Most mid-market businesses we sit with have not deployed an AI teammate. They have Copilot because it came with Microsoft. Or ChatGPT or Claude on a personal login. Sometimes someone has tried Claude Cowork. That is usually a staff member playing, not an enterprise rollout.

Someone uses it to draft emails and summarise documents. That is useful. It is also a long way from what vendors are now selling.

In August 2026 xAI released Grok Bot and described it as an AI teammate. The product has its own computer, holds its own logins, and keeps working after you shut the laptop (xAI, Introducing Grok Bot, August 2026). Other vendors will use different words for the same capability: AI staff, digital colleague, always-on agent. The name is less important than the change in what the software can touch.

A weak draft in ChatGPT or Claude is annoying. Give that same class of software access to CRM, email or job files and it can change a live record, send something outside the firm, or leave a mistake sitting there overnight.

Australian mid-market firms are already stretched. Leadership teams are thin. Operations still run on spreadsheets and what people remember. Salespeople spend a lot of the week on admin. Finance often takes too long to close the month. The few that take a teammate demo are then tempted to buy a seat and hope it soaks up the overflow.

That is the same sequence we already see in failed CRM and digital projects: the system is purchased first, and the work is defined later, if it is defined at all.

In our own firm we have run persistent agents on live work. The useful lesson was not which product we started on. It was that reliability, permissions and a written job matter more than the launch. The businesses getting something out of this will pick a small number of repeatable tasks, limit what the agent can touch, and keep a person on anything that can move money, clients or reputation.

What the data actually shows

Vendor surveys talk as if boards have already put agents into the budget. That is not what we see in mid-market rooms.

In PwC's May 2025 AI Agent Survey of 300 senior executives, 88% said their team planned to increase AI-related budgets in the next 12 months because of agentic AI. 79% said AI agents were already being adopted in their companies. Of those adopters, 66% reported measurable productivity gains, 57% reported cost savings, 55% reported faster decisions and 54% reported a better customer experience (PwC AI Agent Survey, May 2025).

The same survey is less impressive once you get past the adoption headline. 68% of those companies said half or fewer of their employees actually interact with agents in everyday work. Only 17% said agents were fully adopted across almost all workflows. PwC notes that reports of full adoption often reflect excitement about what agents could do, rather than evidence that the work has actually changed.

Vendors will quote the 66% productivity number. In a $15m or $40m Australian business, the 68% is closer to what we see. Copilot is often already on because Microsoft put it there. ChatGPT or Claude is usually a personal seat, without a written policy. A handful of people use it well. Most use it for drafts and summaries. A lot of the professional firms we sit with started on ChatGPT and later moved to Claude because they preferred the quality of the writing. Switching models does not give you an operating model.

Then a demo arrives with an AI teammate that can log into the same systems the office manager uses, and the conversation jumps from writing help to unsupervised staff.

	Copilot, ChatGPT or Claude	An AI teammate
What it is	A chat window you open	A persistent worker you assign
When it works	While you are in the conversation	After you close the laptop
What it can touch	The text you paste in	Email, CRM, files, browsers, internal tools, if you grant access
Typical failure	A bad paragraph or a wrong summary	The wrong update, in the live system, at 11pm
Who should own it	The person using it	Someone who can revoke access
Sensible first use	Drafting, research, rewriting	One narrow, repeatable job with a human check

We wrote about the chat-window version when ChatGPT Agent landed: ChatGPT Agent: The New AI Colleague Your Board Will Soon Demand. An AI teammate does not wait to be asked. xAI is fairly direct about that. Grok Bot is sold as something you can give real work to, on a shared cloud computer, with its own logins, still running when you are not there.

Other products do a similar job on a server you control. There is a genuine debate about which setup is easier to supervise. That debate should come after you have decided whether a particular piece of work is ready to be handed over.

Why this is going to cost people money

For the last two years the default mid-market AI programme has been Copilot licences, a lunch-and-learn, a note telling people to try ChatGPT or Claude, and an AI working group. Then everyone waits for productivity to show up in the monthly pack.

That was already a weak way to spend money. It gets worse once the tool can act inside the business.

The seat gets bought before the job is written

This is the same pattern we keep seeing on CRM and transformation work. The platform is chosen first. The work is mapped later, if anyone maps it at all.

We have written about both: why most CRM projects fail, and why mid-market digital transformation burns millions. Calling the product a teammate does not change the sequence.

We keep seeing a middle step that looks like progress and is not. A few people already have ChatGPT or Claude on personal seats. When you ask what they would use an official tool for, the first answers are emails and first drafts. Leadership wants a proper plan, but there is still no shared view of what the tools actually are. Someone is asked to put a briefing together so the rest of the room can catch up. A live priority takes the calendar. A couple of months later the official conversation is still about which model to buy, and nothing has been deployed.

If nobody can describe the job in a paragraph, including the inputs, the systems, what done looks like, and who checks the output, you are not hiring help. You are buying another demo.

Reliability gets treated as a software preference

In our own firm we have run persistent agents in live work. The first generation could look good for an afternoon and then fail in ordinary ways. It stalled. It lost context. Sometimes it did something almost right, which is harder to catch than an obvious error.

We changed runtime because we needed something we could trust with real work, not because we wanted to be on the newest stack. If an AI teammate is going to sit near client work, it has to be dependable enough to be boring. A clever agent that needs a babysitter is not cheaper than the junior you already employ.

The keys get handed over on day one

The new products ask for logins. Email, Drive, CRM, browser. That is how they do the work.

It is also how a mid-market business can do real damage. Client lists, pricing, unpublished financials and supplier terms often live in the same places an office manager already has open. That inbox is not a test environment.

