AIStall
EditionBy The AI Stall editorsOctober 10, 2026

Three Things in AI: Who Pays When the Agent Does the Work

Three Things in AI: Who Pays When the Agent Does the Work

Three things, all published on October 8, 2026 and read from the vendors' own pages on October 10. They share one theme: when an agent does work on your behalf, somebody is on the hook, either for the bill or for the decision. GitHub decided who pays for a code review. Microsoft set a date for when a sales agent starts costing credits. Anthropic wrote down when a person must stay in the loop. Same three slots as always: what shipped, what it costs, who it changes the week for. ☕

What this note covers, in order.
What this note covers, in order.

One: GitHub lets the organisation pay for Copilot code review

What shipped. On October 8, 2026 GitHub's changelog announced [new organisation billing options and controls for Copilot code review](https://github.blog/changelog/2026-10-08-copilot-code-review-new-organization-billing-options-and-controls). Organisation owners get a setting under Copilot policies that decides whose account a review is charged to.

The default, called Member, bills each review to the requesting member's own Copilot entitlement, and the review fails if that quota is used up. The new option, Organization, bills the review to the organisation that owns the repository, so members' quotas are untouched.

That second option requires paid AI Credits usage to be enabled for the organisation, and a budget can be set.

A second control lets owners and repository admins restrict review requests to authorised users, which stops someone using a personal Copilot licence from outside the organisation to trigger reviews. If an owner turns it on at organisation level, repository admins cannot turn it off.

What it costs. The changelog does not print a per-review price or a premium-request figure; it points to the billing documentation. The honest line is: the mechanism is published, the number is not on this page.

Who it changes the week for: any team where one developer's reviews kept failing at the end of the month because their quota ran out. Switch the setting to Organization, set a budget, and the failure mode moves from "review silently did not run" to "budget alert".

That is a better failure.

Two: Microsoft's Sales Development Agent starts costing credits on November 16

What shipped. On October 8, 2026 Microsoft published [a post on bringing CRM into Teams and agents into sales and service processes](https://www.microsoft.com/en-us/copilot/blog/2026/10/08/bringing-crm-into-the-flow-of-work-and-agents-into-business-process-for-sales-and-service-teams/). The list is long: Dynamics 365 records inside Teams channels, 30 pre-built sales and service skills for Copilot, a Service Agent in Microsoft 365 and a Sales Development Agent.

The Sales Development Agent is the one a small company might actually switch on. Per the post it researches prospects, writes personalised outreach, manages follow-ups and hands qualified leads to a seller. It is managed in Teams, sends email from its own Exchange account, and connects to major CRMs, not only Dynamics.

Availability is a patchwork, and the post says so: skills in Cowork are generally available, skills in Autopilot are private preview, skills in Code are Frontier Program only, the Service Agent is private preview, and the Sales Development Agent is public preview. Teams channel integration with Dynamics is public preview.

What it costs. No dollar figure appears in the post. The one hard number is a date: the Sales Development Agent starts consuming Copilot credits on November 16, 2026.

Until then the public preview is, in effect, free to try; after that it draws from the same credit pool as everything else agentic in Copilot, at a rate this post does not state.

Who it changes the week for. A five-person sales team on Microsoft 365 that wants an inbox-working assistant should trial it now and write down, before November 16, how many credits a week of outreach actually consumed.

The same caution applies to Autopilot, which Microsoft's September 25 announcement bills by usage rather than by seat: the risk of an unattended agent is not the feature, it is the meter.

Three: Anthropic's usage policy now names the human who must stay in the loop

What shipped. On October 8, 2026 Anthropic published [its 2026 Usage Policy update](https://www.anthropic.com/news/2026-usage-policy-update), effective November 12, 2026. Most of it concerns weapons, surveillance and elections. The section that reaches a small business is the one on high-risk uses.

Per the post, uses of Claude that affect someone's health, legal rights, finances, livelihood or access to essential services require a qualified human who can review and change the model's recommendation, and the affected person must be told that AI was involved.

A new section also consolidates rules against fake accounts, fabricated news sites and artificial activity, and it applies to commercial campaigns as well as political ones.

What it costs. Nothing in money. It costs a process: if you use Claude, directly or through a product built on it, to screen tenants, score loan applicants, triage patient messages or shortlist job candidates, you need a named person with authority to overrule it, and a sentence somewhere telling the applicant that AI was used.

Who it changes the week for. Any small firm in lending, lettings, recruitment, insurance, clinics or benefits advice. Also anyone tempted to generate a batch of reviews or testimonials: the commercial-deception clause now covers that explicitly.

The post does not say how these rules apply to a particular setup; the full policy and the Supported Regions page are where to check, and the ownership clause there matters if your company is majority-owned from an unsupported region.

What it costs, in one place

Not in today's three, and why

GitHub also shipped, on October 8, a setting that makes draft pull requests count toward a repository's pull-request limits; it matters to open-source maintainers drowning in low-quality submissions and to almost nobody else reading this.

Anthropic's Cyber Mission, also October 8, offers free periodic scans of critical open-source projects and free Claude Max subscriptions to approved maintainers; good news, but the door is for maintainers, not small companies.

Yesterday's edition covered Haiku 5.5 pricing and the free Copilot sandbox. The next one is about tools appearing inside other tools. The rules for what gets in are still the ones in issue zero.

Bottom line

If your developers use Copilot code review: open organisation settings, choose Organization billing, set a budget, and tick the authorised-users box. If your sales team is on Microsoft 365: trial the Sales Development Agent before November 16 and record what it consumes.

If any part of your business uses Claude to decide something about a person's money, health, home or job: name the human reviewer this week, because November 12 is five weeks away.

None of the three cost money today. All three cost a decision about who is responsible when the agent is working and nobody is watching. Tomorrow, three more.

Sources