Everything-PR followed 30 AI product announcements made between April 1 and June 30, 2025, and checked dated records for 90 days after each one. A separate dated record supported the tracked access promise in 16 of 30 cases. One release was late and one target was later revised. The other 12 had limited evidence, an unverified start date, or no deadline that could be assessed within 90 days.
Updated November 21, 2026. Based on 30 announcements dated April 1 to June 30, 2025, each followed for 90 days.
Download the 30-case evidence workbook (XLSX)
Which AI launches arrived late or changed within 90 days?
Five of 30 cases (17%) had a documented change in release timing, eligibility, pricing structure or access conditions inside the 90-day window. Gemini 2.5 Flash was the one late release, and four others widened or reshaped access.
- OpenAI confirmed Codex access for ChatGPT Plus users in a June 3, 2025 update. The May 16 launch had started with Pro, Enterprise and Team plans and described Plus and Edu as coming soon.
- GitHub opened the Copilot coding agent preview to Copilot Business customers in a June 19, 2025 changelog entry, after the May 19 preview for Pro+ and Enterprise.
- Google's Jules left its free public beta on August 6, 2025 with a free tier and paid plan limits, as its changelog records. That was day 78 of 90.
- Cursor widened Background Agent access in version 1.0, with Privacy Mode disabled as the condition for immediate access, per its 1.0 changelog.
- Google's Gemini 2.5 Flash reached production on June 17, 2025, seven days after the early June window the study coded as June 1 to 10.
Two of the four access changes, Codex and Jules, were signaled at launch, so they widened access without breaking a promise. Salesforce is a sixth case if later evidence is counted. Its Agent Health Monitoring was announced on June 23, 2025 in pilot or beta with general availability planned for August. The reviewed records showed no general availability by day 90 on September 21, and a November 20, 2025 Salesforce update moved general availability to spring 2026. That update is shown as a revised target because it was published 150 days after the announcement.
What counts as kept when an AI vendor promises a rollout start?
A promise to start a rollout is kept when access begins for any eligible customers in the stated period, not when every customer has it. The study scored each case against the wording the vendor used.
Google's May 6, 2025 AI Max for Search campaigns announcement said a global beta rollout would start later in May. The Google Ads developer blog gave a May 27 start date, and Search Engine Land reported the beta in advertiser accounts on May 29. Google confirmed on September 10, 2025 that all advertisers globally could use it. The May records support the promised start. The September record documents the wider milestone, 127 days in.
HubSpot's Breeze Customer Agent expansion worked differently. The announcement named a date and eligible plans, and a June webinar record corroborated Professional and Enterprise access and dated credit use to June 2, 2025. A separate August 4 transition for legacy Service Hub customers was announced but not scored.
Were clearer announcements easier to verify?
Announcements rated Clear were corroborated 13 of 19 times (68%), while Partial ones were corroborated 3 of 10 times (30%). That measures how often the study could corroborate the promise, and clearer wording makes a commitment easier to score. A Partial announcement is often missing the date or access detail that a second record would need to confirm.
| Announcement clarity | Cases | Corroborated | Other outcomes |
|---|
| Clear | 19 | 13 | 3 launch record only, 1 late, 1 available with start unverified, 1 not yet due |
| Partial | 10 | 3 | 2 launch record only, 2 available with start unverified, 1 revised, 1 insufficient evidence, 1 no testable timetable |
| Unclear | 1 | 0 | 1 no testable timetable |
The one late release, Gemini 2.5 Flash, came from a Clear announcement. A clear date made the miss measurable. By sector, corroboration ran 5 of 8 for enterprise AI, 6 of 8 for developer tools, 2 of 7 for cybersecurity and 3 of 7 for marketing software. Those groups are small and uneven, and the cybersecurity and marketing counts mostly reflect thin public records rather than failed launches.
What did AI launch posts leave out about eligibility and regions?
Geography went unstated in 21 of 30 launch records (70%), including cases rated Clear on access stage, timing and buyer group.
The stated conditions often decided whether a team could start. Amazon Q Business named a $3 monthly price, supported browsers and availability tied to Q Business regions. Google Agentspace used an allowlist for Agent Designer. OpenAI's deep research connectors for Enterprise and Edu needed administrator enablement and a supported external service. Cursor's Background Agent needed Privacy Mode disabled for immediate access. A launch can be live and still be out of reach for a given buyer.
Why do some AI launches leave a thin public record?
Five launches had only the launch post saying access was live, and three of those five were Amazon Web Services milestones (the Q Business browser extension, Bedrock Guardrails cross-Region inference and the Guardrails Standard tier). Launch record only means no second dated record was found. It does not mean access failed.
