This Month in AI — August 2026

The 60-second version

July was a pricing month, not a demo month. The models got cheaper, the Microsoft bundles got restructured, and a security control quietly stopped enforcing. Five things changed — two of them are already costing you money:

1. Copilot got about 40% cheaper for small businesses. Microsoft made the bundled "with Copilot" SKUs permanent on 1 July. A 50-person firm on Business Premium now pays about $19,200/yr for full coverage instead of roughly $31,000 under the old add-on math — capped at 300 seats.
2. Per-user licensing is now only half the bill. Agents meter separately in Copilot Credits at $0.01 each. Copilot Cowork went GA on 16 June and is off by default — set tenant, group and per-user caps the moment you switch on usage-based billing.
3. Agents can finally read your structured data. SharePoint list support in Agent Builder hit GA in June — job trackers, equipment registers and onboarding lists are now answerable in plain English.
4. The cheap tier of AI collapsed in price. OpenAI cut GPT-5.6 Luna by 80% on 30 July to $0.20/$1.20 per million tokens. High-volume classification, extraction and triage are now close to free.
5. One security control turned itself off. On 1 July, Copilot Studio and Foundry agent protection moved to Agent 365 licensing with no grace period — tenants that were blocking agent activity stopped blocking that day. This is the single most urgent item in this brief.

Microsoft: the licensing reset

On 1 July, Copilot stopped being an expensive add-on and became a bundled SKU. If you priced Copilot in 2025 and walked away, the math you rejected no longer exists.

Plan	List / user / mo	What you get
Business Premium with Copilot	$32.00	Full Copilot plus the security stack; the only SMB base that qualifies for Agent 365
Business Standard with Copilot	$23.50	Full Copilot in the desktop apps; no Intune, no Defender for Business
Business Basic + Copilot Business	$21.00 promo	Cheapest legitimate route; web apps only
Copilot Business add-on	$18.00 promo	Bolt onto an existing tenant; license the 12 heavy users instead of all 80

The rule of thumb: the bundle wins when most of your staff will use Copilot; the add-on wins under roughly 55% adoption. And note the wrinkle — Business Premium is separately required if you want Agent 365 governance later. A commercial price increase hit Business Basic, Standard, Apps for Business and most E-plans the same day; Premium was not on the list. Do this first: Microsoft's "Copilot in 30" gives organizations under 300 employees a free 25-seat, 30-day Copilot Business trial through CSP partners from 1 August — run it on your single worst bottleneck and measure that one thing. (Sources: Microsoft Partner Center announcements, July 2026; microsoft.com Business plans, retrieved 3 Aug 2026.)

Copilot Credits: the second invoice

Agent billing unified into a single currency worth one cent, shared across Copilot Cowork, Work IQ, Copilot Studio, Business Central agents and Power Platform agents. Your Copilot spend now arrives on two invoices — per-user licensing on the Microsoft 365 bill, agent consumption on an Azure subscription somebody else probably watches. What one run costs: a sales-order email end-to-end ~$0.14; an expense receipt ~$0.50; a 3-line AP invoice ~$0.65; a Cowork light task up to ~$2.00; deep analysis ~$15. Agents that stay inside Microsoft 365 and ground on your own data are effectively free under a Copilot license; agents that call external tools or run autonomously are not. For a 60-person firm where 8 people genuinely need full Copilot, the cost-control shape is 8 licensed seats plus pay-as-you-go agents at $0.01 per billable message for everyone else.

SharePoint: agents can read your lists now

SharePoint list grounding in Agent Builder went GA in June (up to 20,000 items per source) — "which jobs are overdue, and who owns them?" becomes a chat question instead of a Power BI project. Copilot in SharePoint now builds documents, dashboards and reusable site automations from site content, and since 1 July Copilot uses your library metadata to sharpen answers — a direct payoff for firms that invested in structure. The warning that matters: Copilot does not create oversharing; it exposes the oversharing you already have. Run a content assessment before licensing, turn on auditing first, and use Restricted Content Discovery as a scalpel on HR and Finance — five to ten sites, not a tenant-wide switch.

