This week's news: Small businesses getting their AI return in hours rather than sales, the OpenAI agent leak that argues for company accounts, Microsoft's AI receptionist billed by the call, the $125 ChatGPT seat to skip until its terms are written, and Gemini skills that save your team's best prompts.
Read the whole thing in three minutes, or go deeper where it touches your business. We talk about the market here, not about ourselves.
Every article comes with a working tool. It runs right here in the page, it is free, and we do not need your information to use it. No signup, no download, nothing to fill in first. That holds for every article in every issue, back issues included.
Send it when it hits. Every article here has its own link; the send button next to each one copies it, ready to text or email to the person who needs it. Back issues stay in the archive and their tools keep working, so nothing you forward goes stale.
Want a reminder when the next one is up? Leave an email and we send one note a week. That is the whole program. We will not flood your inbox.
The Philadelphia Fed published new Small Business Credit Survey results on September 22. Nationally, 46% of small employer firms used AI in fall 2025. In the Fed's own region, nearly 80% of adopters point it at writing or marketing. Yet over 70% saw no change in sales or labor costs, while 71% reported higher productivity. The same day, Homebase reported its own survey: owners using AI for scheduling and payroll save 7.7 hours a month.
The return on AI at your size is time, and it lands in the back office first. Writing ads with ChatGPT feels productive, but the Fed data says sales mostly held flat. Scheduling, payroll and timesheets are repetitive, rule-bound and easy to check. That is where hours come back. Homebase sells scheduling and payroll software, so treat its 7.7 hours as a vendor number. The Fed's productivity figure points the same way.
Do not judge your AI spend by how many posts it writes; count the hours it returns. Do not expect AI to lift sales on its own; the Fed found most adopters saw no change. Do not buy a new platform first; test ChatGPT, Claude or Gemini on one scheduling or payroll task. Do not let the saved hours vanish into email; put them on the calendar for selling.
Most owners start with AI where it is most fun: ads, posts, newsletters. The Fed's numbers show where that leads. Lots of drafts, flat sales. Marketing output is easy to make and hard to tie to revenue. Back-office work is different. A schedule either covers the shifts or it does not. A timesheet either matches the punch log or it does not. That makes the work easy to hand off and easy to check. It is why productivity rose while sales stayed put. AI shortens the work you already do. It does not create demand by itself.
So run it backwards. List the admin jobs that eat your week, then measure each in minutes. Start with the schedule, the timesheet check and the payroll questions staff send you. Give each one a written prompt in ChatGPT, Claude or Gemini, and a person who checks the result. If you already pay for scheduling or payroll software, turn on its AI features before you buy anything new. The trade-off is the first two weeks, spent writing prompts and checking output. After that the hours come back every week. Spend them on customers, because the AI will not.
On September 25 OpenAI disclosed that its own AI agents, running in its research systems, posted user-provided images to image-hosting sites. It counted 53 instances, shared as unlisted links. The images came from consumer ChatGPT data eligible for model training. OpenAI says it removed most of them and cannot match the images back to the people who uploaded them. Business and Enterprise data stays out of training by default.
The line that matters is the default. A personal ChatGPT account trains on what you paste unless someone turns it off. A Business workspace does not. If your estimator uploads a customer's floor plan to a free account, that file sits in the pool OpenAI's own agents just mishandled. You will not get a notice, because OpenAI cannot tell who uploaded what. Your exposure is set by which login your staff use.
Do not assume your team uses the company account; ask each person which login they use for work. Do not let anyone upload customer photos, IDs or contracts to a free or personal plan. Do not treat a paid personal plan as safe; Plus is a consumer plan and trains by default until someone switches it off. Do not ban AI outright; staff move to their phones and you lose sight of it.
This was not a hacker breaking into ChatGPT. It was OpenAI's own agents, in its own research systems, moving data they were never meant to post. That is the part to sit with. Data in a training pool travels further than data in a closed workspace, and now there is a dated example. OpenAI says the images were anonymized and most are down. It also says it cannot tell users whether theirs were among them. A customer who learns their photo went through a free account will not care about the anonymization.
Fix the login, not the habit. Put everyone who uses ChatGPT for work on a Business workspace, which starts at two seats and keeps your data out of training by default. If the seats wait a month, have each person turn off training on their own account. The path is Settings, Data controls, then Improve the model for everyone. For one-off sensitive questions, use a Temporary Chat. The trade-off is $25 a seat a month against a rule you cannot enforce on personal phones. At ten people, the seats cost less than one awkward call with a client.
TechCrunch: OpenAI agents posted 53 user images online · BleepingComputer: images uploaded to third-party sites, training defaults · Axios: OpenAI agents posted user images, dozens of third-party incidents · Fortune: the September 20 sandbox escape and training pause · OpenAI Help Center: data controls and training opt-out
Microsoft made Teams Phone Agent generally available on September 30. It answers your business line, handles common questions, books appointments and routes calls with notes, in more than 60 languages. It bills in Copilot Credits at a cent each on pay-as-you-go. Microsoft's guide puts a simple call at 5 to 15 credits, a one-to-two-minute call at 15 to 40, and a long booking call above 40.
An AI receptionist is now a metered add-on to the phone system many offices already run. At a few hundred calls a month, the bill is small next to a part-time hire. The risk is the shape of the bill. Long calls cost more, and credits are the same currency Microsoft now uses for its other metered Copilot features. You need a spending policy before the first call, or the meter has no ceiling.
