This Month in AI

The 60-second version

August was a plumbing month. Almost nothing new got demoed; a great deal of billing, licensing and legal machinery moved underneath products you are already paying for. Five things changed:

1. Microsoft deleted the release calendar. On 25 August it retired the twice-yearly release wave model for Dynamics 365, Power Platform and Dataverse. There is no "2026 release wave 2 plan" document to read. Roadmap content moves to one continuous AI at Work roadmap in September, and the Release Planner site retires 15 November 2026. If your semi-annual planning checkpoint was the wave document, that checkpoint no longer exists.
2. The new Copilot Studio bills outside your Copilot licence. Microsoft's August licensing guide splits Copilot Studio into three harnesses. Standard and Copilot Chat stay inside the Microsoft 365 Copilot entitlement under fair-use limits. The new agentic GitHub Copilot harness does not — it draws Copilot Credits at $0.01 each for authoring and runtime, beginning 1 September 2026.
3. Frontier prices halved again, and the increase you budgeted for was cancelled. GPT-5.6 now lists at $2/$10, $1/$6 and $0.10/$0.60 per million tokens — half the 30 July rates. Anthropic scrapped Sonnet 5's scheduled rise; $2/$10 is now the standard price. Anything you costed before August is costed wrong, usually by a factor of two.
4. Two AI compliance deadlines now have dates on them. EU AI Act transparency obligations took effect 2 August 2026 and reach any business whose AI output touches EU users, SMBs included. Connecticut's Public Act 26-15 requires AI disclosure in WARN notices from 1 October 2026, with broader hiring-notice duties from 1 October 2027.
5. A governance control appeared and vanished inside a week. Microsoft announced admin-configurable domain blocklists for Copilot web grounding on 28 July, then withdrew the feature around 4 August with no explanation and no replacement.

Microsoft: a quiet month on price, a loud one on plumbing

No bundled Copilot SKU changed price in August. What changed was the name on the app, the way agents get billed, and who can reach into your tenant. From 18 August the "Microsoft 365 Copilot app" became simply the Microsoft Copilot app, with broad Windows and Mac deployment in mid-September; the URL moves from m365.cloud.microsoft to copilot.cloud.microsoft with an automatic redirect, and the Application ID is unchanged so existing policies keep working. Check your web filter allows *.cloud.microsoft before mid-September.

Three commercial items are worth money: Copilot in 30 is transactable through CSP from 1 August through 31 December 2026 (25 seats, 30 days, free, organisations under 300 employees — mind the two different 300s, since 300 seats is the separate cap on the bundled SKUs); 50% off the Purview Suite was extended on 6 August through 31 December 2026 for Business Premium licensed alongside Copilot; and cross-tenant agent management entered public preview on 10 August, letting a managing partner inventory, install and block agents across every customer tenant from one screen. Ask your IT provider whether they have switched that on for you.

Mark the availability carefully. GA in August: Power BI reasoning in Copilot Chat and Cowork, Cowork effort levels and a /cost skill, Outlook natural-language inbox management. Public preview only, with GA targeted December 2026: SharePoint personal skills and OneDrive files-to-finished-work. PowerPoint SmartArt and Planner AI status reports are Frontier-programme early access, well short of GA.

Copilot Credits: the second invoice got a new line item

Copilot Studio now has three harnesses. Standard (rule-based conversational agents) and Copilot Chat (extending Copilot with your own knowledge) are included under fair-use limits. The GitHub Copilot harness — end-to-end multi-step agentic processes — is not included, and consumes Copilot Credits for building and running. Work IQ Tools API calls bill at a flat 0.1 credit each, a rate not previously published.

The trap is the word "included". A maker who opens Copilot Studio, follows Microsoft's own tutorial for the new agentic experience and builds something useful has left the entitlement without being told — and the bill lands on an Azure subscription, not the Microsoft 365 invoice. Set tenant, environment and per-user caps before anyone builds.

