AI Leadership Weekly · Issue #97 · Tuesday 18 August 2026 · 08:00 BST
Good morning.
This week, the AI race turns up in three places a business can recognise: the model doing the work, the team helping deploy it and the documents somebody needs to find.
Google has another Flash release. IBM and OpenAI are joining forces on enterprise delivery. ChatGPT is adding a more direct route into Google Drive. Taken together, these announcements suggest a useful way to review a proposal: separate capability, implementation and access to information. Improving one does not automatically fix the other two. That distinction could save a surprisingly long procurement meeting.
IN 60 SECONDS
Models: Test the latest Gemini release against the work, rather than the announcement. Delivery: A new IBM–OpenAI partnership adds another implementation route. Information: Google Drive access in ChatGPT has specific boundaries. The useful question is which of these changes removes a real obstacle for your team—and what remains to be done.
CEO / COO / CXO CHECKLIST
CEO: Identify whether the current obstacle is capability, delivery capacity or missing information.
COO: Include the handover into day-to-day operation in any implementation proposal.
CXO: Test document permissions and model changes separately; one successful demo does not validate both.
TOP STORIES
1. Gemini 3.7 Flash gives buyers another upgrade decision
Google · 13 August 2026
What happened. Google introduced Gemini 3.7 Flash on 13 August, with claimed improvements for coding and agent tasks. Its introductory token prices are lower than 3.6 Flash’s original launch prices. Those comparisons do not establish the cost or quality of a completed business process.
Why it matters. Our take: a fast release cycle makes a standing evaluation more useful than a fresh strategy discussion each time. Keep a small set of accepted examples, some difficult cases and a record of the current configuration. Then a release becomes a question the team can answer with evidence. There is no prize for adopting every model before understanding what the previous one was doing.
What to do. Choose a stable task and compare old and new configurations with the same inputs. Count corrections and retries as well as consumption. Record the result and retain a rollback route before changing a workflow that colleagues or customers depend on.
2. IBM and OpenAI announce an enterprise delivery partnership
IBM / OpenAI · 13 August 2026
What happened. IBM and OpenAI announced their partnership on 13 August, combining OpenAI technology with IBM’s consulting and enterprise-delivery capabilities. Plans include specialist implementation teams and work across core business operations. The announcement describes an intended offer, not independently established customer outcomes.
Why it matters. Our take: buyers should use implementation support to accelerate a specific change, not outsource the definition of success. A partner can help build the service, but somebody in the business still needs to decide how it works, accept exceptions and own the outcome. The handover matters: access, documentation and practical knowledge should not disappear when the initial team moves on.
What to do. Ask a supplier to show the deliverables for one bounded process, the responsibilities retained by your staff and the capability left behind. Include operating documentation and a supported handover in the acceptance criteria, alongside whatever the software itself must achieve.
3. Google Drive becomes easier to reach from ChatGPT
OpenAI, 13 August entry · 13 August 2026
What happened. OpenAI’s 13 August release notes add Google Drive browsing to ChatGPT’s Library for supported paid plans. The initial scope includes My Drive and directly shared files, but excludes Shared Drives. A connected account and its permissions remain part of the access path.
Why it matters. Our take: bringing documents into the working conversation can remove a tedious search-and-upload step. It also creates a practical test of information quality. Can the user distinguish the latest approved policy from an old draft? Does the answer identify what it used? A convenient connector is helpful, but it is not a substitute for knowing which document is authoritative.
What to do. Try a small approved document set containing a current version and an obsolete version clearly marked as such. Check what is retrieved and cited. Test with the intended user’s permissions, not an administrator’s account, and keep confidential material outside any unapproved trial.
SIGNALS FROM THE LAST MONTH
10 August: Google’s advertising updates bring more questions into the product, with beta and rollout limits. Google · 10 August 2026
6 August: Apptio’s value-and-cost preview puts baseline comparisons back on the agenda. IBM · 6 August 2026
4 August: The UK institute’s incident report illustrates why a test’s configuration belongs beside its findings. UK AI Security Institute · 4 August 2026
4 August: Lightwell’s research-access offer has defined institutional and geographic eligibility. IBM / Red Hat · 4 August 2026
IN BRIEF
Further dated updates, including recent context worth keeping in view.
11 August: IBM and Together AI announce an infrastructure agreement, with the new capacity expected in early 2027. IBM / Together AI · 11 August 2026
14 August: Microsoft announces Copilot app changes beginning on 18 August, distinguishing work and personal experiences. Microsoft, 14 August entry · 14 August 2026
10 August: Sonnet 5’s introductory API pricing becomes permanent; check the dated update, not an old price plan. Anthropic, dated update · 10 August 2026
10 August lookback: Multi-tenant agent management is in public preview, not a universal production capability. Microsoft, 10 August entry · 10 August 2026
THE 15-MINUTE PLAYBOOK
Minutes 0–4: Pick one delayed AI use case. Write the obstacle in plain English: the answers are poor, the information is missing, or nobody can get the change into operation.
Minutes 4–8: Find an example that demonstrates that obstacle. Keep it specific enough for another person to reproduce.
Minutes 8–12: Choose a test that addresses only that problem. A new model, a connector and a delivery partner are different interventions; do not change all three and then guess what helped.
Minutes 12–15: Assign the test and decide what would count as improvement. Record the things it does not establish. Removing a search problem does not prove the whole process is ready for automation—and acknowledging that makes the next decision easier.
DATA WAVE MOMENT
Our starting point at Data Wave is the work that is stuck. Sometimes it needs better software; sometimes it needs a better information source or a clearer decision. Separating those problems makes the delivery brief more honest and the build more useful. The aim is a service the team can use and understand after the initial excitement has passed.
QUESTION FOR READERS
What is holding your most promising AI use case back: the model, the information or the ability to deliver the change?
Brought to you by Data Wave — your AI & Data Team as a Subscription.
