What it actually costs to run a company on AI agents

Most writing about AI agents comes from people who sell them but don’t live with them. The costs quoted are either fantasy (“replace your whole team for $99/month”) or fog (“it depends, book a call”).

We can do better than that, because we run one. nollie is a SaaS CRM for hospitality (restaurants, cafes, bars), live in four markets (Australia, New Zealand, Singapore, the UK), built and operated by a solo founder. The day-to-day operating layer is a fleet of AI agents: finance close, customer onboarding, support triage, marketing production, error monitoring. Humans supervise; agents do the work.

Here is what that actually costs, by category, with the failures included. Treat our numbers as one honest data point, not a universal price list. Your workflows, volumes and risk tolerance will move every figure.

Category 1: Model API spend

The headline most people want first: our model API bill runs to a few hundred Singapore dollars a month. It moves month to month (heavier in months with big onboarding pushes or one-off migrations, lighter in steady-state months), but it has never been the number that matters.

For external calibration: Sandlabs, a Melbourne consultancy that builds agents for clients and is one of the few firms anywhere publishing run-costs, quotes A$200–600 per month to operate a deployed agent. That range is consistent with our lived experience across a fleet doing real operational work.

Two honest caveats. First, costs scale with carelessness faster than with volume: a poorly designed agent that re-reads its entire context on every task can cost ten times a well-designed one doing the same job. Second, model pricing has been falling for years, but don’t build a business case that only works if it keeps falling.

Category 2: Tooling and SaaS

Agents don’t run in a vacuum. Ours need the same subscriptions a small ops team would: task management, a knowledge base, comms, email infrastructure, monitoring, hosting for the orchestration layer. Call it a few hundred dollars a month again, much of which we’d be paying anyway to run the company with humans.

The honest line item people miss is the orchestration glue: schedulers, runbooks-as-code, logging. None of it is expensive in dollars. All of it is expensive in the next category.

Category 3: Human supervision time

This is the cost nobody puts on the pricing page, and it’s the one that decides whether your agent operation is leverage or theatre.

Our standing rule is that no agent takes an unsupervised external action. Nothing reaches a customer, a supplier, a bank or a regulator without a human approving it. That rule costs real time: reviewing drafted customer communications, checking the finance close before anything is filed, scanning triage decisions, reading the overnight run reports.

In steady state, that supervision runs to roughly an hour or two a day: more when something new is being brought into the fleet, less in quiet weeks. It is not optional, and anyone who tells you supervision rounds to zero is selling something. The supervision time is precisely why the fleet works: review is where errors get caught, and caught errors are how runbooks improve.

Category 4: Build time, amortised

Every agent in the fleet cost real hours before it earned any. A workflow has to be documented (often for the first time: agents are ruthless at exposing processes that only existed in someone’s head), the agent built, tested, run alongside the manual process, and given a runbook covering what “done” means and what to do when it fails.

A simple workflow takes days. A gnarly one (our finance close took several rebuilds) takes weeks, spread across months of iteration. Amortised over the agent’s working life the cost per month is small, but if you’re budgeting an agent programme, the build hours will dwarf the API bill in year one. Plan for that, or you’ll mistake the project for failing when it’s merely in its expensive phase.

What broke

A transparency piece without a failure list is a brochure, so:

• Onboarding stalled for six weeks after launch, not because the agent erred, but because the surrounding workflow still assumed a human and had no written completion criteria. The fix was process redesign, not a better model.
• A marketing production agent confidently used stale brand assets for a customer who’d rebranded. Caught at human review (which is the system working), but it cost a rework cycle and earned a new runbook step: verify source material is current before anything is produced from it.
• Silent failures. Early on, scheduled jobs would die quietly and we’d discover it days later. The fix was boring: every run reports, including the successful ones, and a missing report is itself an alert.
• Cost spikes from sloppy design. One early agent re-ingested an entire data set on every run. The bill flagged it before we did. Now token spend is reviewed monthly like any other supplier.

None of these were existential. All of them were the kind of thing that kills agent programmes in companies where the first failure becomes the reason to stop rather than the next line in the runbook.

The comparison that actually matters

Add it up: model spend plus tooling in the mid-hundreds of SGD per month, an hour or two a day of supervision, and a build investment measured in founder-weeks, amortising down over time.

What does the fleet replace? Honestly assessed: the operational work of roughly three to four full-time hires: an ops/admin function, a finance/bookkeeping function, a support triage layer, and a chunk of a marketing producer. In Singapore, those roles fully loaded would run somewhere in the S$15,000–25,000+ per month range depending on seniority and market.

Against that, the fleet’s cash cost is a rounding error and the real cost (supervision plus build) is a fraction of one salary. That is the leverage, plainly stated.

And the limits, equally plainly: agents do not give you judgement, taste, relationships, or accountability. The supervision hour exists because the fleet still makes mistakes a junior employee wouldn’t, and avoids mistakes a junior employee would. It substitutes for headcount in well-defined operational work. It does not substitute for the person who decides what the work should be.

We’d also note what made our case favourable: a solo technical founder, a software business with digital workflows, and a high tolerance for iterating on our own operations. A 40-person business with paper processes will spend more on the build and the redesign than we did. The economics still tend to work, but through fewer future hires and reclaimed hours, not through firing anyone on day one.

Do this next week

Don’t start by buying agents. Start by costing one workflow.

Pick a single recurring process: invoice handling, customer onboarding, weekly reporting. Write down: hours per week it consumes, the hourly cost of whoever does it, and the steps as they actually happen (you will discover at least one step that exists only in someone’s head, and that discovery alone is worth the exercise).

Then run the comparison: a realistic agent cost for that workflow is tens to low hundreds of dollars a month to run, plus build time, plus a slice of daily review. If the workflow burns ten-plus hours a week of someone’s time, the maths will usually be loud. If it burns one, leave it alone. Not everything deserves an agent, and knowing which is which is most of the skill.