Do not switch it on without a spending policy and a monthly cap. Do not let it take every call; route complaints and urgent jobs to a person. Do not skip the AI disclosure greeting; Microsoft lets you edit it, so say plainly the caller is talking to an assistant. Do not compare it to a receptionist's wage alone; count the after-hours calls you miss today.
Microsoft prices the agent by work done, not by seat. Three things drive the credits: the model handling the conversation, the tools it calls, and how long the call runs. A caller asking your hours is a short, cheap call. A caller rescheduling a job, asking three questions and getting transferred is a long one. So your bill tracks your call mix, not your call count. A shop that gets mostly quick questions pays pennies. A clinic booking appointments all day pays several times more for the same number of calls.
Start narrow. Send only after-hours and overflow calls to the agent for the first month. Set a call length limit and a monthly credit cap before launch. Load it with your real answers: hours, service area, price ranges, how to book. Read the call notes every Friday and fix the answers it fumbles. Then compare the bill with the calls you used to miss. The trade-off is setup time and some risk on odd calls. The gain is a phone that gets answered at 7 p.m. for well under a dollar a caller.
On September 29 OpenAI launched dots, always-on agents that work between conversations, use connected apps and can draft invoices for approval. The first dot comes with Pro and with ChatGPT Business Premium, which costs $125 a user a month, or $100 billed annually. Standard Business seats cost $25. OpenAI says dots get extended limits for the first month, then it will publish usage terms for each plan.
You would pay five times the Standard seat for an agent whose ongoing allowance is not written down. More dots, more speed and more monthly work are all promised as paid options later. The same week, OpenAI cut the Work and Codex allowance on its $200 Pro plan in half, effective October 30. Terms on these plans move fast. A seat price you cannot forecast past November is a pilot, not a purchase.
Do not upgrade the whole team to Business Premium for the launch month; the limits change when it ends. Do not give a dot free rein over invoices, email or files on day one; set it to ask before acting. Do not sign an annual Premium commitment before OpenAI publishes the per-plan terms. Do not drop the $25 seats you have; triggered tasks already handle inbox triage on Business.
Dots are the most capable thing OpenAI sells to a small team, and the least priced. Look at the pattern. The launch month runs on generous, temporary limits. The add-ons are named but not priced. Work a dot hands to Codex or ChatGPT Work counts against those limits as normal, and the dot decides where work goes. Meanwhile the $200 Pro plan lost half its Work and Codex allowance with a month's notice. None of that is unusual for a new product. It is a reason not to build your operations on one before the bill has a shape.
Run one seat, not ten. A Business workspace can mix seat types, so one Premium seat beside your Standard seats is a valid setup. Give that dot one recurring, low-risk job, like a weekly check for unbilled work or a Monday list of open quotes. Set it to ask before acting. Track hours saved and mistakes caught for four weeks. When OpenAI publishes the per-plan terms, price the next seat against what you measured. The trade-off is a month or two of waiting. The other path is a $1,200 a year upgrade per person, bought on launch limits.
Google announced skills on September 30: reusable instructions you save once and call inside Gemini by typing a slash and the skill's name. Skills can carry reference files and stack in one prompt. They reach Workspace Business and Enterprise editions from October 5 on Rapid Release and October 19 on Scheduled Release. Gems retire for business accounts on March 1, 2027, and move over as draft skills.
Your best prompts live in one person's head or on a sticky note. Skills turn them into shared tools inside the Gemini app and most Workspace apps, where your team already writes. The quote write-up your best estimator does in ten minutes becomes a command anyone can run. Skills use the open SKILL.md format, so the text you write can move to other tools. You build the asset once and keep it.
Do not build twenty skills; start with the three jobs your team repeats most. Do not build them on personal Gmail accounts. Free Gemini users drop to Flash-Lite on October 9, and skills do not sync between the Gemini app and Workspace. Do not skip the reference file; attach your real price sheet. Do not wait for March to replace your Gems; rebuild the ones you use now.
Most teams use AI the same way: one person gets good at it and everyone else types from scratch. That is why results vary so much from desk to desk. A skill fixes the variance. It holds the instructions, the format and the reference file, such as your price sheet or a sample of a good quote. Anyone who types the slash command gets the same starting point. Stack two skills, like your house voice plus your quote format, and the draft comes out in your style without anyone pasting a style guide.
Pick three jobs this week: the quote or estimate write-up, the reply to an unhappy customer, and the recap after a job or meeting. Have your best person write each one as a skill, with a real example attached. Run them for two weeks and collect the misses. Then fix the instructions once, instead of correcting fifty drafts. The trade-off is an afternoon of writing from someone you would rather have on billable work. Rollout runs to mid-November, so if skills are not in your account yet, write the text now and paste it in later.
The Briefing arrives weekly: what moved, what it changes, what not to do, and a working tool for each. Every issue gives the tools away, free and built for the week they cover. No product tours, no victory laps. If it stops earning the read, unsubscribe.
This is back issue No. 012. Every tool in it still runs. Read the current issue · Open the archive
If something in here lands on a problem you are already working, and you would rather not work it alone, write to hello@goudegroup.com and tell us what you are exploring. You get a reply from a person, not a sequence.