AI Leadership Weekly · Issue #103 · Tuesday 29 September 2026 · 08:00 BST
Good morning.
The final September edition brings another model release, more connections to workplace applications and a UK acquisition focused on secure digital delivery.
It would be easy to summarise this as “AI gets faster and goes everywhere”. A more useful reading is that businesses have more choices about where to place the work. A faster model might suit an existing task. A connector might remove a handoff. Specialist delivery support might make a sensitive use case feasible. Each changes a different part of the service—and each needs a different acceptance test.
IN 60 SECONDS
Efficiency: Anthropic’s Sonnet 5.5 announcement concerns task efficiency, not a blanket token-price cut. Connections: Gemini adds more workplace services, subject to connection settings. Delivery: IBM acquires a UK cybersecurity consultancy. The practical opportunity is to finish one useful end-to-end workflow before expanding the estate of tools surrounding it.
CEO / COO / CXO CHECKLIST
CEO: Choose the improvement that changes a customer or colleague’s result, not just the technology inventory.
COO: Check the full handoff between connected systems, including failure and correction.
CXO: Verify regional delivery and information-handling requirements against the actual proposed service.
TOP STORIES
1. Sonnet 5.5 claims faster work—not cheaper tokens
Anthropic · 28 September 2026
What happened. Anthropic introduced Claude Sonnet 5.5 on 28 September, reporting faster performance and up to thirty per cent lower task cost in its tests. API token prices remain $2 per million input tokens and $10 per million output tokens. The claimed saving is not a thirty per cent list-price reduction.
Why it matters. Our take: efficiency can come from fewer steps, fewer retries or less generated material, so a unit-price comparison may miss something useful. The reverse is also true: a favourable supplier test may not resemble your workflow. Separate price, consumption and the quality of the accepted output before deciding whether an upgrade changes the economics of the service.
What to do. Repeat a representative task set using the current model and the proposed replacement. Record total consumption, elapsed time and correction effort. Include unsuccessful runs in the comparison. Upgrade only where the result is satisfactory and the operational benefit survives that fuller calculation.
2. Gemini adds connections to more workplace applications
Google · 23 September 2026
What happened. Google announced new Gemini connected apps on 23 September, including services such as Airtable, Linear, monday.com and Zoho. Access depends on connecting the relevant service and its settings. An announced connector does not automatically confer authority over every record or action in the connected application.
Why it matters. Our take: the appealing opportunity is to remove a handoff between thinking about work and updating the system where it belongs. That only helps when the right record is changed, the permitted fields are clear and the user can see what happened. Treat a connection as the beginning of workflow design, not proof that the workflow has been completed.
What to do. Choose one small internal handoff, such as preparing a task update for review. Confirm the user’s permissions and show the proposed change before committing it. Check the destination record afterwards and decide how a mistake would be corrected before increasing the scope.
3. IBM adds UK security capability through Logiq Consulting
IBM · 24 September 2026
What happened. IBM announced its acquisition of Logiq Consulting on 24 September, adding UK secure-transformation and cybersecurity expertise. Financial terms were not disclosed. The announcement concerns delivery capability; it does not automatically establish that any particular customer service meets UK data-sovereignty requirements.
Why it matters. Our take: sensitive work may require more than a product configuration. Buyers need to understand where information is processed, who can support the service and how operational responsibilities are divided. A supplier’s local presence can be relevant, but it is not a substitute for examining the actual architecture, access model and contractual commitments for the proposed deployment.
What to do. For one sensitive use case, document the required processing locations, support-access limits and approval responsibilities. Ask a supplier to map its proposed service against those requirements, identifying exceptions explicitly. Avoid accepting a general statement about regional presence as the complete answer.
SIGNALS FROM THE LAST MONTH
21 September: IBM’s survey puts critical thinking and the work of validation into the workforce discussion. IBM / Oxford Economics · 21 September 2026
18 September: Embedded evaluation arrangements need clear scope, funding disclosure and responsibility for findings. Anthropic / Accenture · 18 September 2026
15 September: Google’s live-model options suggest different balances between responsiveness and deliberation. Google, updated 17 September · 15 September 2026
1 September: Anthropic’s planned customer-controlled safeguards remain a staged future deployment option. Anthropic · 1 September 2026
IN BRIEF
Further dated updates, including recent context worth keeping in view.
22 September: Anthropic launches Opus 5.5, another option to evaluate against demanding tasks. Anthropic · 22 September 2026
22 September: OpenAI expands the GPT-6 family with Sol and Luna; keep model versions explicit in evaluations. OpenAI · 22 September 2026
15 September lookback: Gemini Notebook’s study tools broaden the formats available for learning from source material. Google · 15 September 2026
3 September lookback: Daybreak’s defender-support programme combines access and assistance, with geographic rollout limits. OpenAI · 3 September 2026
THE 15-MINUTE PLAYBOOK
Minutes 0–4: Pick one workflow that crosses two applications. Identify the useful outcome and the person who is accountable for it.
Minutes 4–8: List the information passed between the systems. Mark which fields the assistant may read, draft or change, and which require a person’s decision.
Minutes 8–12: Walk through a normal completion, a wrong-record selection and an interrupted update. Require a visible result and a practical correction route for each case.
Minutes 12–15: Define a small supervised trial and the measure of improvement. Count the coordination work removed as well as the checking work added. Do not expand to the next five applications until the first connection produces a dependable result. A short chain that finishes is worth more than a long chain of promising integrations.
DATA WAVE MOMENT
The point of connecting tools is to improve the work between them. Data Wave combines the business process, information and application design so that an AI-enabled service has a clear job and an understandable outcome. Start with one worthwhile workflow, build it properly and use the evidence from that service to decide what should come next.
QUESTION FOR READERS
Which handoff would you remove first if the connected workflow still had to be clear enough for a colleague to check?
Brought to you by Data Wave — your AI & Data Team as a Subscription.
