Issue #3: 5 AI-agent tricks for bookkeeping, support, follow-ups and ops

5 new agent tricks worth trying this week. For business owners, operators and anyone running AI agents. This week: less time on bookkeeping, support triage, meeting follow-ups and reports. Five minutes to read, each trick checked against its source.

1. Let OpenAI's Decisions API sort your bank transactions into categories

What it does: the Decisions API went generally available on 22 Oct. It doesn't write text. It answers bounded questions: yes/no (with a probability), one option from a list you supply, or a score on a rubric. OpenAI says its own Personal Finance team uses it to categorize transaction descriptions. It costs $0.10 per million input tokens, with no charge for output tokens.

Who it's for: anyone who categorizes bank lines, expenses or receipts by hand, or pays someone to. You need an OpenAI API key, or you can test first in the Decisions Playground.

How:

In the Decisions Playground, paste 20 real transaction descriptions and add a choice question whose choices are your expense categories, each with a one-line description. Include an other choice so odd lines have somewhere to go.
Move it to code with POST /v1/decisions and the model gpt-6-luna (the only model it supports). Each answer comes back with a confidence value. Keep your category list in the shared input so repeat requests can use the prompt cache (the reused part has to be at least 1,024 tokens).
Pick a confidence cut-off by checking it against a month you've already categorized. Book the high-confidence lines automatically and send the rest, plus everything marked other, to whoever does your books.

Why bother: in OpenAI's launch post, Ramp says it attributes AI costs 4x faster and 30% cheaper than with plain GPT‑6 Luna, and Sierra reports about 90% lower cost per call than its previous approach.

Source: OpenAI, "Build more responsive apps with the Decisions API" https://openai.com/index/decisions-api · developer guide https://developers.openai.com/api/docs/guides/decisions

2. Triage your support inbox on your own computer, for free, with an open model

What it does: Ollama v0.40.11 (24 Oct) added d3, a family of open decision models from vLLM Semantic Router, released under Apache 2.0. Like the API above, d3 doesn't chat. It picks a category, answers yes or no, or rates something, and gives a probability for each option. It runs on your own computer, so customer emails stay there and there's no per-message bill.

Who it's for: small teams who get a steady flow of support or sales email and don't want to send customer data to a cloud model. The default 4B model is a good starting point on a 16 GB machine. The 2B model was tested on an 8 GB Mac.

How:

Install or update Ollama (d3 needs 0.40.9 or later), then try it on one message: ollama run d3 "A customer says their card was charged twice." --ask "Which team should handle this?" --choice 'billing=Invoices and charges' --choice 'returns=Refunds and replacements' --choice 'technical=App errors'
To run it from a script or an automation, call the local endpoint http://localhost:11434/v1/systemone (or ollama.systemone in the Python and JavaScript libraries). You can ask several questions about the same email in one request, such as which team, whether it's urgent and whether it's a refund request.
Ollama warns that scores aren't guaranteed real-world frequencies. Test it on a batch of past emails you've already sorted before you trust a threshold. Keep messages short, because these packages use a 4,096-token context.

Source: Ollama v0.40.11 release notes https://github.com/ollama/ollama/releases/tag/v0.40.11 · d3 model page https://ollama.com/library/d3

3. Get meeting decisions and action items posted to Google Chat automatically

What it does: Google is rolling out meeting recaps in Google Chat. When "Take notes for me" is on during a Meet call, Gemini posts a card in the meeting's Chat conversation when the call ends. The card has a summary, key decisions, action items with owners, and links to the notes, recording and transcript. Nobody has to write the "here's what we agreed" message.

Who it's for: teams on Google Workspace Business Standard or Plus, Enterprise Standard or Plus, or Frontline Plus, with Gemini turned on. Rapid Release domains started getting it on 16 Oct. Scheduled Release domains start on 27 Oct, and it can take up to 15 days to show up.

How:

Start "Take notes for me" in the Meet call. The recap only posts when the meeting has a Chat conversation linked to it.
Check the host's "Send notes to" setting in Meet. If it's set to "Hosts and co-hosts only," no recap card is posted.
After the call, open the meeting's conversation in Google Chat and find the recap card at the bottom. Turn its action items into your client follow-up or CRM tasks while the call is still fresh. If you delete the card, the summary is gone, but the Gemini notes doc stays in Drive.

