Agentic retrieval lands on Amazon Bedrock Managed Knowledge Base

Small teams should expect new choices that affect cloud search costs, per-seat AI spending and developer workflows — plan for short testing cycles and a small budget to validate changes.

Amazon Bedrock: agentic retrieval for managed knowledge bases

Amazon Bedrock Managed Knowledge Base now supports agentic retrieval, which breaks a multi-part user query into sub-queries, runs separate searches, and judges whether it has enough evidence before returning a composed answer. The blog post compares standard vector similarity search with agentic retrieval, shows execution traces and compares the two retrieval paths on cost and when the cheaper approach is sufficient. Why it matters: Agentic retrieval will increase the number of searches and model planning steps for complex queries, raising query volume and operational cost but improving coverage of multi-part questions compared with a single-vector RAG query.

Excel: Microsoft removes the COPILOT() worksheet function

Microsoft will retire the COPILOT() cell function on 14 September 2026 and directs users to the Copilot side pane as the supported interface. The company updated its roadmap and told customers it will not move the function forward to general availability. Why it matters: Workflows that embedded AI calls directly in cells will stop working after the removal; teams that relied on COPILOT() must migrate formulas and macros to the side pane or alternative automation before the retirement date to avoid disruption.

Salesforce and Anthropic: Claudeforce and the flex-credit signal

Salesforce reported $11.3 billion in revenue for the quarter ending 31 July and announced a tie-up with Anthropic to deliver Claudeforce, which uses Claude alongside Salesforce data, workflows and governance. The offering includes a Plugin with 37 prebuilt “sales skills” and can be embedded in Salesforce products such as AIforce, Headless 360, Data 360 and Tableau. Why it matters: Customers deploying AI inside Salesforce will have a ready set of prebuilt skills to automate sales tasks; the company also flagged that a substantial share of bookings came from customers “refilling the tank,” signalling ongoing operational spend on consumable credits.

OpenAI: Business Premium seats raise the per-seat price and limits

OpenAI introduced ChatGPT Business Premium seats that deliver higher usage limits — five times the usage of a standard Business seat and exemption from a five-hour-per-day limit on advanced features — and carry a higher monthly price. The published prices are $125 per month for Premium seats or $100 if billed annually, versus $25 per month ($20 annually) for standard Business seats; help pages for Business model limits do not yet include GPT-5.6 usage details. Why it matters: Organisations that exceed standard Business limits will face materially higher per-seat costs if they upgrade to Premium; budget owners should measure current usage against the new caps before switching seats.

GitHub: ReviewBench for AI code review

GitHub published ReviewBench, an open benchmark designed to evaluate AI code reviewers across real pull-request diversity, review findings by severity and precision-recall trade-offs, and offline signals that predict production improvements. The benchmark aims to show what different reviewers catch and miss so teams can choose a reviewer aligned to their priorities. Why it matters: Development teams assessing code-review agents can run ReviewBench to compare tools on categories that matter (critical bugs versus small style notes) before wider rollout, reducing the risk of noisy or incomplete automated reviews.

Anthropic Cowork: the VM and model run in the cloud now

Anthropic changed Cowork so model inference and the previously shipped VM now run in the cloud, with each session getting an isolated sandbox; the desktop app remains responsible for file-access tool calls to fetch local files when needed. The move was made to avoid local disk, battery and performance costs and to keep sessions running when devices close. Why it matters: Teams that previously relied on local VM execution gain persistent sessions and lower endpoint resource use, but must review the desktop app’s file-access behaviour and data-mapping controls to ensure it meets their security and compliance requirements.

What we don't know

• Exact per-query or per-session pricing for agentic retrieval on Amazon Bedrock Managed Knowledge Base and how it scales with sub-query count.
• How Microsoft will support migrations for heavy COPILOT() users and whether any export tools will be provided.
• Salesforce contract-level terms for Claudeforce (billing model, credit use or bundling) and whether the 37 prebuilt skills carry extra fees.
• How OpenAI will price or limit access to GPT-5.6 under Business and Premium tiers — the help pages do not yet list GPT-5.6 usage.
• The precise security boundaries and audit trails for Anthropic Cowork’s cloud sandboxes and the desktop file-access tool.

What to do next

1. Inventory any workflows that call COPILOT() in Excel and convert them to side-pane processes or export logic before 14 September 2026 to avoid disruption.
2. Run short experiments: test agentic retrieval on a representative set of multi-part queries in Amazon Bedrock Managed Knowledge Base and measure query volume and relevance against your current RAG approach.
3. Audit per-seat usage for ChatGPT Business: compare current Business seat consumption to the Premium limits and decide whether to upgrade specific heavy users or optimise usage to avoid Premium costs.