Business Central: the ERP agents grew up

Version 28.3 landed 1 July. The Sales Order agent is GA with multiple instances, relevance filtering and a 50-page PDF limit; the Payables agent is GA with a 50-invoice trial mode that needs no mailbox setup; the Expense agent stays preview-only — pilot it, don't rely on it. Worked example at published credit rates: 100 sales emails ≈ $16.50/month, 500 AP invoices ≈ $325/month — compare that against the loaded cost of the clerk hours displaced, not against zero. Two fine-print traps: agent requests may now be processed in other Azure geographies (review the residency setting if client contracts care), and unallocated credits silently draw from your default Power Platform environment.

Governance: five traps in the fine print

1. Your agent blocking rules stopped enforcing on 1 July. Tenants configured to Block on old Agent 365 rules stopped blocking that day; rules must be redefined under Settings › Security for AI › Policies. Check it today.
2. Agent 365 costs $15/user and quietly requires Business Premium (or E5/A5) as the base — which changes the SKU decision above.
3. SharePoint Advanced Management eligibility is genuinely unclear for SMBs — get entitlement confirmed in writing before designing governance around it.
4. The EXTRACT usage right will make Copilot look broken — encrypted content labeled years ago without it silently fails on exactly your most important documents.
5. One thing got quietly better: Copilot-generated files now inherit the sensitivity label of their source, and Purview's DSPM view of every active AI app went GA.

Frontier models: the price floor collapsed

OpenAI's GPT-5.6 family shipped 9 July; three weeks later the cheap tier was cut 80% — Luna now runs $0.20/$1.20 per million tokens, Terra $2.00/$12.00, Sol $5.00/$30.00, all with a ~1M-token context. Hard monthly spend caps (added 22 July) are the single most practical control against a runaway agent invoice. Anthropic shipped Claude Opus 5 on 24 July at unchanged pricing with the freshest knowledge cutoff here (May 2026), while Sonnet 5's introductory $2.00/$10.00 rate ends 31 August — a real budget line. The honest chooser: OpenAI for high-volume cheap work and small-team governance (ChatGPT Business has a 2-seat minimum; Claude Enterprise floors at 20 seats); Anthropic for document-heavy retrieval and long agentic runs (flat pricing to a full 1M context, 950+ MCP connectors). Both vendors' compliance pages are marketing — the exclusion lists are the contract, and neither's "HIPAA-ready" survives contact with its own feature list. If you handle PHI, read the exclusions first.

Open models: further behind, carrying more traffic

Two facts that sound contradictory and are not: open weights are now about four months behind the closed frontier (Epoch AI), and they carry a majority of tokens on OpenRouter (Mozilla). Extraction, classification, summarization and RAG grounding show no meaningful gap; hard reasoning and computer use remain decisively closed-model territory. On one graphics card, the picks that change your economics: Gemma 4 26B-A4B (best single-GPU model), Granite 4.1 8B (classification and RAG), OpenAI's Privacy Filter (PII redaction at a 96% F1 score, on a CPU box), and Granite Speech (transcription that never leaves the building). At 100M tokens/month of extraction, open-hosted runs roughly $20 against $2,500 frontier — the right architecture routes the hard 10% to a frontier model and everything else to open. Watch the license traps: Llama's attribution rule, revenue-tiered Kimi, and CC-BY-NC embedding models that quietly poison commercial RAG stacks.

Automation of the month: the weekly marketing clinic

A nine-location retail and services group (~$14M revenue, two-person marketing team) was losing most of every Monday reconciling five dashboards — Search Console, GA4, Meta Pixel, Conversions API and Semrush — into observations instead of decisions. We built a scheduled pipeline that pulls all five into one table, does the week-over-week math deterministically, and uses one model call to rank ten actions into three tiers with the evidence attached. Monday reporting fell from 6.5 hours to 24 minutes; two quarters in, organic non-brand clicks are +34%, blended acquisition cost −19%, and 81% of Tier-1 items got done. It runs for under $15 a month in model spend. The design decision that made it work: the model never sees raw analytics — the numbers are computed, not generated; the model's only job is ranking and explanation. (Client anonymized; figures from that engagement, not a benchmark.)

Everything above is dated and attributed; features are marked GA only when the vendor says so. The designed 15-page PDF edition, with the full tables and charts, is free below.

Corrected 8 October 2026: Business Standard with Copilot is a permanent $23.50, not a $22.00 promotion; the $21.00 Business Basic route is promotional; SharePoint list support reached GA in June, not on 29 July; and the bundle saving is about 40%, not half. Smaller fixes cover the Expense agent, the E-plan price rise, Mozilla's token figure and the Privacy Filter score. The PDF edition still has the original text.