What changed this summer, and what to check before October

August is a strange month to run a small business: half the team is away, the phone is quieter, and nobody has the appetite to read anything that opens with the word "regulation". Which is a shame, because three things shifted this summer that land squarely on your desk: what it costs to reply on WhatsApp, what your site has to tell customers, and what you pay per automated answer. Here's the short version of each, what to do about it, and — just as usefully — what you can safely ignore.

One: from 1 October, replying on WhatsApp stops being free

If you use the WhatsApp Business Platform (the API, not the free phone app), service messages — the ordinary replies you send inside the 24-hour window — become billable per message on 1 October. We pulled that apart when it was announced, in this piece on the pricing change, so I won't repeat the whole thing.

What's new since then is that the other half is already live. Since 1 August, Meta's own Business Agent is billed by tokens at $2 per million, which Meta puts at roughly 4–5 cents per reply. That's the only price in this article I'll quote with a straight face, because it's published in Meta's pricing documentation.

The number you actually need is duller than it sounds: how many messages you sent last month, split between templates and in-window replies. Your provider has it. If there's no dashboard showing it, email them and ask. Once Meta publishes the rate for your market you can do the arithmetic yourself — and note that rates only change on the first day of a quarter, with notice, so there are no mid-month surprises.

One reassuring detail: messages the customer sends you still cost nothing. You only pay for what leaves your side.

Two: AI transparency is no longer a date in the diary

On 2 August 2026 the EU AI Act's transparency obligations became applicable and enforceable, with no extension this time. The European Commission announced it that day in its note "Safer and more transparent AI", which also spells out who polices it: national market surveillance authorities, with fines up to €15 million or 3% of global annual turnover, applied proportionately to SMEs.

Before that number ruins your afternoon: for an ordinary business, compliance is a sentence. Two questions:

1. Does your assistant say it's an assistant in its first message? If a customer could reasonably think they're talking to a person, you're not compliant. This is a one-line fix.
2. Do you publish AI-generated content that needs marking? That covers images, audio or video resembling real people or events, and text published to inform the public where nobody reviewed it. The EU has even published a set of icons for labelling it.

If your AI answers questions and books appointments, your duty starts and ends with question one. If you also generate social creative with AI, give question two a careful read. The full walkthrough, with wording you can copy, is in our article on the transparency rules.

Three: models keep getting cheaper, which changes the maths

The third change has no deadline and no official gazette: the price per answer from language models has been falling for months and is still falling. I'm not going to quote figures from third-party comparison tables — they move every few weeks and half of them are stale the day they publish. Open your own provider's pricing page and write down today's number and today's date.

The interesting part isn't the saving, it's what the saving unlocks. Tasks that didn't pay for themselves a year ago — summarising every conversation, classifying every inbound email, drafting a quote for every enquiry — now cost little enough to be worth revisiting. If you ran those numbers in 2025 and the answer was no, run them again. The answer may have flipped.

And here's the quiet irony of this quarter. Put the two together — the message that starts costing money in October, the model that keeps getting cheaper — and from the autumn it's quite possible that the envelope costs more than the letter: more to carry the message than to think it. Don't assume that without checking it against your own rates. But if it's true for you, the lever changes: it stops being "find a cheaper model" and becomes "resolve it in fewer messages".

What you can ignore without guilt

Nobody writes this part, so here it is.

• High-risk obligations. They've been pushed to 2 December 2027 for standalone systems and 2 August 2028 for AI embedded in products like machinery or medical devices. Unless you're screening CVs, scoring credit or similar, this isn't about you.
• Headlines saying "the EU AI Act has been delayed". Half true. The heavy block moved; transparency didn't.
• Switching provider over the October change. At a few dozen messages a day you're talking about a few euros a month. Migrating over that trades a cents problem for a weeks problem.
• Rebuilding your agent to save tokens. If your model bill is under two figures a month, there's nothing to optimise. That time earns more almost anywhere else.

The half hour that's actually worth it

Block out a Friday afternoon slot, laptop open, phone face down. Five things:

1. Ask for your volume (10 minutes). Messages sent last month, templates separated from service replies. Write the number somewhere you'll find it again in January.
2. Read your agent's opening line (2 minutes). Does it identify itself? Does it offer a route to a human? If either is missing, fix it today.
3. Check your AI provider's pricing page (5 minutes). Note the current rate and the date. It's the only way to know in three months whether you gained or lost ground.
4. Count the messages in a typical conversation (5 minutes). Take three real closed threads and count. If it took eight messages and could have taken four, that's your October saving, found.
5. Diary it for November (1 minute). Repeat every quarter. That alone puts you ahead of most businesses on your street.

None of this requires understanding the technology. It requires looking at four numbers once a quarter — which is exactly the work we do with anyone who sets up a receptionist with us: she resolves things in as few messages as possible, introduces herself as what she is, and anything doubtful lands on your screen as a draft. You approve, she executes.