01 / EFFICIENCY
The cost of capable AI is falling, but the useful unit is a finished task.
Anthropic's September 28 release says Claude Sonnet 5.5 produces output more than 30% faster than Sonnet 5 and can cost up to 30% less per task because it often needs fewer tokens. The listed token prices did not fall, which is an important distinction: efficiency improved because the model can complete some work with less effort. For an operator, the comparable measure is not the advertised price per token or per seat. It is the full cost of producing an acceptable estimate, analysis, customer response, or other completed deliverable.
02 / SPEND
AI budgets are separating into predictable access and variable work performed.
Microsoft's September 23 operating update describes two cost patterns: subscriptions that provide broad access at a predictable per-person price, and usage-based billing for longer agent work that varies by task. Small businesses will increasingly encounter both. A flat monthly plan may look inexpensive while a frequently repeated workflow creates variable charges; a usage fee may look high while replacing hours of manual effort. Without a task-level measure, owners can cut productive work or keep paying for activity that does not improve an outcome.
03 / CONTROL
Scaling AI requires an improvement loop, not a one-time rollout.
Anthropic's October 1 announcement about Barclays emphasizes disciplined governance and real operating outcomes as the bank expands AI across complex systems. Microsoft's guidance makes a related point: an agent is not finished on deployment day; teams must define good work, measure it, and improve it over time. The scale is different for a small company, but the management lesson is the same. One owner, one quality standard, and one recurring review matter more than a large implementation program.
THE CJC VIEW
Implementation beats experimentation.
AI should earn its place in the operating budget the same way any process change does: by improving a result after the cost of review, correction, and follow-up is included. Faster output is useful only when the finished work is accurate, usable, and delivered with less total effort.
Small businesses do not need elaborate AI accounting. They need a short scorecard for each recurring workflow: volume, direct tool cost, human review time, rework, cycle time, and the business result. That makes model, plan, and automation decisions much easier to defend.
PRACTICAL NEXT STEP
Build a one-page AI workflow scorecard
- Choose one recurring AI-assisted task that happens at least weekly and has a clear finished result.
- Record the current monthly volume, software or usage cost, average completion time, reviewer time, and common corrections.
- Define one quality threshold and one business outcome, such as an approved estimate, a resolved ticket, an invoice sent, or a report delivered on time.
- Track ten completed jobs and calculate the total cost per accepted result, including human review and rework rather than counting generated drafts.
- Review the scorecard monthly and change the model, instructions, approval step, or workflow only when the evidence shows a better operating result.
Lower model costs make more AI experiments affordable. They do not remove the need to manage the work. Measure what reaches the finish line, include the human effort around it, and use that evidence to decide what deserves to scale.
Sources and further reading
- Introducing Claude Sonnet 5.5 — Anthropic
- Building the system for AI at work — Microsoft
- Barclays scales Claude to upgrade operations and improve client experience — Anthropic