Business Central's published consumption rates were unchanged in August: a sales order email runs 14 credits, 19 with a priced attachment; an AP invoice with three lines runs 65; an expense receipt runs a flat 50 regardless of line count, and that agent is still preview only. Worked through at $0.01 per credit: 100 sales emails a month is about $16.50; five hundred three-line AP invoices about $325.

SharePoint: your lists can now become live dashboards

On 5 August Copilot in SharePoint gained the ability to generate an interactive HTML dashboard directly from a list, an Excel file or a CSV — one that stays connected to the underlying list and refreshes when opened, rather than a static export. Microsoft's own examples are project status, inventory trackers and hiring pipelines. Button web parts can now be wired to a saved Copilot prompt, which is the cheapest way to give non-technical staff a reliable prompt instead of a blank box. All of it still requires a paid Microsoft 365 Copilot licence.

The dated item that will hurt if ignored: Restricted SharePoint Search has been closed to new enrolment since 31 July 2026 and retires completely on 31 January 2027. Allow-lists do not carry over to Restricted Content Discovery. If you used it to fence off HR or Finance from Copilot and you do not migrate, that content becomes discoverable again the day enforcement stops. Separately, an RCD rollout completing in late September removes AI entry points entirely from RCD-governed sites and stops recently-opened files surfacing in Copilot and Microsoft 365 Search there.

Business Central: update 28.4, and the end of release waves

Update 28.4 shipped 3 August 2026. The Payables Agent gained Known Senders (GA) — a policy of always review, never, or only untrusted senders, plus a per-sender Ask / Approve / Reject list. No default is preselected, so it does nothing until an admin chooses one. The Expense Agent remains public preview and is now gated behind an Azure OpenAI privacy notice an admin must accept first.

Good news for tristate firms specifically: the cross-geography data movement setting we flagged last issue does not move your data if you are a United States tenant — Microsoft states processing stays within the United States. The routing applies to EU, UK, Australia, India, APAC, Canada, Japan and Korea tenants.

And a correction worth making loudly: a widely syndicated roundup claimed promotional credit allowances drop on 1 September (Business 3,000 to 1,900, Enterprise 7,000 to 3,900). Those figures trace to GitHub Copilot, a different SKU from Microsoft 365 Copilot, Copilot Studio or Business Central agents. Real if you have developers on GitHub Copilot seats; otherwise it does not touch your bill.

Governance: two real deadlines, one vanished control

The EU AI Act reached you on 2 August 2026. The EU AI Office gained enforcement powers over general-purpose model providers, and Article 50 transparency obligations took effect: disclose that a person is interacting with an AI unless obvious, mark synthetic media machine-readably, label deepfakes. It applies by where the output lands, not where you are incorporated — a 40-person Stamford firm with EU customers, vendors or website visitors hitting an AI chat widget is in scope as a deployer. Fines reach €15M or 3% of turnover; existing systems have until 2 December 2026 for the marking obligations.

Connecticut moves first among the tristate states. Public Act No. 26-15 requires AI disclosure in WARN Act notices from 1 October 2026, and written notice before AI is used in a hiring decision from 1 October 2027. There is no small-employer exemption. New Jersey has no new statute — its Division on Civil Rights guidance dates from January 2025. New York's RAISE Act takes effect 1 January 2027 but targets frontier model developers, not Copilot deployers.

Two licensing gotchas worth pricing now: Entra Agent ID governance is not in Business Premium — conditional access templates, lifecycle workflows and orphaned-identity cleanup all need Entra ID P2 plus a Governance add-on. And Purview retention recommendations for Copilot reach GA in September, but those policies can require pay-as-you-go billing. One thing did get better: Purview DLP for Microsoft 365 Copilot is now generally available commercially, and auto-labelling throughput rose five-fold to 500,000 files a day.

Frontier models: halved again, and a browser disappeared

OpenAI's current API pricing, verified 2 September 2026: GPT-5.6 Sol $2.00 / $10.00, Terra $1.00 / $6.00, Luna $0.10 / $0.60 per million input/output tokens. The pricing page notes that Sol's promotional pricing is available at least through 21 November 2026. Two secondary outlets reported a 22 August cut taking Sol to $4/$20; the vendor's own page shows $2/$10. We print the vendor figure and tell you the other exists.

