What Changed in AI This Week? OpenAI Ads, Astra for Law, Salesforce AIforce, and Claude Safeguards
What Changed in AI This Week: OpenAI Ads, Astra for Law, Salesforce AIforce and Claude Safeguards
The biggest AI developments for businesses this week, from conversational advertising and specialist legal AI to CRM agents, measurement and governance.
AI is moving deeper into advertising, analytics, specialist professional work and core business systems. For businesses, the focus now needs to shift towards trusted data, clear permissions, measurable outcomes and governance around what AI is allowed to do.
Is this the week AI turns sentient and kills us all? Well, it’s made headlines all week for fears of exactly that happening within a decade, which is far from a cheerful way to start this instalment.
So, let’s pivot away from that and look at how it’s moving into the places where businesses sell, measure, govern and carry out specialist work.
This week has seen OpenAI turn ChatGPT into a more serious advertising channel, give administrators better ways to connect AI spend with business outcomes and launch a legal-specific version of its latest model. Salesforce used Dreamforce to make CRM data and actions available through other AI interfaces. Anthropic introduced verified access for sensitive life sciences work. Microsoft’s main Copilot announcement was aimed at US government teams, but its bundled approach to AI and governance is still worth noticing.
The common thread is fairly straightforward, for once: the market is moving beyond asking whether AI can produce a useful answer. The next question is going to be whether it can work inside a real process, use the right data, stay within permissions and produce enough value to justify the cost. Those are much healthier questions for businesses to be asking.
Headlines at a glance
OpenAI Ads
Sponsored Agents turn selected US ads into branded conversations, while HubSpot and Shopify integrations bring campaign management closer to existing sales and ecommerce workflows.
AI ROI
New OpenAI admin analytics connect usage, spend and task data, giving leaders a better starting point for measuring whether AI is improving real outcomes.
Astra for Law
OpenAI has configured GPT-6 Astra for legal research and professional workflows, with specialist controls, a US legal index and early access for selected firms.
Salesforce AIforce
Salesforce wants its CRM data, workflows and permissions to follow staff into Claude, Slack and other AI interfaces instead of keeping work inside a fixed application screen.
Claude safeguards
Anthropic’s new verification programme gives approved life-sciences teams broader model access in exchange for vetting, scoped use and monitoring.
Microsoft G7
Microsoft bundled Copilot, agents and governance for US government customers, signalling how regulated AI may be packaged.
OpenAI turns ChatGPT ads into conversations
OpenAI has announced a new set of ChatGPT Ads features, led by Sponsored Agents. When someone clicks a participating ad, they can choose to start a clearly labelled conversation with an agent representing that business. The test is currently limited to selected advertisers in the United States.
The advertiser side is becoming more automated, too. Businesses can use natural-language prompts in ChatGPT to create, update and analyse campaigns through an Ads Manager plugin. OpenAI is also adding suggested copy and imagery, optional text customisation and automatic translation. Advertisers still review and approve what goes live.
The integrations may matter most to smaller businesses, as HubSpot users can connect a ChatGPT Ads account, manage campaigns, track performance and follow up leads using CRM context. A Shopify app is available now to US merchants, with international availability planned from the 23rd of September in markets where ChatGPT Ads operates.
This is not simply another advertising dashboard, but more so a bet that product discovery will become conversational: a potential customer sees an offer, asks detailed questions and moves towards a purchase without starting a new search. For businesses, that could shorten the gap between interest and qualification. It could also make accuracy more important. A branded agent that gives a confident but incorrect answer about price, availability or suitability creates a customer-service problem before the sale has even happened.
Treat conversational advertising as a sales and service channel, not just a media buy. Keep product data current, define which claims the agent may make, preserve a clear route to a person and measure lead quality rather than celebrating clicks alone.
OpenAI gives businesses better tools to prove AI value
OpenAI also released new analytics for ChatGPT Work and Codex. The Admin Console now brings together adoption, credits and token use with task-level insights. Administrators can see the kinds of work teams are using AI for, which models and tools support those tasks and where usage is concentrated. Codex analytics can also show contributions to merged commits and lines of code alongside review activity.