Source: Google Workspace Updates, "Follow through on your meeting tasks with meeting recaps in Google Chat" (19 Oct) https://workspaceupdates.googleblog.com/2026/10/follow-through-on-your-meeting-tasks-with-meeting-recaps-in-Google-Chat.html · Google Chat Help https://support.google.com/chat/answer/17140251?hl=en

4. Have Grok Bot keep a live "who owes us money" page

What it does: Grok Bot v0.75.2 (22 Oct) added Pages: reports, memos and dashboards that you edit together or have the Bot keep up to date. QuickBooks Online became available to connect in v0.70.2 (11 Oct), and Zoho CRM in v0.75.2. Put them together and you get one page with unpaid invoices and open deals that updates itself, instead of a Monday spreadsheet session.

Who it's for: owners and operators who run their books in QuickBooks Online or their pipeline in Zoho CRM and want one place to check cash and follow-ups.

How:

Connect QuickBooks Online (and Zoho CRM if you use it). For Zoho, pick your account's regional data center before you sign in.
Ask the Bot to build a Page of open invoices by customer, amount and days overdue, plus open deals with their next step, and to keep it up to date on a schedule, for example every Monday morning. Your schedules appear in the Routines tab of the Bot's details pane.
Ask it to draft reminder emails for the most overdue invoices. Since v0.75.5 (23 Oct), emails written on a routine run wait as drafts for you to approve. To share the page, use Download (PDF), and use the dashboard's Pages section to unshare it later.

Source: Grok Bot changelog (v0.70.2, v0.75.2 and v0.75.5) https://x.ai/changelog/bot

5. Make sure n8n catches up on scheduled jobs it missed while it was down

What it does: from n8n 2.45.0 (20 Oct), self-hosted n8n runs schedules from a database queue. Before that, scheduled runs lived in memory, so if the server restarted at 09:00, the 09:00 invoice-reminder or daily-report run never happened. Now upcoming runs are saved in advance, and each Schedule Trigger can catch up on a missed run instead of skipping it.

Who it's for: anyone self-hosting n8n with scheduled workflows that cost money when they skip, like payment reminders, stock syncs or end-of-day reports.

How:

Upgrade to 2.45.0 or later, where the durable scheduler is on by default. From 2.36.0 you can switch it on yourself with N8N_SCHEDULER_ENABLED=true.
On each Schedule Trigger that matters, set the "If Execution Is Missed" option to "Run the Most Recent Missed Execution". The default still skips missed runs. Set the "Missed Execution Grace Period" option to how long a run can be late before it counts as missed (the default is 60 seconds). According to the docs, only Schedule Trigger nodes added in 2.36.0 or later have these options.
Check your "every N minutes" schedules. They now run at fixed gaps counted from when you published the workflow, not on clock boundaries. If something has to fire exactly on :00 and :30, set N8N_SCHEDULER_TRIGGER_NODE_MODE=legacy.

Source: n8n docs, "Durable scheduler" https://docs.n8n.io/deploy/host-n8n/configure-n8n/durable-scheduler · n8n release notes https://docs.n8n.io/changelog

⚠️ One thing to watch

If anything you run uses LiteLLM, the popular proxy for routing requests to different AI models, check the version. On 21 Oct, attackers took over LiteLLM's release pipeline and published litellm==1.107.0 to PyPI and Docker with malware that steals credentials. The PyPI package was live for about two hours, and the latest, main-latest and main-stable Docker tags were also affected. If you installed or started it in that window, LiteLLM says to treat the machine as compromised, rebuild pinned to 1.106.2 and rotate every key the machine could reach. Their write-up: https://docs.litellm.ai/blog/security-incident-oct-21-2026

Sources credited above. All summaries are ours. Got a trick that worked for you? Reply and tell us.

Get next week's five, free · All issues

Latest Daily Frontier Digest: 29 Oct 2026: An unreleased Claude model built working exploit chains for Chrome, Safari and Linux