Two new AI models landed. Here's what actually changes in a business like yours.

On September 1, Anthropic released Claude Fable 5.1. On September 3, OpenAI released GPT-6 Astra. Between them, those two companies power most of the AI a small business ever touches.

You can find a hundred articles arguing about which one won which test. This isn't one of them. Benchmarks don't run your business.

Two questions are worth your time. What did people actually get these things to do? And what do you now have to decide, because a tool that can open your software and change records is a different thing than a chat window, and it needs an answer from you before someone on your team points it at a customer file.

What people are doing with them

OpenAI's release went hard at one thing: getting an AI to operate the software you already own. Week-one examples (lmspedia, OpenAI):

• Open the customer database, find every record that breaks your own rules (no phone number, no owner, duplicates), fix them, and log every change so you can check the work.
• Fill in supplier and vendor onboarding portals from the source documents, then stop before submit and flag the fields a human needs to look at.
• Triage the shared inbox against your written rules: draft the replies, propose the meeting times, send nothing.
• Match the bank statement against the ledger for a date range and hand back the exceptions list instead of the whole month.
• Read forty contract PDFs and return the fields you asked for in a table, with the page number each one came from.
• Go find current pricing and positioning across five competitors, and return a comparison table with a source link on every row.

Anthropic went the other way: less clicking, more long unattended work on messy material (Anthropic). Documents, spreadsheets and slides from a blank page to a finished draft. Reading charts and tables buried inside ugly PDFs. Multi-step research that runs while you do something else.

They also cut the cost of the model re-reading material it has already seen by 75%. That sounds like a technical footnote, but it's the exact cost that piles up when an AI works on one long job for an hour, so that kind of work got roughly 25 to 45% cheaper to run.

The one that made the rounds: a creator handed Astra a brief, not instructions, and got back a finished video. It checked its sources, wrote the script, generated the voiceover, cut the timeline including the sound design, then transcribed its own output and compared it to the script to catch its own mistakes. About 50 minutes, roughly $60 in usage (MindStudio).

Read the fine print on that one, because it's the whole article in miniature: it worked because the tools and accounts were already set up and connected, and because someone wrote a brief.

The model didn't build the workshop. It used one that existed.

Three examples from businesses that look like yours

Generic use-case lists are easy. Here's what this actually looks like in the kind of shop we work with: 3 to 30 people, owner still on the tools.

The retyping chain

This is the most common shape of waste we find, and it's nearly universal in the service trades.

A lead comes in on the website form. It lands in the office inbox. Someone forwards it to the owner. The owner decides who takes it. The office types it into the job list. The work happens. The owner says "done." Only then does someone build an invoice.

Every arrow in that chain is a person moving information from one box to another by hand, and nothing in it has a status anybody can look up. The failure isn't dramatic. It's a finished job sitting three days between complete and invoiced because the only person who knew was up a ladder.

Where AI fits: the retyping. Intake into the job list, job record into the draft invoice, chasing what's finished but unbilled.

Where it doesn't: the owner deciding whether this job goes to an apprentice, a journeyman, or himself. That call is priced on his read of the customer, the site and his crew. It is the business. Anyone selling you automation for that is selling you a problem.

The stack of quotes that never went out

In formwork, tree work, electrical, the pattern is identical. Every quote runs through the owner. Plans arrive as PDFs, he does the takeoff by hand at the kitchen table at 9 p.m. after a full day on site, and there's a stack he hasn't priced. Nobody else in the shop can build one. When he's buried on a job, nothing goes out, and the work goes to whoever answered first.

Where AI fits: everything around the number. Pulling dimensions and quantities off the plans into a takeoff sheet he checks. Turning his scribbled numbers into a clean quote in his format. Drafting the follow-up on the ones that went quiet. Building the job-cost comparison afterward, quoted versus actual, so next year's prices come off evidence instead of memory.

Where it doesn't: the risk. What this GC is like to get paid by. Whether that removal near the house is a four-hour job or an eight-hour one. Whether to hold the price when steel has moved 20% since you quoted it.

And before any of it, check that you have the problem you think you have. If half your quotes never convert because you're the third of three bids, quoting twice as fast just gets you to no faster. Do you actually lose jobs on speed, or on price? Most owners have never checked, and it's a different fix.

The compliance dates that live in a folder and a memory

Every shop is carrying dates attached to named people. Certifications with expiries. Working-at-heights training that lapses after three years. In Quebec construction, the CCQ monthly report is a filing and a payment due by the 15th, with a penalty of 7% if you're up to a week late, 11% at eight to fourteen days, and 20% past that (CCQ). Almost nobody has those dates anywhere but a folder and their head.

Where AI fits: building the register. Point it at the folder, have it pull every certificate, expiry date and named person into one sheet, and set the reminders.

This is the best first job to give it. It's read-only, you can verify every line against the source document, and if it gets one wrong you catch it before it costs anything. Compare that to letting it touch invoicing on day one.