Two Google and CrowdStrike cases show the other kind of gap. Google described Asset Studio as coming soon on May 21, 2025. Trade reporting found the beta in advertiser accounts on August 22, three days after day 90, so first access is unknown. CrowdStrike said Charlotte AI Agentic Response would be generally available in the coming weeks. The records reviewed documented April access for a different feature, Agentic Workflows, and could not date Agentic Response itself.
Three more cases were available by day 90 with the original start unverified: Google's Agentspace Agent Designer, Anthropic's Claude Code and SentinelOne's Purple AI Athena. For Athena, SentinelOne said on April 29, 2025 that Auto Triage had gone generally available that week as the first phase, and a technical post updated May 2 described the wider Athena preview.
What do two real AI launch timelines look like?
Google AI Max shows a rollout that began on schedule and reached global availability later. The study scored the May start, and the September milestone sits outside the 90-day window.
| Date | Days from announcement | Event |
|---|
| May 6, 2025 | 0 | Google says the global beta rollout will start later in May. |
| May 27, 2025 | 21 | Start date specified by the Google Ads Developer Blog. |
| May 29, 2025 | 23 | Trade coverage shows the beta appearing in advertiser accounts. |
| Aug 4, 2025 | 90 | Day 90. Rollout start had been corroborated. |
| Sep 10, 2025 | 127 | Google confirms all advertisers globally can use the beta. |
Salesforce Agent Health Monitoring shows a target revised after day 90. The August general availability goal passed with no general availability found in the reviewed records, and the later update is recorded as a revised target, not as proof of when release happened.
| Date | Days from announcement | Event |
|---|
| Jun 23, 2025 | 0 | Pilot or beta with general availability planned for August. |
| Aug 31, 2025 | 69 | End of the promised GA month. |
| Sep 21, 2025 | 90 | Day 90. GA remained unverified in the reviewed records. |
| Nov 20, 2025 | 150 | Salesforce says GA will come in spring 2026. This is a revised target. |
Three more timelines: Gemini 2.5 Flash, GitHub Copilot coding agent and SentinelOne Purple AI Athena
Gemini 2.5 Flash
| Date | Days from announcement | Event |
|---|
| May 20, 2025 | 0 | Preview open. Production GA promised for early June. |
| Jun 10, 2025 | 21 | End of the study's early-June coding window. |
| Jun 17, 2025 | 28 | Google developer and Vertex AI records date production GA. |
| Aug 18, 2025 | 90 | Day 90. Production availability had been documented since June 17. |
GitHub Copilot coding agent
| Date | Days from announcement | Event |
|---|
| May 19, 2025 | 0 | Public preview announced for Pro+ and Enterprise. |
| Jun 19, 2025 | 31 | Changelog confirms the May preview and opens access to Business. |
| Aug 17, 2025 | 90 | Day 90. Original preview and broader eligibility had been documented. |
SentinelOne Purple AI Athena
| Date | Days from announcement | Event |
|---|
| Apr 29, 2025 | 0 | Vendor announces a phased Athena release and says Auto Triage was made GA that week. |
| May 2, 2025 | 3 | Technical post, updated on this date, describes Athena preview and identifies Auto Triage as its first phase. |
| Jul 28, 2025 | 90 | Day 90. Preview was documented. First enablement date and broader access remained unverified. |
What does the full 30-case data show?
The table lists every case in the study with the milestone tracked, the announcement date, the clarity rating and the timing outcome. Day 90 is the announcement date plus 90 calendar days. Dates are 2025.