The first permission set should be tighter than the vendor will suggest. Read access before write access. One system before five. A separate mailbox before the MD's.

Messy work gets sped up rather than cleaned up

An agent will run a messy process faster. If quoting still lives in three spreadsheets and someone's head, the teammate will scale the confusion. If month-end depends on one person remembering the exceptions, the agent will miss those exceptions at 2am and still produce a file that looks finished.

The work does not need to be perfect before you delegate it. It does need to be clear enough that a competent stranger could follow it.

The stack gets changed every quarter

Open-source runtimes, vendor clouds, in-app agents inside Microsoft and Salesforce, and a new AI staff product every month. The category is moving quickly.

If you chase it, you will spend the rest of 2026 installing tools and the first half of 2027 explaining why nothing actually shipped. Pick something you can supervise, stay on it long enough to finish a job, and then judge it.

What we are seeing work

We are not seeing a neat AI programme land across mid-market businesses. We are seeing a smaller group treat the teammate the way they would treat a new hire who is fast, cheap and not yet trusted.

The job is written before the seat is bought

The work that holds up is usually unglamorous. Someone already reviews a weekly pack of jobs, invoices or sales activity, and the agent can assemble the first version. Follow-ups already sit in the CRM, and the agent can draft them in the firm's language. A leadership pack is built from sources that already exist. A recurring report needs a first draft, then a person. A mailbox or folder needs exceptions flagged, not resolved.

If you cannot describe the job without a slide deck, it is the wrong first job.

A person still sits on anything with a consequence

Anything that can move money, change a client record, go outside the firm, or commit the business to a date should stay behind a person. Let the agent prepare the work, then have someone release it. That is slower than the demo, and it is how you avoid spending the next quarter unwinding a confident mistake.

Each job has one number

Use one number you already understand, not a vague productivity claim. Hours from the start of month-end to a usable pack. Number of stale CRM records. Time a BD person spends on first-draft follow-up. Error rate on the weekly ops summary.

If that number has not moved in 30 days, the teammate is not working on that job. Change the job before you change the slogan.

Access looks like contractor access

A contractor does not get the MD's password. An AI teammate should not either.

Separate credentials, the least access required, a named owner, and a way to shut it off on a Friday night without calling the vendor. In a mid-market firm that owner is usually the GM, the ops lead or the FD. It should not be the person who enjoyed the demo.

Nobody asks the agent to fix the operating model

An AI teammate can take work off a capable team. It cannot give you a management structure, a decision right, or a clean set of numbers. If the business is already straining under growth, key-person risk or a reporting lag, the agent sits on top of that strain.

We have already made a version of this point for fund managers, where the regulatory setting is tighter: AI Agents for Fund Managers. In a mid-market operating company the version is simpler. If the process is not written down, do not automate it. If it is written down and still depends on one person, deal with that first.

The sequence we keep coming back to looks like this.

1. List the weekly work that is structured, repeatable, and currently done by a relatively senior person because nobody else has time.
2. Remove anything that changes a live client, cash, or an external commitment.
3. Write a one-page job for what remains: source, steps, what done looks like, who checks it.
4. Give the teammate access to only what that page requires.
5. Run it for 30 days next to the current process, not instead of it.
6. Keep it, narrow it, or kill it. Do not expand it because the vendor shipped a new feature.

What this means over the next 18 months

The category is not a fad. PwC's respondents already think agents will reshape work more than the internet did. Whether that forecast is right matters less than the buying behaviour it will create. Every software vendor will ship an agent. Every MD will be asked in a board pack what the firm is doing about it.

Some firms will spend 2026 buying three products, finishing no jobs, and then deciding that AI does not work here. More will stay on Copilot and a personal Claude login and call that a programme.

The businesses that get something useful will look a bit behind from the outside. They will still be on one runtime. They will have two or three jobs that actually ship. Their people will trust the output enough to stop redoing it from scratch. That is a commercial advantage, even if it does not photograph well.

Two things are worth being clear about.

This is an operating decision, not an IT purchase. The GM or MD has to own the job list. If it sits with "the person who likes AI", you will get experiments rather than capacity.

And the cheap lesson is available now. While the teammate is new, you can keep it in a small box. Once it is in email, CRM and finance, taking it back is a political event.

You can hire an AI teammate this week. Several vendors will be happy to sell you one. The work that makes it useful is older than the product: know the job, decide what a new hire is allowed to touch, check the output, and keep the keys.

Questions operators are actually asking

What is an AI teammate?

A persistent AI agent assigned to a job. It can use a computer, log into tools, and keep working when you are not in the chat. Vendors also call this AI staff, a digital colleague, or an always-on agent.

How is that different from Copilot, ChatGPT or Claude?

Copilot, ChatGPT and Claude wait for you. An AI teammate is set up to act. A bad answer in a chat window is usually a bad paragraph. A bad teammate answer can be a bad action in a live system.

Should a mid-market business deploy AI agents in 2026?

On a short list of boring, repeatable jobs, with a human check, yes. As a general-purpose staff replacement, no.

Does Grok Bot change the decision?

It makes the product category obvious. It does not decide the job. A vendor cloud computer, a self-hosted runtime, or an agent inside Microsoft or Salesforce can all cause trouble if the work is unclear.

What should we automate first?

Work that is already written down, happens every week, and currently consumes a senior person's time without needing their judgement. Reporting packs, follow-up chases, exception flags. Not pricing, client commitments, or anything you would not give a contractor in week one.

What is the main risk?

Access. Once the teammate has logins, treat it like staff.

Do we need a big transformation programme first?

No. You need one written job and one owner. A transformation programme is how this waits until the next budget cycle. Firms already spending heavily on transformation should look at the sequence first before they put an agent on top.