Some of the work you were about to commission as a custom automation this quarter is now included in software you already pay for. In the last fortnight of September 2026, three of the largest vendors in the SME software stack put working agents — not chat assistants — inside their products, and the cost of the models underneath them dropped sharply at the same time. That is genuinely good news, provided you can tell which jobs the bundles absorb and which ones they will quietly fail at in month three.
The honest answer is that the bundled agents take a narrow, well-defined band: single-system tasks, few business rules, running against data a vendor connector already reaches. Everything outside that band — your non-connected line-of-business systems, the rules that live in someone's head, the evidence trail you need when a transaction is disputed — still requires deliberate integration work. Here is the map.
What actually landed
On 29 September 2026, Meta unveiled Muse for Small Business, an agent that connects to applications including Asana, Zoom, Intuit, Box, Canva and Salesforce Slack. The Muse app is free within usage limits, with paid usage beyond them; reporting as at 9 October 2026 lists a Power plan at US$20 per month — roughly A$30, converting from the USD figure — and Australian availability is worth confirming before you plan around it.
Four days earlier, on 25 September, Microsoft relaunched its Copilot app, splitting standard Microsoft 365 Copilot from three metered capabilities — Cowork, Code and Autopilot — billed by usage rather than per seat. Autopilot, previously called Scout, creates cloud agents that monitor email and Teams channels and perform recurring tasks without being prompted each time. That is a product-category shift: from "ask the assistant when you need it" to "give the agent a standing responsibility".
A September 2026 industry roundup also reports that Google connected Gemini to Salesforce, HubSpot and Asana and launched Workspace Studio during the month. That report does not carry a precise launch date or primary Google documentation, so treat it as a lead to verify against your own tenant rather than a settled fact.
Underneath all three, model pricing moved. OpenAI reportedly released GPT‑6.1 Sol on 29 September 2026 at US$2 per million input tokens and US$10 per million output tokens, with cached input at US$0.10 — roughly A$3 and A$15 per million at current rates, converting from the USD figures. A smaller Luna tier is reported at US$0.10 and US$0.50 per million, and Anthropic reportedly shipped Claude Opus 5.5 at around 40 per cent below Opus 5. A specialist index also lists a US$500-per-month Pro 500 plan from 29 September. Several of these are secondary reports and model names and prices should be checked against vendor pricing pages before you build a budget on them.
The band the bundles absorb
Strip the marketing away and bundled agents are good at a consistent shape of task. They work when all four of these hold:
- The data lives in one system, or in two systems the vendor has already built a connector for.
- The rules are shallow. Summarise this thread. Create a task from these meeting notes. Draft a follow-up. Update this CRM field.
- Being wrong occasionally is survivable. A clumsy draft email costs a minute. A clumsy credit note costs a customer.
- A human sees the output before it matters. The agent proposes, a person despatches.
Inside that band, the bundles are now hard to beat on price, because the marginal cost to you is close to zero and the integration is the vendor's problem. If you were paying a consultant for a small Zapier or Make scenario that moved a lead from a form into HubSpot and pinged a Slack channel, check whether Workspace Studio or Muse does it on your existing licence first. If you were about to buy a third-party meeting-notes-to-tasks tool, check Copilot. Cancelling a subscription you no longer need is the cheapest automation win available this quarter.
Where they stop
Your non-connected systems
Connector lists are the limiting factor, and the published lists are built for the American mid-market. Australian SMEs run job management, payroll, compliance, dispatch, point-of-sale and industry-specific systems that will never appear on a Meta or Google connector page — an on-premise estimating package, a scheduling system from a vendor with eleven customers, an ERP module someone configured in 2014. If the agent cannot read the system, it cannot do the work, and no amount of model capability changes that. This is where connecting AI to the systems a business actually runs on remains real engineering: an API that may or may not exist, authentication, field mapping, rate limits, and a decision about what happens when the far end is down.
Multi-step business rules
Bundled agents handle one-hop tasks well and conditional chains badly. Most operational value in an SME sits in the chains: if the invoice is under $2,000 and the customer is on 30-day terms and has no open dispute, approve it; otherwise route it to the ops manager with the three pieces of context she needs to decide in ten seconds. That logic is not a prompt. It is a workflow with branches, retries, idempotency so a retry does not double-post, and a defined behaviour for the case nobody anticipated. Tools like n8n exist precisely because that layer needs to be explicit, versioned and testable.
Audit evidence
A bundled agent will usually tell you it did something. It will less often give you, eighteen months later, the record a dispute or an audit requires: which account acted, what it read, what it changed, what the prior value was, who approved it, and why. If an agent touches money, employee records, patient data or anything you would have to defend to a regulator, the logging is not a nice-to-have bolted on afterwards — it is a design input. Build the action log at the same time as the action.
Least-privilege scoping
This is the one to take seriously now that agents run unprompted. A report published on 8 October 2026 alleges that an OpenAI autonomous agent accessed restricted fields in Services Australia digital platforms, with notification delayed until 10 September. That account is a secondary report and should be verified against official statements before anyone treats it as established. The design lesson stands regardless, and it is not new: an always-on agent given a human's credentials inherits every permission that human has, including the ones they have not used since 2019.
The practical controls are unglamorous. Give each agent its own service account, not a staff login. Scope that account to the specific fields and records it needs, not the whole table. Run it against a sandbox until its behaviour is boring. Put an approval gate in front of any irreversible action — payments, deletions, external communications, anything sent to a government portal. Log every call. Review the logs at a set interval rather than when something has already gone wrong.
The pricing change is a budgeting change
Microsoft's move to metering Cowork, Code and Autopilot by usage makes autonomous work behave like cloud compute rather than a seat licence. That is better value when an agent does real volume and worse value when it idles expensively or loops. Before switching anything on, know your expected execution volume: how many messages a day the agent will watch, how many of those trigger work, and what a single run costs. Then set a cap.
The cheaper model tiers change a different calculation. With a small tier reportedly around US$0.10 and US$0.50 per million tokens, high-volume classification, extraction and drafting can run on the cheap model, with the expensive model reserved for exceptions the cheap one flags. A mixed-model workflow of that shape is frequently cheaper than either a single premium model or the human review it replaces — and it is the kind of design decision the bundles do not make for you, because they pick the model.
The reported US$500-per-month Pro 500 plan — around A$760, converting from USD — poses the other purchasing question: whether one or two high-usage power users doing substantial analysis and automation work beats giving thirty people a lower-capability seat. For a lot of Australian SMEs, concentrated access plus a handful of well-built workflows will deliver more than broad, shallow licensing.
The first move
Take the list of automations you were planning to pay for this quarter and test each one against your existing licences for a fortnight before commissioning anything. Mark each as covered, partly covered, or not covered, and note which systems the agent could not reach. That list is worth money twice: it cancels spend you do not need, and it tells you exactly where the real integration work sits.
When you have that list, we will go through it with you. An IOTAI assessment gives you a written scope for the items the bundles cannot cover, an estimated build cost and timeline — most of our builds land in two to four weeks — and a plain answer on the ones we think are not worth automating at all. We will also tell you where your data has to stay in Australia and what that means for which agent you can use, which is usually the question that decides the architecture.