| ID | Vendor | Milestone tracked | Announced | Clarity | Timing outcome |
| E01 | Microsoft | Azure AI Foundry Agent Service GA | May 19 | Clear | Corroborated |
| E02 | Salesforce | Agentforce 3 Agent Health Monitoring | Jun 23 | Partial | Revised timetable |
| E03 | OpenAI | Deep research connectors for Enterprise and Edu | Jun 4 | Clear | Corroborated |
| E04 | AWS | Amazon Q Business browser extension subscriber expansion | Apr 3 | Clear | Launch record only |
| E05 | AWS | Amazon Q Business hallucination mitigation | Apr 14 | Clear | Corroborated |
| E06 | Google | Agentspace Agent Designer | Apr 9 | Clear | Available by day 90, start unverified |
| E07 | Notion | AI Meeting Notes | May 13 | Clear | Corroborated |
| E08 | Google | NotebookLM as a Workspace for Education core service | Apr 3 | Clear | Corroborated |
| D01 | OpenAI | Codex cloud research preview | May 16 | Clear | Corroborated |
| D02 | Anthropic | Claude Code GA | May 22 | Partial | Available by day 90, start unverified |
| D03 | GitHub | Copilot coding agent public preview | May 19 | Clear | Corroborated |
| D04 | Google | Jules public beta | May 20 | Clear | Corroborated |
| D05 | Google | Gemini 2.5 Flash updated model GA | May 20 | Clear | Late |
| D06 | Google | Gemini Diffusion experimental demo | May 20 | Partial | Corroborated |
| D07 | Cursor | Background Agent early preview | May 15 | Partial | Corroborated |
| D08 | OpenAI | GPT-4.1 API model family | Apr 14 | Clear | Corroborated |
| C01 | Microsoft | Entra Agent ID initial inventory preview | May 19 | Clear | Corroborated |
| C02 | Palo Alto Networks | Prisma AIRS AI Runtime Security platform announcement | Apr 28 | Unclear | No testable timetable |
| C03 | SentinelOne | Purple AI Athena rollout | Apr 29 | Partial | Available by day 90, start unverified |
| C04 | CrowdStrike | Charlotte AI Agentic Response | Apr 28 | Partial | Insufficient evidence |
| C05 | AWS | Bedrock Guardrails April safety controls | Apr 8 | Clear | Corroborated |
| C06 | AWS | Bedrock Guardrails cross-Region inference | May 14 | Clear | Launch record only |
| C07 | AWS | Bedrock Guardrails Standard filter tier | Jun 24 | Clear | Launch record only |
| M01 | Adobe | GenStudio Firefly image generation beta | Jun 16 | Partial | Launch record only |
| M02 | Salesforce | Marketing Cloud Next GA | Jun 11 | Partial | Launch record only |
| M03 | HubSpot | Breeze Customer Agent expanded access | May 8 | Clear | Corroborated |
| M04 | Google | AI Max for Search campaigns beta | May 6 | Clear | Corroborated |
| M05 | Google | Google Ads Asset Studio | May 21 | Partial | No testable timetable |
| M06 | Google | Marketing Advisor for Chrome | May 21 | Clear | Not yet due |
| M07 | Adobe | Firefly Image Model 4 Ultra on web | Apr 24 | Partial | Corroborated |
How was the AI launch availability study designed?
The study is a purposive sample of 30 named, buyer-facing access milestones from dated vendor material, with quotas of eight enterprise AI, eight developer tool, seven cybersecurity and seven marketing software cases. It describes this sample only and is not a market estimate.
Large vendors are overrepresented. Google accounts for 8 of the 30 cases and Amazon Web Services for 5, followed by OpenAI with 3 and Microsoft, Salesforce and Adobe with 2 each. Amazon's Bedrock Guardrails milestones sit in the cybersecurity group, which is a judgment call, and they account for three of that group's seven cases.
Scoring rules and limits
Day 0 is the publication date of the sampled announcement, and day 90 is day 0 plus 90 calendar days. A month such as August means through August 31. Early June is coded June 1 to 10. A rollout starting in May is scored on whether access began for any eligible customers during May. Vague wording such as coming soon supplies no deadline. A waitlist is judged on the access route promised, and one documented successful invitation shows the route worked for at least one user.
A match requires a separate dated record or follow-up. A dated update on the same URL counts as distinct when publication and update dates are identifiable. Later pricing and evergreen setup pages count as context only. A later forward-looking update shows a timetable was revised but does not prove when a release actually happened.
This is a documentary audit with no account-level activation tests. An archive check was attempted for AI Max, Asset Studio and HubSpot's Customer Agent, and the snapshots could not be retrieved, so no archive supports a scored outcome. A second coder has not yet reviewed the clarity ratings. The study did not follow cases past their dated records, so Salesforce's spring 2026 target, now past, was not re-checked for this update.
The study will repeat quarterly, with quotas and cohort frozen before later outcomes are checked and results published after every case reaches day 90.
What can buyers and communicators do with these findings?
Buyers should confirm four details before planning around a launch, and communicators can supply them in the announcement itself.
- Find the access stage: immediate, public beta, waitlist, private preview or a dated future release.
- Check the eligible plan, allowlist or administrator setting that applies to your account.
- Look for a region list, since 21 of 30 launch records did not state one.
- Tie any promised date to a usable milestone, and find the changelog or release note that records it.
An announcement that states those four things can be checked later by a second record. A changelog entry, an activation guide or an observed account change is what turns a promise into something a reporter or buyer can cite. Teams can run the same check on their own releases before publishing. Everything-PR's CEO departure communications study applies the same document-first method to press releases and SEC filings.
Summary: what should readers remember?
Sixteen of 30 AI launches were backed by a separate dated record, one was late and one was revised after the window. Clearer announcements were corroborated more often, geography went unstated in 21 of 30 launch records, and the launch post was the only record for five. The strongest launches paired a stated stage, a named buyer group and a dated milestone. Related Everything-PR research includes the AI launch claims study, the emerging industry adoption casebook and the press release coverage audit.
Disclosure
Virgo PR, which published the earlier AI launch claims study referenced here, has overlapping ownership with Everything-PR.