AI Leadership Weekly · Issue #104 · Tuesday 6 October 2026 · 08:00 BST
Hello.
What would you want to see before renewing an AI contract? A bank can count searches, an engineering team can count qualifications and a model supplier can quote a price per token. The harder figure to find is what it costs to give one customer an answer they can act on. This week, Barclays, Anthropic and OpenAI give us three parts of that calculation.
In 60 seconds
Barclays reports over a million searches on its colleague knowledge assistant. Follow a sample through to the customer’s next step.
Anthropic is funding deployment skills. Its engineering residency ties training to a project each participant takes back to their organisation.
OpenAI quotes lower API rates for GPT-6.1 Sol than for Astra. Include the reviewer’s time when comparing models for a particular job.
This week’s decision. Choose the completed service outcome that will determine whether your next AI pilot earns more budget.
What changed
Barclays gives its colleagues a place to look
Anthropic says more than 16,000 Barclays colleagues have adopted the UK bank’s internal knowledge assistant since it went live in 2025. Staff use it to find information when helping retail customers. Elsewhere in the bank, models help sort and route incoming Global Markets emails. These are different jobs with different measures. For a knowledge search, look for an accurate answer staff can trace to the current policy. For a routed email, check whether it reached the right desk with the details needed to act. A fast transfer to the wrong team adds another queue. Anthropic on the Barclays rollout · 1 October
Training gets a named project
Anthropic announced a $100 million commitment to its Frontier Academy and an aim to train 10,000 engineers by the end of 2027. Participants nominated for the first programme bring a project from their employer. After an initial practical assessment, successful participants spend 12 weeks working on it. For a manager buying training, that suggests a good question to ask before paying the fee. When the engineer comes back, which service owner has agreed to give them access, time and a real problem to solve? Anthropic’s academy announcement · 2 October
The model quote is only one line of the bill
OpenAI lists GPT-6.1 Sol at $2 per million standard input tokens and $10 per million output tokens, against $10 and $50 respectively for Astra. Consider an illustrative batch using one million uncached input tokens and 100,000 output tokens. At those rates the model charges would be $3 and $15. The $12 difference is small enough that a few extra checks could change the choice. Run both models on the same completed cases, mark which answers pass, and time how long it takes to put the others right. OpenAI launch pricing · 29 September
CEO / COO / CXO checklist
CEO. Ask which customer or colleague gets a better result when the pilot succeeds.
COO. Record a reopened case against the original request so the extra work stays visible.
CXO. Ask the service owner who updates the source material when a policy changes.
The 15-minute playbook
Minutes 0–5. Bring five recently closed enquiries to the team meeting. For each, write down what the customer needed and how the team knew the work was finished.
Minutes 5–10. Pick one enquiry and trace every handoff. Mark the minutes spent searching, checking, correcting and waiting. Circle the step where the proposed assistant would help.
Minutes 10–15. Write the pilot’s scorecard on one line. Use cost per correctly resolved enquiry, paired with the share that needed a second contact. Give a colleague the five cases and see whether they would score them the same way.
Question for readers
Which cost disappears from your AI business case when a customer has to ask the same question twice?
Brought to you by Data Wave — your AI & Data Team as a Subscription.

