AI with real value: three tests before you sign anything.

By our count, seven separate AI vendors have knocked on the door of most Australian mid-market operators this year. They all pitched the same demo. A synthetic voice answered a call. A CRM record appeared. Someone in a sales deck circled a number that read "70% productivity gain". Everyone shook hands and nobody bought anything.

The gap between the demo and the deployment is the whole story. And the reason it keeps happening is that most AI is being sold on features, and most operators are trying to buy outcomes. Those are different conversations, and the second one is harder to have.

Here is how we would run the conversation if we were buying, not selling.

Test one: does it ship work, or just answer the phone?

The easiest AI to buy is the one that picks up a call and holds a conversation. The reason it is easy to buy is that it is already a solved problem. Voice models have been good enough for eighteen months. What most operators discover after the first invoice is that the AI answered the call and did nothing after it.

The value is not in the call. The value is in the next twelve steps. Did it update the CRM. Did it generate the invoice. Did it send the confirmation text. Did it flag the exception to a human. Did it write the outcome back to the file so the next person who picks up the account is not starting from scratch.

A voice agent that answers a call and does not follow through is a fancy voicemail. Ask any AI vendor for a live end-to-end deployment where the AI closed the loop. Not the call. The loop.

Test two: does it fit the system you already run?

The second-easiest AI to buy is the one that comes with its own dashboard. You get a login, you get a report, and you get a second inbox to check every morning. That is not automation. That is a tax on your operations team's attention.

Real value writes back into the system your team already opens. If your firm runs Smokeball, the matter should update inside Smokeball. If your clinic runs Cliniko or Nookal, the appointment and the fund verification should land inside Cliniko. If your broker desk runs Insight or WinBEAT, the renewal answer should log against the risk. If your workshop runs MYOB Advanced or NetSuite, the PO confirmation should hit the ERP.

Ask the vendor for a bi-directional integration screenshot on your specific stack. Not a webhook diagram. An actual write, from a live call, into your live system. If they cannot show you one, you are the pilot.

Test three: does it pay for itself before it does anything clever?

The right cost frame for an operator is not "how much does AI cost". It is "what am I paying for the coordination this AI would replace". In practice, that is one to three FTEs at the front office of most mid-market businesses. Somewhere between $80,000 and $250,000 a year in wages, before you count the revenue you lose to slow turnaround or dropped follow-up.

An AI that costs less than that and does the same coordination pays for itself before it does anything clever. If a vendor cannot draw the sum on a napkin against your current headcount, they are asking you to buy on faith.

The clearest way to run this test is to insist on a fixed monthly commercial per deployed worker. Per-minute pricing sounds cheap and ends up expensive. Per-seat pricing pretends AI is a tool your team uses when the point is AI doing work your team does not. Fixed monthly per worker, with a written scope, lets you calculate the return in the meeting, not six months later.

What real value looks like when you find it

The AI deployments that show up on a P&L six months later share four things.

One, they are scoped to a specific job. Matter intake. Fund verification. Renewal chase. Not "AI transformation". A defined job with a defined output.

Two, they were co-designed with the practitioners who used to do the job. Not with the CTO, not with the vendor, and not with the McKinsey deck. The person on the phone shaped the workflow.

Three, they were piloted alongside the existing manual process for a month before anything was switched off. You watched the numbers before you trusted the machine.

Four, they wrote back into the systems the business already ran, and the vendor took responsibility for the integration. No swivel chair, no shadow dashboard, no second inbox.

Heya is built for that shape of deployment. The Cordi engine sits on top of the systems you already run and closes the coordination loop end to end. Each digital worker has a written scope, a fixed monthly commercial, and a clear escalation to a human. The pilot lasts four weeks before anything gets switched off. The value shows up before anything clever does.

The three tests above are how we sell against ourselves. If you are talking to us and to another vendor, ask us the same three questions. If our answers do not hold up, do not sign.

---