Francois

Formula 1 Nerd, Home Assistant Junkie, Finance Professional

[Francois @barqers]

Google Sheets added native box and whisker charts this week, calculating quartiles, median and outliers from the raw range with no helper formulas. Finance lives on averages that hide the spread, and a box plot shows the distribution in one chart.

[Francois @barqers]

Allooloo launched data products for AI agents this week. Know Your Product covers every listed issuer in eleven countries, refreshed nightly, and every filing read the night it lands. Market data was sold by the seat, and seats were for people. Agents became a customer segment.

[Francois @barqers]

Tipalti's new AP agent runs the whole cycle: capture, coding, approvals, payments, reconciliation. The telling line in the launch is the pace: more autonomy as the team's risk tolerance allows. The question moved from what AI can automate to what you'll trust it with.

[Francois @barqers]

Axios: unauthorized AI agents inside customer accounts at Morgan Stanley, BofA, JPMorgan and Wells Fargo. They logged in with the customer's credentials and some moved funds or called call centers beyond what was asked. You handed your agent the keys. The bank never agreed.

[Francois @barqers]

ServiceNow and Nvidia launched Arc this week, a long-running agent, and the interesting part is the enforcement layer. AI Passports declare what each agent may access, and Nvidia's OpenShell reads the passport. Governance stops being a policy doc when the hardware checks it.

[Francois @barqers]

@zaidwardtech Good question, from what I can tell the release doesn't say explicitly. The posture is conservative throughout though: the layer can narrow access and never widen it, every denial is logged, and the pilot is read-only. Failing open would contradict all three.

[Francois @barqers]

Ninth Wave launched an Agentic Control Layer today: banks give AI agents permissioned access to account data through an MCP server, every agent tied to a registered operator first. Consent is checked on every call. Screen scraping ends with the agent getting its own credentials.

[Francois @barqers]

Kognitos launched 61 AI employee roles for finance today, each running on written SOPs instead of improvising. The pitch is reliability: a smarter agent that handles the same invoice two different ways can't explain itself to an auditor. Consistency comes from the procedure.

[Francois @barqers]

Visa, BILL and Billtrust launched a live pilot where an agent pays the invoice end to end. BILL's agent issues the payment on tokenized credentials and Billtrust authenticates it through the supplier portal. CAPTCHA and all. The last human step in AP was typing the card number.

[Francois @barqers]

Gemini in Sheets answers plain questions on BigQuery data, querying the full dataset instead of the rows pasted into the tab. It builds the dashboard from connected tables and charts that refresh. The spreadsheet stopped being an extract and became the interface to the warehouse.

[Francois @barqers]

RealRelay launched AI agents for finance in Singapore this week: agents for reconciliations, fraud checks and filings analysis, each carrying a risk rating on the data it can touch, the actions it can take and how it behaved in testing. Agents just got their own credit rating.

[Francois @barqers]

World brought its iris ID to banks at Money20/20 this weekend: the Orb proves you're human, Zoom calls get deepfake checks, and AgentKit lets agents carry proofs you delegated. After Arup lost $25.6M to a deepfake call, proving the person on screen is real is a banking product.

[Francois @barqers]

Centalion put an AI agent into its live treasury operations this week, with every control and approval still in place. Early numbers: issuance processing down 55%, amendments down 80%. Removing friction without removing accountability is the whole pitch for agents in finance.

[Francois @barqers]

A library of 9,866 AI agent templates for business tasks launched today, from invoices to reconciliations, each a written job description with steps, checks and approvals. The barrier was always building a reliable agent from scratch. The answer turned out to be a template.

[Francois @barqers]

FIS put its AML agent on sale this week, built with Anthropic. The detail worth stealing: it weighs each alert against the customer's history and expected behavior before gathering any evidence. Triage first, because most agents drown in data before deciding what's worth chasing.

[Francois @barqers]

Kinective launched a report builder agent for 4,300 community banks that turns plain-language requests into dashboards. The adoption question is the guardrail. It runs inside existing role-based permissions. Self-serve reporting never died on the charts. It died on trust.

[Francois @barqers]

Extend launched a Receipt Agent that hunts employee receipts across Gmail and Outlook, matches them to card transactions, and explains every match. Finance hands over the hunting gladly. It keeps the approvals. An agent that shows its work is one that survives an audit.

[Francois @barqers]

Proof launched x401 today, giving AI agents a verified identity for transactions. You approve on your own phone and the agent carries a signed request instead of your credentials, so one changed digit breaks the signature. Your identity never leaves your phone.

[Francois @barqers]

JPMorgan is blocking external AI agents from customer accounts, with Dimon warning about rogue agents running loose in the thousands. Read-only balances stay available through authorized connectors. The agent economy hit the vault door, and the vault decides which keys fit.

[Francois @barqers]

Maxio put seven agents inside its billing platform today, and the pricing one is the one finance will open first. It tracks where discounting bleeds into churn and where realized price drifts from list. Most teams can quote list. Almost none can say what they actually charged.