the sources, define a measurable pilot and identify who will operate it after handover.
Product pricing checked against the linked announcement on 12 October 2026. Prices and product capabilities may change. Business scenarios and calculations are illustrative, not measured XYZ customer results.