Cheaper AI, Bigger Bills — How Nonprofits Can Plan Ahead
As AI gets cheaper, nonprofits may use more of it. Guardrails set early keep agent costs predictable.
Software has traditionally been a predictable cost for nonprofits. A licence or subscription, plus the occasional setup fee, that fits neatly into a fixed budget. AI agents have a setup cost too, but once they are running, the bill isn’t fixed. Their costs rise with how much information they process and generate. This leaves nonprofits in an odd position, where getting more work out of an agent can also mean paying more for it.
The challenge may only grow as the price of AI continues to drop. As AI becomes cheaper, nonprofits may use agents for more tasks, reflecting a phenomenon known as Jevons paradox. Even as the cost of individual tasks falls, the growing volume of work could push total spending beyond budget. Managing that trade-off begins with understanding what drives an agent's costs.
Where the costs come from
To understand this, let's follow an agent through a single task. Take a grant-writing agent asked to draft a funder's report. It reads the funder's guidelines, reviews the nonprofit's background documents, searches for evidence, drafts sections, checks requirements and makes revisions — all using the most capable model available.
The first round is inexpensive. But by the third, the agent is re-reading the guidelines, the earlier drafts and its own notes to address an unmet criterion. Each retry involves more reading, reasoning and writing, even when the work appears to be progressing normally. This helps explain why an agentic task can cost several times more than a simple chatbot interaction. Multiply that across a month's worth of funder reports, and a seemingly modest expense can quickly grow.
Looking closer, three factors drive the bill, with repeated use multiplying their impact:
The size of the task: The more information an agent reads, processes and produces, the more it costs. A report drawing on several lengthy documents will cost more than one using a handful of sources.
The number of rounds: Every additional round of research, drafting, or revision adds to the bill. An agent that repeatedly revisits documents or corrects its own work can turn a relatively affordable task into a more expensive one.
The choice of model: More capable models generally cost more. Running every step on the most powerful model means paying premium rates even for routine tasks such as sorting files.
Keeping costs predictable with governance guardrails
Understanding what drives an agent's costs is a starting point. But who decides how far an agent can go before those costs become excessive?
Consider the funder report again. Without limits, the agent could keep retrying until it finds the missing source or gives up, accumulating costs either way. Governance means deciding those limits in advance, rather than discovering them when the bill arrives. Nonprofits can start with three practical guardrails.
1. Set limits in advance
Decide how much an agent can spend and when it must stop or seek approval. This could mean setting budget alerts, limiting retries or restricting how much information an agent processes in one run. For the funder report, a retry limit could prevent repeated searches for a missing source. Someone should also be responsible for reviewing alerts and deciding when additional spending is justified.
2. Build in staff checkpoints
Sometimes, asking a person is cheaper than letting an agent keep trying. If the grant-writing agent cannot find a source, a staff member might provide the missing context in seconds. Checkpoints before costly steps and final submission can reduce unnecessary retries while catching errors that could damage the nonprofit's credibility.
3. Split work with delegated agents
Not every step requires the most capable AI model. A lighter-model agent could organise background documents and prepare a summary, leaving a more capable agent to draft the report’s narrative. By establishing which tasks warrant more expensive models, nonprofits can avoid paying premium rates for routine work.
Follow a grant writing agent through the creation of a funder report and the pattern becomes clear. Agent setup is a cost nonprofits can plan for, but running costs depend on the size of the task, number of rounds, and the model used. Governance helps nonprofits manage these costs before they become surprises: setting spending limits, assigning staff where judgment is needed and assigning routine work to less expensive models. As AI subscriptions become cheaper, nonprofits are likely to use more of it. Setting these guardrails early can help them scale their use of AI agents without losing control of their budgets.