Where the return comes from

Every investment committee eventually asks the same thing of an AI programme: show me the return. Most programmes answer with anecdotes, hours "saved" and a productivity survey. None of these survive a finance director. The return has to be found where money actually moves.

Three places the money is

1. The run rate. What the company pays to run its software, including the people who administer, maintain and feed it. This is the return that shows up on its own, in the next quarter's accounts.

2. Capacity. The hours agents take off your people: logging calls, drafting follow-ups, sorting tickets, chasing invoices. Real, but not yet cash. It becomes cash only when the hours go somewhere: more customers served, more deals worked, a hire you no longer need to make.

3. Speed. Follow-ups sent the same day, tickets answered in minutes, a forecast that is current every morning. Speed shows up later and indirectly, in win rates, retention and cash collected sooner.

The run rate is the first proof

The fastest return is the cheapest to measure. When Kintoro moved its own company onto piMonk, our software run rate fell 83 per cent in three months, counting admin and maintenance staff time, not only licences.¹ Our cost per CRM seat fell by more than three-quarters.¹

At list prices, piRevenue costs 89 per cent less per sales seat than Salesforce with its AI add-on.² The point is not the discount. It is that software priced for high-wage markets forces companies elsewhere to ration seats, and rationed software never becomes the way work is done.

Hours saved are a promise. Run rate is a fact. Lead with the fact.

Measuring capacity honestly

Capacity is where most AI cases are inflated. A fair measure has three parts, agreed before the start:

The job. One named task, such as logging a sales call or answering a routine question.

The baseline. How many of those tasks the team does today, and how long each takes.

The redeployment. Where the freed hours will go. If nobody decides, they disappear into the working day and the return with them.

The resulting equation is simple. Illustrative, for any one job: hours returned per month, multiplied by the loaded cost of an hour, less what the software costs to run. Anything that cannot be written this way is a hope, not a return.

What the machines cost to run

Agent work has a running cost, and it varies more than most buyers realise. By Kintoro's estimates, 1,000 agent actions cost about $1 on a small AI model trained for the job, about $2.30 on a mix of mostly small models with some large general ones, and about $13 on large general models alone.³ The choice of model is largely the choice of margin. Those trained small models are still being built; today's apps use a mix.

The case against the case

Sceptics say returns from AI are always two years away, and that savings get quietly absorbed. They are often right. That is why the run rate should carry the first case, why capacity should be counted only when it is redeployed, and why every programme needs a stop date. A return that cannot be stopped at day 30 is a subscription to a hope.

What this means for investors

Operating leverage. A company whose agents carry the routine work can grow revenue without headcount rising in step. That is the prize in a value creation plan, and it compounds across a portfolio that adopts the same way.

A repeatable playbook. One team, one app, 30 days, measured the same way in every portfolio company. The second company is faster than the first, because the approval rules and the baseline method carry over.

Downside you can see. Usage off by default, ceilings the company sets, data returned on request. The worst case is known in advance.

The return from applied AI is not a mystery. It is a run rate you can read, capacity you choose to redeploy, and speed that shows up in time. The map for taking it one function at a time is The applied-AI grid.