Anthropic published a write-up on 28 August 2026 describing how its own staff use Claude Tag, the Slack integration where you mention Claude in a thread the way you would mention a colleague. It is a more useful document than most vendor case studies, because it names the people, gives the actual prompts, and includes the bits that did not work.
For a small Australian team the value is not the scale. Anthropic has thousands of Slack channels and a product engineering culture most mid-market businesses do not have. The value is that three of the patterns described are reproducible by two people and a Slack workspace, and the setup mechanics behind them are the part worth copying.
Pattern one: a messy thread becomes a review-ready document
A product marketer took a Slack thread of more than fifteen messages, with several people adding asks and some disagreement about what was actually needed, and turned it into a customer-facing one-pager in about 45 minutes. The first prompt was close to trivial: go through this thread and produce the one-pager being asked for. A draft came back in roughly two minutes.
The interesting step was the second prompt, which asked Claude whether everything in the document was factual and correct. Claude sorted its own claims into ones it could verify against public documentation and ones that were its own framing, and flagged the second group for a product lead to sign off. The marketer then supplied two approved source documents and Claude rewrote a section to match the sanctioned wording.
That is the transferable move. Asking a model to separate what it verified from what it constructed turns an unreviewable block of text into a short list of things a human actually has to check.
Pattern two: consolidating requests scattered across months
An operations staffer needed every customer request for a feature that was about to go generally available, spread across Slack channels and a product feedback hub, going back months. She gave Claude the search targets, a one-sentence definition of what counted as a match, and a pasted example of the output format from a previous launch.
Claude ran roughly twenty search variants and returned about twenty-four accounts in around 26 minutes, one line per requester with handle, team, account and a link to the original message. A larger job followed: read every incident, escalation, support and product-feedback channel for the previous week and report what customers had hit. Roughly fifty minutes later it posted 23 open and 14 resolved issues condensed from about 120 raw findings, organised by product area with a source link on each.
Then she asked Claude to check its own work, and it surfaced fifteen more issues it had missed the first time. Worth sitting with that. A first pass that misses fifteen items is still enormously useful, and it is also not complete, which is exactly the posture to adopt with this kind of work.
Pattern three: a review queue that runs in a channel
Anthropic's legal team reviews marketing material before public release, and in launch weeks that queue backs up badly. A product counsel with no engineering background created a dedicated Slack channel where Claude examines every asset first against rules she wrote herself. Reported turnaround went from a day or longer to about 30 minutes per asset.
Claude flags issues such as unsubstantiated marketing claims, checks factual statements against the company's internal knowledge index and the public web, and works directly with the person who submitted the asset to fix what it found. Anything still needing a legal decision gets tagged to the right counsel with the specific statement attached.
The structural point is that the reviewer defined the rules, not an engineer. That is the difference between an automation that survives and one that stops being maintained the moment its author moves teams.
What can a small Australian team actually copy from this?
All three patterns, at a smaller scale, without writing code. Each one is a Slack channel, a set of written standing instructions, and a scoped set of connectors. A Brisbane agency of twelve people can run a marketing review channel, a client-request consolidation job and a thread-to-document workflow with the same setup Anthropic describes, because none of the three depends on custom engineering. The constraint is not the size of your team, it is whether someone writes the instructions down properly.
The table maps each pattern to a mid-market equivalent and an honest read on effort.
| Anthropic pattern | Your equivalent | Setup effort | Reported time saved |
|---|---|---|---|
| Thread to customer collateral | Client thread to scope note or proposal draft | Under an hour, no config | Hours per document |
| Consolidating scattered requests | Every client ask for a feature or fix, across channels | Half a day to define a match | Around a week of manual work |
| Legal review channel | Any gated queue with written rules: quotes, copy, compliance | 1 to 2 days of rule writing | Day or more down to about 30 minutes per item |
The setup mechanics that make the difference
Reading across all three, the same handful of habits show up. These are the parts most teams skip and then wonder why the results are vague:
Give the full instruction up front. Where to search, what counts as a match, and a pasted example of the output format you want. The vague version of the same request produces a vague answer and a second round trip.
Keep a private channel with Claude and use separate threads per request. Both staff profiled work this way, which keeps context clean and stops one long thread from poisoning the next task.
Ask it to check its own work as a standing second step. It surfaced fifteen extra issues in one case and separated verified claims from its own framing in another.
Scope access deliberately and expect to be told when it is missing. Claude works only from the channels and documents it has been granted, and says when it cannot reach something. Treat a blocked source as a gap in the answer, not a solved problem.
Let the person who owns the process write the rules. The legal review channel works because a lawyer wrote the criteria, not because an engineer built a workflow.
What this costs and where it pays
Claude Tag is in beta for Claude Enterprise and Team customers in Slack, with Microsoft Teams flagged as coming. So the real question for most Australian businesses is not the integration, it is whether you are already on a plan that includes it.
The consolidation job is where we would start, because it is the easiest to value honestly. If a senior person would have spent a week compiling a picture of client issues and does not, that is roughly $6,000 of loaded time on a $150,000 package, once. The more interesting version is the work that simply never got done because nobody had the week to spare. Our ROI calculator handles both cases, and will tell you when the answer is that it is not worth it yet.
Two honest caveats
Anthropic's figures are Anthropic's own, drawn from a company that builds the product and staffs itself accordingly. The 45 minutes and the 26 minutes are real reported numbers from named employees, and they are also best cases from people with unusually good instincts about how to ask. Expect your first attempts to be slower.
The second caveat is about completeness. In one example the product feedback hub blocked direct access and Claude worked around it through cross-references in Slack. That is resourceful and it is also a silent partial answer. If you put this on anything where missing an item has consequences, name the sources it could not reach in the output and have a human close the gap.
You can read Anthropic's own write-up of how its employees use Claude Tag for the full prompts. We have also covered the two structural changes that make these patterns possible: Claude Tag reading a whole channel rather than a single message, and the agent identity and access model that governs what it can reach.
If you want help deciding which of your own queues is the right first one to put in a channel, book a time with us. We will pick the one with the clearest rules, because that is the one that works.