That sounds like administration, because.. well, it is. But it’s also one of the more commercially useful releases of the week. Many AI programmes still measure success by licences assigned, prompts sent or hours supposedly saved. None of those, on its own, tells you whether the business is better off.
OpenAI’s own guidance is more sensible: pick an outcome, establish the current baseline, compare the changed process over a defined period and include review and correction time. It also suggests combining the Admin API with business data, such as comparing support-agent credit use with ticket-resolution time. The worked sales example is explicitly hypothetical, so it should be treated as a method rather than a promised return.
A separate OpenAI study of workplace use adds useful context. Among roughly 6,200 workers observed consistently from April to July, previously used cross-occupation tasks rose from 13.1% to 25.9% of occupation-specific AI activity. Marketing and customer-facing tasks were among those people most often returned to. The research doesn’t prove productivity gains, but it does suggest AI is quietly broadening roles before job titles change.
Choose two or three repeatable workflows and measure them properly. Track cost, completion time, quality, rework and the downstream result. If the only dashboard metric is ‘active users’, you know adoption is happening, but not whether it is useful.
Astra for Law shows the move towards specialist AI
OpenAI’s Astra for Law combines GPT-6 Astra with legal instructions, a search index and controls designed for professional legal work. The index covers US case law, statutes, regulations, court rules and administrative decisions. The product will initially be available to selected law firms through a Trusted Access programme in ChatGPT and Codex, with API access coming later.
OpenAI says Astra for Law passed the overall correctness check on 54.0% of 200 questions from a private validation set of the Vals AI Legal Research Bench, compared with 38.7% for GPT-6 Astra using web search alone, which is a meaningful improvement on OpenAI’s test, but it isn’t a reason to remove professional review; a 54% pass rate rather makes the opposite point.
The controls and ecosystem are as important as the model. Eligible firms can access Zero Data Retention via the API (use with ChatGPT Enterprise is excluded from human review by default), and 26 partner-built plugins connect to tools such as Relativity and Clio. OpenAI also made ChatGPT for Word generally available for legal users.
The broader business signal is that one general assistant will not be enough for every high-stakes workflow. Legal, finance, healthcare and other regulated work need specialist sources, permissions, audit trails and review processes. The frontier model is becoming one component in a controlled system rather than the whole product.
Don’t confuse sector branding with automatic suitability. Before using specialist AI for legal work, ask which jurisdictions and sources it covers, how current they are, where data is retained, how citations are checked and who remains accountable for the final advice.
Salesforce AIforce takes CRM beyond the Salesforce screen
At Dreamforce, Salesforce unveiled AIforce, an interface layer intended to make Salesforce data, workflows, business logic and governed actions available wherever people and agents work. The launch includes Claudeforce, Slackforce and Agentforce Coworker.
The practical concept for this one is that a user shouldn’t have to open a CRM dashboard and navigate records to get something done. They could ask Claude or Slackbot which accounts are going quiet, see the support history behind the change, create a follow-up task and draft a win-back email. Salesforce says requests follow existing permissions and business rules, and that the model provider does not retain business data.
Salesforce in Claude includes 37 prebuilt sales skills and is now available to all customers in beta. The company is also offering more than 40 skills through a Salesforce Development plugin for Claude Code. This is a substantial ecosystem move: Salesforce is accepting that the interface where work begins may belong to another vendor, while trying to remain the governed system of record underneath it.
For growing businesses, that could reduce the gap between owning a CRM and actually using it well. It could also expose poor data hygiene very quickly. An agent grounded in duplicate accounts, missing ownership and stale opportunity stages will automate the confusion with impressive speed.
Prepare the system of record before adding a conversational layer. Clean key records, confirm ownership, review permissions and decide which actions should require approval. The quality of the agent will depend heavily on the quality and governance of the CRM beneath it.
Anthropic introduces verified access for life sciences
Anthropic has launched its Life Sciences Verification Program in beta for teams and institutions. Approved organisations can use Mythos, Opus and Sonnet models with safeguards adjusted for legitimate biology work that may be blocked in the general product, including drug discovery, clinical development, manufacturing, quality assurance and regulatory work.