On 24 August OpenAI added a Premium seat to ChatGPT Business at $125 per user per month, or $100 annually, with five times the usage of a Standard seat and no five-hour cap. Standard stays at $25/$20 and the two-seat minimum holds. The standalone Atlas browser shut down on 9 August and nothing migrated automatically. Retirement dates to diarise: whisper-1 and the gpt-4o-transcribe family were deprecated 26 August with a hard shutdown 26 February 2027; gpt-5, gpt-5-mini, gpt-5-nano and o3 shut down 11 December 2026.

On the Anthropic side, the headline is a cancellation. The docs now read: "The $2/$10 per million input/output token pricing for Claude Sonnet 5 … is now the standard price. The previously scheduled increase to $3/$15 … on September 1, 2026 will not occur." Fable 5.1 and Mythos 5.1 shipped 1 September with base pricing unchanged at $10/$50 but the cache read price cut 75%, from $1.00 to $0.25 per million. Claude Team is $25/$20 standard and $125/$100 premium for 2 to 150 seats — a two-seat minimum, not the twenty some secondary sources still quote; Enterprise floors at 20 seats. OpenAI's new Business Premium seat matches Team Premium to the dollar, so seat price no longer separates the two vendors.

Read the exclusions, not the trust page

Both vendors publish a compliance page designed to reassure. The contract is the exclusion list, and in both cases it excludes precisely the features that make agents useful.

Anthropic: excluded from BAA coverage even on a HIPAA-enabled Enterprise account are Claude Console, Claude Cowork, betas including Claude in Office and Design, MCP connectors, Enterprise Search, Claude in Chrome, and on the API the Batch, Files and Skills APIs, Code Execution, Computer Use and Web Fetch. Zero Data Retention does not cover consumer plans, and web search is excluded even under a BAA.

OpenAI: the Trust Portal makes no mention of HIPAA or BAAs at all. The enterprise privacy page says BAAs are available but scopes that to the API Platform, not ChatGPT Business or Enterprise. Endpoints that may still store application state with ZDR on include /v1/conversations, /v1/assistants, /v1/threads, /v1/vector_stores, /v1/files and /v1/fine_tuning/jobs; /v1/videos is blocked for ZDR entirely.

The pattern, stated plainly: connectors, file stores, web fetch and code execution are what make an agent useful — and on both platforms they are the first things carved out of the strictest data guarantees. If you handle PHI or restricted client data, plan for the duller architecture up front, not in an audit.

Open models: Meta blinked, and the licence got simpler

August was the busiest month of the year for open weights. On 5 August Meta shipped Muse Spark 1.2 as a closed model; five days later it released the same-generation Muse Glimmer 30B as open weights under plain Apache 2.0 — the monthly-active-user threshold gone, the acceptable-use annex gone, the redistribution conditions gone. Every Llama licence caveat we have warned about no longer applies to Meta's current open release.

Also landing: Nemotron 3.5 Lightning (NVIDIA, 11 Aug, 30B/3B active, OpenMDW-1.1), DeepSeek-V4-Pro-0813 (MIT), Qwen3.8-27B (14 Aug, Apache 2.0) and Qwen3.8-2.4T-A95B (custom licence), Granite 4.2 in 3B/8B/30B (IBM, 25 Aug, Apache 2.0), and GLM-5.3-Flash (MIT) and GLM-5.3 (bespoke licence) from Z.ai. On the Artificial Analysis Intelligence Index snapshot dated 1 September, the best closed model scored 65.7 against the best open model at 59.5 — a 6.2-point gap. The UK AI Security Institute separately puts open weights four to seven months behind on cyber tasks, narrowed from six to ten months in early 2025. Coding remains where closed models stay well ahead; extraction, classification and RAG grounding show no real gap.

