Zenith Consulting Services

Articles

Where AI agents work in your business, and where they don't

James Pepplinkhouse / / 8 min read

Advice from my real world experience

Everyone is telling you to use AI. Hardly anyone has enough experience to tell you how to embed it in your business. I run AI agents every day in my own software businesses. I started those businesses specifically to build on this new wave of tooling. I have watched the agents do excellent work, and I have watched them do the wrong thing with total confidence. This is what I have learned, written for a business owner rather than an engineer.

Unattended AI is slow and expensive - use it sparingly

AI is powerful, but it is slow and it costs money every time it runs. An ordinary program, one that does the same thing the same way every time, will always beat it on speed, quality and reliability. I keep AI to a minimum and use it for tasks that need judgement, content creation, or interpretation. Everything before and after those tasks should be plain, deterministic software.

Many of the quick wins we build contain little or no AI. They are point solutions: staff click a button, or an automated trigger fires, and the work happens. In most cases this is superior to slow, costly, non-deterministic AI outputs. Where a task needs some reasoning, we add an AI for that task only. That rule is for AI running on its own. The next section is the exception.

Attended AI is the opposite - use it everywhere

Sparingly applies to AI inside your systems, where it runs unattended. The opposite is true when a person is driving it. The biggest use of AI in my businesses is not in any product. It is Claude Code on my desk, with agents, skills and tools I have built up over months, doing ad hoc work all day: reproducing and diagnosing a support issue, testing a UI change in the browser, writing development tickets and then implementing them, drafting a LinkedIn post, researching a prospect before a meeting, or building a throwaway tool to migrate a client's messy historical data into our platform. A person asks, watches, corrects and decides. That loop is fast and cheap because the human is the safety net, and it is where most of the value sits today.

The difference is who is in the loop. Unattended AI needs to be rare and well fenced. Attended AI can be everywhere.

Start with the human loops

The best place to begin is not the cleverest idea. It is the most repeated one. Become critically aware of the loops in your business: the things your people do every day, several times a day, in the same shape. The Monday report. The re-keying between two systems. The quote assembled from the same five sources every time.

Your team is good at these jobs. That is exactly why they are worth automating: the work is well understood, the rules are known, and a small tool can take the mechanical part and leave your people the high-value parts that need a human.

The Australian Bureau of Statistics found that only around 12% of Australian businesses used AI in 2024-25, and about 11% of small and micro businesses (ABS, June 2026). For all the noise, almost nobody has started. That is an opportunity: a few small, targeted tools put you ahead of most of your market, without a 24-month transformation programme.

Map each workflow, then grade each task

Do this exercise this week. Take one workflow and write out its tasks. Paper is fine. Better still, ask an AI to interview you and surface how things actually get done. For each task, mark one of three things: AI can do this today, AI cannot do this yet, or AI should never do this.

The "not yet" column is the interesting one. Check it again next month. The pace of improvement is staggering, and a task that was out of reach in March is often routine by June. The "never" column is yours to set. You decide what AI can and cannot do for your business, not the vendor.

Once the map exists, you know where to start and which tasks need a person to sign off.

Give it knowledge, or it will guess

The biggest factor in whether an AI agent does a good job is the knowledge and context you give it. Not the model. Not the prompt. The information.

In my business every process, decision and rule lives in a shared set of written notes that the agents read before they act. We use Obsidian, but the tool does not matter. What matters is that the information is living and current. Stale documentation produces stale decisions, and the agent will never tell you it is working from old information.

MIT's NANDA initiative reported in 2025 that about 95% of enterprise AI pilots delivered no measurable impact on profit. The lead researcher's diagnosis was not model quality but a learning gap: generic tools do not learn from or adapt to the way a particular business works (MIT NANDA, The GenAI Divide, reported by Fortune, August 2025). Give the agent your context and you are already ahead of most of that 95%.

AI can do more than you think. When a task seems out of reach, it is usually a case of adding the right knowledge, skill or tool, not waiting for a smarter model.

A person signs off on anything that leaves the building

This is the one rule I will not bend. Human approval gates go wherever the consequence of a mistake is real: anything sent to a customer, anything that spends money, anything that changes a record you cannot easily change back.

In my own engineering team, AI workers write code, run the tests, review the work and present the result for approval. They never put it into the live product. A person does that, every time. If a worker hits a problem it cannot verify, it stops and writes up exactly what it needs a human to decide, rather than guessing. Agents do the work; a human owns the decision. That is how you get the speed without the risk.

Gartner predicted in June 2025 that over 40% of agentic AI projects will be cancelled by the end of 2027, citing escalating costs, unclear business value and inadequate risk controls. All three are avoidable with small scope, measured loops and approval gates.

The pattern that is working now: a chief of staff

We arrived at our current setup in three steps, and each one was a bigger lift than the last.

First, we prompted individual agents one at a time. Useful, but every task needed a person to start it, interactively prompt, watch it and pick up the result. That caps how many agents one person can run at around five to ten. Beyond that the context switching is really exhausting.

Second, we gave the agents a Kanban board. Tickets, columns, a claim on each card. The human and AI team could now collaborate on the same board: we write the tickets in collaboration with an agent, other agents pick them up and raise a pull request, and a human engineer reviews and merges it. That merge is the gate: nothing an agent writes reaches production without a person signing off. That was a big uplift, because the work became visible and the hand-offs became explicit. This is a bigger step change than it sounds. As long as there are enough tickets on the board, we can run two to four agents on each engineer's machine, and they will keep going day and night if you let them. As a business owner that is a massive unlock: I take a meeting or a phone call and come back to find a pile of work finished.

Third, we put a manager on top. Now I talk to a single "chief of staff" agent. I say what I want. It breaks the work into pieces, hands them to several sub-agents that run at the same time, tracks what is in progress, and tells me what is waiting on my decision and what is on my to-do list. That was an even bigger uplift than the Kanban board (honestly, it feels like 100x).

The part that matters for a business owner: I can talk to it day or night (even from my phone) and it responds, but it does not run off and spend money on things I did not ask for. We deliberately do not run always-on, act-on-their-own agents. They are not needed, and an agent with nothing to do and a budget to spend is a liability, not an asset.

That is the shape to aim for: one assistant you speak to plainly, with a team of specialists underneath it that you never manage directly, and that only works when you give it work (either by asking, or by seeing that there are tickets to action on the board).

The longer journey

Point solutions get you efficiency now, and I strongly recommend you start enjoying those quick wins today! The step change described above is harder. For that, AI has to sit at the centre of the business, with the company's knowledge, skills and tools available to it, as a member of the team rather than an add-on. Most businesses will take a long time to restructure that way, and that is fine.

This is not about replacing people. It is about multiplying the expertise you already have so a small team can compete with a much larger one.

Two notes on the people side. In my experience a small percentage of staff "get" AI almost immediately; the rest take months to become comfortable and to change how they think about their own work. Give them that time. You must come to understand what each model can do before you can look at a task and recognise it as one AI could handle. That understanding is the real skill, and it is learnable.

Trial tools and be prepared to swap them. The state of the art in how to use these tools changes at least monthly, and loyalty to last year's choice is expensive.

Where to start

Pick one loop. Map its steps. Grade each one. Build the smallest tool that removes the mechanical part, with a person approving anything that matters. Then do the next loop.

If you would like help finding the first one, or building it, that is the work we do at Zenith Consulting Services. Tell us what you are thinking about, even if it is rough, and we will tell you honestly what AI can do for it today.

Sources