The interesting part is the access model. Organisations are vetted for research credentials, security standards and ethical oversight. Standard Use covers broad day-to-day work; High-risk Use is tied to a specific project, receives additional scrutiny and must be renewed every six months. Cyber safeguards remain in place.
Anthropic is shifting some enforcement from immediate blocking to offline monitoring across patterns of use. That reduces disruption to legitimate research, but it comes with a trade-off: data associated with flagged activity is retained for 30 days for review. Anthropic says that data is compartmentalised and not used for model training.
The programme is available through Anthropic’s first-party API console and Claude Enterprise and Team plans, but not individual plans or third-party platforms. It isn’t available to BAA-enabled organisations during the beta, so customers handling protected health information need to pay close attention to that limitation.
This is a useful model for sensitive AI access: verify the organisation, scope the use case, monitor behaviour and make responsibilities explicit. It is also a reminder to read the data-retention and compliance detail before a pilot begins, particularly where health data is involved.
Microsoft bundles Copilot and governance for government
Microsoft’s main Copilot announcement this week was Microsoft 365 G7, a US government offering that combines Microsoft 365, Copilot, agents, identity, security, compliance and governance for Government Community Cloud environments. It becomes available for GCC customers to purchase on the 1st of October, with capabilities arriving in phases as authorisations are completed.
This is not an SMB product launch, and it shouldn’t be presented as one. The reason it matters more widely is the packaging, as Agent 365 is positioned as a control plane to discover, register, manage, block or remove agents, while Copilot and Work IQ provide the user experience and organisational context. Microsoft is selling the productivity layer and the control layer together.
That reflects the direction Microsoft has been taking across Copilot: as agents become easier to create, organisations need a reliable inventory of what exists, who owns it, what it can access and how it is behaving. Regulated government environments make that need obvious, but the same questions apply to a 100-person business with staff building agents in several tools.
Keep an agent register even if your organisation is nowhere near government scale. Record the owner, purpose, data access, actions, approval points, supplier and review date. Governance is much easier when it starts with a simple inventory rather than an incident.
The practical takeaway for businesses
This week’s releases point in the same direction. AI is moving into specialist work and everyday systems, while vendors compete to own the interface, the data connection and the control plane. That may make the tools more useful, but it also makes casual adoption harder to defend.
For most SMBs and growing businesses, don’t go chasing every new product. Pick a small number of workflows where the outcome matters, connect AI to trusted data, define where people must review or approve an action and measure the result against a baseline. Then keep an eye on the less exciting details: permissions, data retention, variable spend, system ownership and what happens when the answer is wrong.
This week's AI news FAQs
What are OpenAI Sponsored Agents?
Sponsored Agents are part of OpenAI’s ChatGPT Ads tests. They allow someone interacting with a participating advertisement to start a clearly labelled conversation with an agent representing that advertiser.
How can businesses measure AI ROI?
A useful approach is to start with a defined business outcome and baseline, then compare cost, completion time, quality, rework and the downstream result after AI is introduced rather than relying only on adoption metrics.
What is Astra for Law?
Astra for Law is a specialist version of GPT-6 Astra configured for legal research and professional workflows, supported by legal sources, controls and integrations for selected firms.
What is Salesforce AIforce?
AIforce is Salesforce’s interface layer for making CRM data, workflows, permissions and governed actions available through AI environments such as Claude and Slack rather than limiting work to the traditional Salesforce interface.
Why does AI governance matter more as agents become common?
Agents can access data, use tools and take actions across business systems. Organisations therefore need visibility over what agents exist, who owns them, what information they can access and where human approval is required.
Build AI around real business outcomes
We can help you identify useful workflows, connect the right data, establish governance and turn AI adoption into something your business can actually measure.
Talk to our team
Bring us the workflow, AI idea or governance question you’re working through and we’ll help you make sense of the next step.
AI for business
See how Fifosys approaches AI adoption, automation, governance and practical business use cases.