Optimus Consulting
The Week in AI
Edition 11
UK / Service SMEs
AI split into tiers this week. The skill is matching the model to the job.
Anthropic shipped a workhorse and a flagship in the same week. Every advisory voice said the same thing: let the best model plan and cheaper models execute. The single-model era is ending, and knowing which grade of AI to use where is becoming an operational skill.
Top stories
5 to knowTwo new Claudes landed: one for every day, one for the hard yards.
Anthropic released Claude Sonnet 5, a near-flagship model at everyday pricing, then followed with Claude Fable 5, its most capable model for agents and long-running work. Access to the restricted Fable and Mythos models also returned after US export controls were lifted.
Optimus angle
Don't default everything to the flagship. List what your team actually uses AI for and give the everyday work to the workhorse tier. Reserve the top model for the few jobs where quality moves money.
Strong models plan, cheap models do.
The week's clearest advice: use your most capable model to design and review work, and hand execution to cheaper models. Analysts framed three lanes for AI usage: cheap volume, balanced work, and frontier only, with fallback plans in case access changes.
Optimus angle
Use your best model once to write the runbook for a routine task, then let a cheaper model run it. Review the output weekly, not per task, and you cut cost without cutting quality.
Notion put AI agents on the team task board.
Notion 3.6 added External Agents, so Claude, Cursor and other agents can be assigned tasks on a shared board alongside human teammates, with audit-log visibility for enterprise users. Cursor and OpenClaw also shipped mobile apps for supervising agents from a phone.
Optimus angle
Pilot one board, one agent, one task type. Give the agent its own column and keep a human review step as the final column before anything ships.
The three powers rule: read, ingest, communicate. Pick two.
A widely shared governance briefing argued AI risk now comes from three powers: reading your data, taking in outside content, and communicating externally. An assistant holding all three is exposed to prompt injection and data leakage. The advice: limit any one assistant to two.
Optimus angle
Check your main assistant's settings against the three powers today. If it can read your files, browse outside content and send messages, switch one power off until you've tested the risk.
Cloudflare put a toll gate on AI crawlers.
Cloudflare moved to block mixed-use AI bots from ad-hosting pages from September and is pushing AI firms to pay publishers and separate search crawlers from training crawlers. How AI systems reach, ingest and cite content is being renegotiated, and that decides whose answers customers see.
Optimus angle
Ask whoever runs your website one question this week: can AI crawlers reach our content? If you're blocked, you're invisible in the AI answers your customers increasingly read first.
From the Insurance Desk
Sector signal for brokers, motor claims and credit hire companies.
Item 01
The FCA told insurers how it wants AI used.
The FCA's interim insurance director Graeme Reynolds urged insurers to innovate with AI while keeping consumer protections front and centre. It's active encouragement with conditions, not a warning shot, and it sets the tone for how firms will be assessed.
Optimus angle
Write a one-page note of everywhere AI already touches customer outcomes in your firm, quotes, claims decisions, correspondence. That list is the first thing a regulator conversation will need.
Item 02
Instant claims edge closer to straight-through processing.
Industry reporting says instant claims and straight-through processing are moving from concept to rollout, even as LexisNexis warns most motor claims handling is still fragmented and reactive. The gap between automated and manual books is widening.
Optimus angle
Map which of your claim types could settle with no human touch and which never should. For credit hire firms, faster insurer settlement changes hire-period economics, so know which of your files STP would shorten.
Editor's column
The Optimus Take
For two years the AI question in most firms was binary. Use it or don't. This week made the real question operational: which tier of AI does each task deserve? Anthropic now sells a daily driver and a flagship side by side, and the advice from every corner was the same. Plan with the best. Execute with the cheapest safe option.
For UK service SMEs this is good news. Most of your AI workload is routine: summaries, first drafts, first-pass reviews. The workhorse tier handles it at a fraction of the price. Triage your tasks the way a good claims team triages files. The discipline is knowing which few jobs deserve the expensive brain.
- StrategyDecide which few jobs deserve the flagship model, and make the workhorse tier your default for everything else.
- OperationsMatch every routine task to the cheapest model that can do it safely, and reserve the premium tier for work that earns it.
- ServiceClient-facing judgement work keeps the best model and a named human sign-off. Never economise on the work a customer sees.
“Stop asking which AI is best. Start asking which AI is best for this.”