The decisions you actually have to make

This is the part the launch coverage skips, and it's the part that determines whether any of the above works or blows up in your face.

1. Where does it stop?

Look at the good use cases again: fills the form, doesn't submit. Drafts the reply, doesn't send. Produces the exceptions list, doesn't touch the books.

Decide, in writing, which actions in your business are irreversible: money moving, something going to a customer, a record being deleted, a filing being submitted. Put a human on every one of them. Everything before that line can move fast.

That's not caution, it's design. You get the speed without betting your reputation on a machine that is occasionally, fluently wrong.

2. What is never allowed to go in?

Write the short list of things nobody pastes into an AI tool. Customer names and addresses. Employee files. Bank details. The contract with your biggest client. Anything covered by an NDA.

If you're in Quebec, this one has teeth. Under Law 25, an employee pasting client personal information into the wrong AI tool is a confidentiality incident. If it presents a risk of serious injury you have to notify the Commission d'accès à l'information and every affected person, and you're required to keep a register of incidents regardless (P-39.1, ss. 3.5–3.8). Most owners have never been told this applies to them.

Outside Quebec there's no equivalent statute today. The exposure is still real: it's your customer's information sitting in a system you don't control, and your customer didn't agree to that.

3. Who is already using it, and on what?

Not whether. Who. It's already in your shop. Somebody is drafting quotes with it right now and hasn't mentioned it, because nobody asked and they weren't sure it was allowed.

The pattern we keep finding: one long-running chat per function. One for quotes, one for invoices, reused across every customer. That has two consequences. It costs more than it should. And one customer's numbers end up in another customer's quote, because the tool is reading the whole conversation and your last three jobs are in it.

That's not a hypothetical. It's the single most common real failure we see, and it produces a wrong number on a document with your name at the top.

Ask the question out loud, without making it a trial. You'll learn more in ten minutes than from any policy you write afterward.

4. Free or paid, and don't assume paid is safer

Free tiers generally handle your data differently, and that matters the moment someone pastes anything about a real customer. Pay for the business version.

But do not assume that paying settles it. We tear these tools down for clients, and we have found products where the paid tier retains your data indefinitely while the free tier expires it in 30 days. "We pay for the business version so we're fine" is a belief a lot of owners hold, and no vendor is going to correct it. Read the terms for the tier you're actually on, or have someone read them for you.

5. Is AI making decisions about people?

Screening job applications. Ranking candidates. Anything that sorts humans.

In Quebec, if a decision about someone is based exclusively on automated processing, you have to tell them, and they can ask for an explanation and to have it reviewed (Law 25, art. 12.1). That's a legal obligation, not a nice-to-have. Even where it isn't law, an AI that quietly filters out a good candidate is a hiring problem you won't detect for a year.

Watch them yourself

Both labs put out a short announcement video. Between them they take less time than finishing this article, and you get to see the thing move rather than read a description of it.

Watch them the way you would watch a supplier's product video, because that is what they are. Every demo is one attempt that worked, with the failed takes left out, and about a fifth of the most-shared Astra demos circulating that week turned out to be OpenAI's own marketing reshared as though it were independent (explainX). They are worth your time for what is possible, not for how easily.

So which one should you use?

Honestly, it doesn't matter. Flip a coin.

Both are excellent. Both are far better than what you were using a year ago, and whichever you land on will do good things in your business. The gap between them is smaller than the gap between a Tuesday where you use it and a Tuesday where you forget.

Already paying for one? Stay there. Neither? Pick the one your team will actually open, pay for it, and get on with it.

Because the technology was never the variable.

Go back through this article and notice that not one recommendation depended on which model you picked. Where it stops. What never goes in. Who's already using it. What stays yours. Every one of those is a decision about your business, and every one of them would be identical if both companies had shipped nothing this month.

There's one more, and it's the one that decides whether any of this works. Every use case above runs on something somebody already wrote down. The record cleanup works because there are rules about what a good record looks like. The inbox triage works because someone decided what gets answered how. The reconciliation works because there's a ledger.

If your quoting process lives entirely in your head, no model on earth can run it. That isn't a reason to wait. It's the actual first task.

The model is the fast part. The process is the part that makes it worth anything, and the part nobody can sell you.

If you're spending your evenings reading comparison charts and waiting for the next release before you start, you're optimizing the one input that barely moves the outcome. The businesses that get real results this year aren't the ones that picked correctly in September. They're the ones who wrote down how their work actually gets done, decided where the machine stops, and then handed it to whichever model was sitting in front of them.

Pick one. Build the process. That's the job.

---

Modus is an operations and AI consulting firm based in Sherbrooke, Quebec, serving founder-led businesses across Canada. We diagnose how a business runs, fix the most expensive friction, and use AI only where the return is real. We don't resell either of these tools and we're not paid by either company. Nothing here is legal advice: the Law 25 points are the general rules, not a read on your situation.