The follow-up nobody sent. Here is what this week's model news does about it.
A new AI model came out this week. If you run a plumbing firm or a salon, the clever demo is not the bit that matters. What matters is that a job which runs over days, like chasing one enquiry until it answers, just got cheaper for an AI to hold. You do not need to do anything about it. If you have a crew, the job is already inside it.
The enquiry that costs a small business the most money is not the one you lost. It is the one you never answered twice.
Someone fills in your form on Tuesday. You reply Wednesday, because Tuesday was a nightmare. They do not answer. By the following Monday you have forgotten them, and so has whichever tab you had them open in. They booked someone else on Thursday. Nobody did anything wrong. The job needed a second and third touch, and the only person who could do that was you, and you were up a ladder.
That is a long job. Quick to do. Long to hold.
What happened in the news
One of the big model makers released a new version on Monday. Most of the coverage is about how clever it is. The line that matters for your week is smaller and more boring: the maker says the model is now much cheaper to run on long, repetitive work, where an AI re-reads the same standing instructions every time it picks the job back up. Their own note says the saving is largest on jobs that run for hours or days rather than seconds. Anthropic's release note and VentureBeat's write-up say the same thing in different words.
Take that out of the demo. Until recently, an AI that knew your tone of voice and your no-Sundays rule had to be reminded of all of it every single time it did anything, and the reminding was the expensive part. This week the reminding got a lot cheaper. That is good news for you, because holding a job open for a week is mostly reminding.
What can come off your desk now
The jobs that suit this are short to do and long to remember. The ones we hear about most from owners:
- The enquiry follow-up. Reply, wait, nudge, wait, nudge once more, then close it or book it.
- The invoice chase. Send, wait for the due date, remind politely, remind less politely, flag it to you.
- The reschedule. Customer cancels, offer two slots, hold them, confirm the one they pick, update the diary.
None of that is hard. It dies when the only person holding it is also the only person doing the work. We wrote about that shape in everything still comes back to your desk. The news this week is that the holding part is finally cheap enough to hand over without a meter running on every nudge.
What it still gets wrong
Worth knowing before anyone sells you a miracle.
It can stop mid-job and not tell you. The other big model maker said this week that its next model comes with a safety monitor which may, in their own words, occasionally flag legitimate activity and slow, pause or stop it. Wired has the quote. Sensible thing for them to build. For you it means an AI that was chasing an invoice might go quiet on Thursday, and if nothing shows you that, the customer will. We described that failure in the quiet failure. A smarter model does not remove it.
It can be confidently wrong about your business. Cheaper memory is not more accurate memory. Tell the model the wrong price once and it will now repeat the wrong price cheaply and often.
It can go too far. A model that can hold a job for a week can also send five follow-ups in a week. There is a line between persistent and rude, and the model does not know where yours is.
Does the WorkMate crew do this job?
Yes, and there is nothing to set up. This is how it runs, in detail, because the question deserves more than a feature list.
The sales mate owns open enquiries. When a lead comes in it drafts the first reply from the shared Brain, in your voice, using the prices and rules you told it once. That draft goes to your review queue. You approve, it sends. If the customer goes quiet, the sales mate drafts the nudge on the day you would have, and that goes to review too. When the customer answers, the thread updates in the Brain, so by the time the bookings mate puts the job in the diary it already knows what was promised. You never brief the second mate on what the first one did.
The invoice chase runs the same way through the accounts mate, with the invoicing module doing the chasing on the dates you set. Rescheduling runs through the bookings mate and the booking module. Same shape every time. The mate holds the job and drafts every outward step. You press send.
Back to the three failure modes. A paused or empty run shows in the activity log as a missing result on a day you expected one, so you see it in your own log rather than hearing it from the customer. A wrong price gets caught in review, and when you correct it the correction sticks in the Brain for every mate, not just that one draft. The nudge cadence is yours to set, because the mate never sends anything you have not approved.
The model underneath the crew is our problem, not yours. When a better one arrives, as one did this week, we move the crew onto it. Your Brain, your brand kit and your approval rules stay put. Nothing to install, nothing to switch on.
The one thing worth doing this week
Count the enquiries from the last month that got exactly one reply from you and then nothing. That number is the job. It was never a technology question. If the honest answer is more than a couple, hand that work over first, before you read another word about which model is cleverest.