On one graphics card: Muse Glimmer 30B at Q4_K_XL is about 19GB and the strongest new 24GB pick for agentic tool-calling; Qwen3.8-27B at Q4_K_M is about 17.1GB and better at coding. What will not fit, despite the name: GLM-5.3-Flash carries native FP8 weights of 306–328GB. Rent that one.

The licence traps, briefly. Apache 2.0 or MIT with no triggers: Muse Glimmer, Qwen3.8-27B, Granite 4.2, GLM-5.3-Flash, DeepSeek V4. Qwen3.8-Max requires displaying the model name above 100M MAU or $20M monthly revenue and a separate Qwen licence if model-as-a-service revenue tops $50M over 12 months. GLM-5.3 (753B) requires a Z.ai security review above $10B MaaS revenue. Kimi K3 requires displaying "Kimi K3" above the same thresholds plus a separate agreement above $20M a year from paid inference hosting. The distinction that gets almost every SMB off the hook: those gates attach to running a model-as-a-service business — weights run in-house for your own staff fall outside the definition in all three. And the family name tells you nothing: read the licence from the repository, at the exact size you will deploy.

Automation of the month: the daily collections desk

A 41-person commercial mechanical contractor in Bergen County and lower Westchester, about $18M a year, was carrying $412,000 past ninety days — not because customers would not pay, but because nobody could tell which calls mattered. The controller's weekly ritual was exporting the QuickBooks aging report, cross-checking completion in ServiceTitan, searching Outlook for who had promised what, and dialling from the top of a list sorted by amount. Three failure modes repeated: invoices chased on jobs not signed complete, payment reminders sent to customers who had already disputed a change order, and failed ACH payments that looked identical on the aging report to customers simply ignoring them.

Five sources are pulled on a weekday schedule — QuickBooks Online, ServiceTitan, Stripe, HubSpot and Microsoft 365 via Graph — into a staging schema. Postgres then does the joins and every number: invoice to job to payment to owner, days past due, aging bucket, ACH return status, weighted exposure. No model runs in that stage. About forty surviving rows go to a single model call that classifies each email thread, ranks, and writes one line of reasoning, producing twelve tiered actions in Teams and the controller's inbox by 06:40.

The design decision that made it trustworthy: the model never sees a ledger and never produces a dollar figure. Every amount and date passes through verbatim from the reconciliation table. The model can be wrong about which call matters most; it cannot be wrong about how much is on the invoice, because it was never asked.

Results over the two quarters after go-live: DSO fell from 61 days to 43, AR past 90 days from $412,000 to $148,000, and weekly hours on collections from 11.5 to 2.25 — roughly $890,000 of cash pulled forward. It took 52 engineering hours over three and a half weeks and costs $39.49 a month to run, of which model spend is $1.49. Client anonymised; these are one company's figures, not a benchmark.

Six things worth an hour each

1. Re-price every automation you costed before August. Frontier prices halved again and the Sonnet 5 increase was cancelled. If you rejected an automation on cost this year, the arithmetic that killed it no longer exists.
2. Set Copilot Credit caps before anyone opens the new Copilot Studio. The GitHub Copilot harness bills outside your Microsoft 365 Copilot licence from 1 September.
3. Migrate off Restricted SharePoint Search. New enrolment closed 31 July; the feature retires 31 January 2027 and allow-lists do not carry over.
4. Answer one question about the EU AI Act: does our AI output reach anyone in the EU? If yes, you have been in scope since 2 August. If no, write down why and move on.
5. Connecticut employers: put AI disclosure in your WARN template now. Mandatory from 1 October 2026.
6. Grep your codebase for whisper-1. Transcription models were deprecated 26 August with a hard shutdown 26 February 2027.

Everything above is dated and attributed, and features are marked GA only where the vendor says so. Three items are reported with their uncertainty attached rather than left out: OpenAI's exact August price chronology for Sol, the precise GA date for Purview DLP for Microsoft 365 Copilot, and whether a Business Premium tenant without a Copilot licence can buy SharePoint Advanced Management standalone. The designed 15-page PDF edition, with the full tables and charts, is free below.