Daily AI · 2026-06-26

Useful AI Daily - June 26, 2026

Today's useful AI signal is control around real work. Microsoft pushed Copilot in Excel deeper into finance workflows, GitHub made Copilot for Jira generally available, GitHub added stricter plugin-marketplace controls, OpenAI's Broadcom-built inference chip became a cost-and-capacity signal, and Copilot code review got cost-efficiency and review-depth updates.

Try AI where the work is traceable: a copied spreadsheet, a low-risk Jira ticket, a reviewed pull request. Watch permissions, plugin sources, billing, and capacity. Skip any tool that cannot show inputs, edits, costs, and the human review path.

Updated 2026-06-26 · ai-daily, enterprise-ai, developer-tools, ai-governance

The Short Version

Try with audit trails: Copilot in Excel is getting finance skills, financial data connectors, and traceability features aimed at reviewable spreadsheet work.
Try one ticket: Copilot for Jira is now generally available with live agent progress in Jira and follow-up steering on the same draft pull request.
Lock down plugins: GitHub Enterprise admins can now restrict Copilot CLI and VS Code plugin installs to known marketplaces.
Watch infrastructure: TechCrunch reports OpenAI unveiled its first custom Broadcom-built inference chip, a capacity signal rather than a product to try today.
Tune review depth: Copilot code review says file-exploration changes cut review costs by about 20% while adding clearer Medium-depth controls.

5 Updates Worth Your Time

Try with audit trails Microsoft Copilot in Excel finance update

Copilot in Excel is moving AI into reviewable finance work

What changed: Microsoft said on June 25 that Copilot in Excel is adding finance-oriented skills, new financial data connectors, and improved traceability. The update includes sample finance skills, support for custom skills stored as SKILL.md files in OneDrive, and connectors from providers such as CB Insights, Daloopa, and FactSet.
Why it matters: Who should care: analysts, finance teams, operators, founders, and anyone who lives in spreadsheets. AI in finance has to show sources, formulas, and changed cells because a confident answer is useless if the workbook cannot be reviewed.
Try, watch, or skip?: Try it only on a copied workbook first. Ask for one variance analysis, one forecast refresh, or one formula explanation, then compare every cited source and changed cell before using the result in a live report.

Read source

Try one ticket GitHub Copilot for Jira GA

GitHub Copilot for Jira is ready for a cautious handoff test

What changed: GitHub made Copilot for Jira generally available on June 25. The GA release adds real-time agent progress in Jira, follow-up instructions from the Jira chat panel after a draft pull request opens, and simpler onboarding for connecting GitHub organizations and repositories.
Why it matters: Who should care: product teams, engineering managers, agency teams, and indie builders using Jira plus GitHub. The useful part is not a smarter chat box; it is moving from ticket intent to visible implementation progress without losing the review trail.
Try, watch, or skip?: Try one low-risk Jira issue with a clear acceptance test. Keep customer data and secrets out of the ticket, watch the streamed progress, and require a human review of the draft pull request before merge.

Read source

Lock down plugins GitHub strictKnownMarketplaces update

Copilot plugin governance moved closer to the user device

What changed: GitHub said enterprise-managed settings now support strictKnownMarketplaces in VS Code and GitHub Copilot CLI. In public preview, enterprises can define allowed plugin marketplaces so Copilot only permits installs from known sources for licensed Copilot Business or Copilot Enterprise users.
Why it matters: Who should care: security leads, IT admins, platform teams, and anyone approving AI tools that can run near source code. Plugins are where AI assistants become operational, so marketplace governance is now part of AI risk management.
Try, watch, or skip?: If you run a team, set an allowlist before users install plugins at random. If you are solo, use the same habit manually: install only tools you can name, audit, update, and remove.

Read source

Watch infrastructure TechCrunch on OpenAI's Broadcom chip

OpenAI's Broadcom chip is a capacity signal, not a tool to try

What changed: TechCrunch reported on June 24 that OpenAI unveiled its first custom chip built by Broadcom. The chip, called Jalapeno in the report, is designed for inference, meaning the high-volume work of running AI models after they have been trained.
Why it matters: Who should care: AI tool buyers, founders, investors, and anyone wondering why AI subscriptions, limits, and latency keep changing. Custom inference hardware is about cost, capacity, and dependence on a small number of chip suppliers.
Try, watch, or skip?: Do not try anything today. Watch whether lower inference costs eventually show up as more generous usage limits, better reliability, or lower prices. Until then, keep budgets and fallback providers in place.

Read source

Tune review depth GitHub Copilot code review update

Copilot code review got cheaper, but review settings matter more

What changed: GitHub said Copilot code review now uses file exploration tools from the Copilot CLI and SDK, including grep, rg, glob, and view. GitHub says the change reduced code review costs by about 20% while keeping review quality steady, and Medium analysis depth users now get attribution plus an organization-level default setting.
Why it matters: Who should care: developers, maintainers, and teams trying AI review at scale. Cheaper review can make AI checks more routine, but depth settings decide whether a review is a light lint pass or a deeper safety net.
Try, watch, or skip?: Try it on one repo with known recent bugs. Compare Medium-depth AI review against human review notes, then set defaults by repo risk instead of using the same depth everywhere.

Read source

Tool Worth Trying Today

Copilot for Jira two-pass ticket handoff

Pick one small Jira issue, let Copilot produce a draft pull request, then use the Jira chat panel for exactly one follow-up instruction before human review.

Best for: Teams already using Jira and GitHub who want to test agent work without moving planning, progress, and pull-request review into separate places.

Watch out: Do not include private customer reports, unreleased roadmap detail, legal text, credentials, or production incident data until Jira, Confluence, GitHub, and Copilot permissions are reviewed together.

Source: GitHub Copilot for Jira GA

Privacy / Cost Watch

Finance spreadsheets often contain payroll, forecasts, customer revenue, acquisition plans, and board material. Use copied files and approved connectors until retention, admin controls, and audit requirements are clear.
Jira and Confluence context can expose more than the active ticket. Clean the issue, remove secrets, and check project permissions before asking an agent to work from it.
Plugin marketplaces are now an AI security boundary. Restrict Copilot CLI and VS Code plugins to known sources before letting assistants run near private repositories.
Usage-based AI coding can hide cost until the invoice arrives. GitHub's cost-center update lets enterprise teams attribute usage to teams and set budgets; smaller teams should still track cost per shipped change.
AI code review is a useful second pass, not a replacement for ownership. Keep human review for security-sensitive changes, migrations, auth code, payments, and customer data paths.

One Practical Workflow

Run a 20-minute AI work-control review

Pick one AI workflow you might use this week: spreadsheet analysis, ticket handoff, plugin install, coding assistant, or code review.
Write the exact output you want in one sentence.
List the data it can read and mark anything private, regulated, financial, customer-related, or unreleased.
List what the tool can change: workbook cells, Jira text, pull requests, repo files, settings, or installed plugins.
Add one cost limit, one permission limit, one review owner, and one rollback step.
Run the workflow only if all four controls are visible before the tool starts.

Builder Note

The product lesson today is that AI tools earn trust through boring controls. Source tracing in Excel, ticket-visible agent progress, plugin allowlists, team budgets, and review-depth labels all make a product easier to adopt than another impressive demo.

Ignore For Now

Ignore autonomy that hides controls

Skip AI tools that promise to finish work across spreadsheets, tickets, repos, or plugins but cannot show data scope, permission source, cost owner, review status, and a fast stop path.

Bottom Line

Bottom line: the practical AI story today is not a new model race. It is whether useful assistants can survive contact with finance review, Jira tickets, plugin risk, coding budgets, and pull-request quality. Try one workflow, but make the controls visible before the work begins.

Sources

Past issues RSS