When an organization says it reduced emissions or conserved a forest, nobody takes its word for it. It has to show what it measured and how, and let someone else check. That's the logic behind measurement, reporting and verification (MRV) systems.
AI agents are now reaching tasks with real consequences: answering farmers, preparing donor reports and summarizing data for decisions. Yet they are often accepted on the strength of one good demo. We ask for more evidence about a tonne of carbon than about a system that works every day.
What an agent is, and why it changes things
An AI agent does more than answer. It can chain steps together: query data, use tools, write a draft and decide what comes next. That makes it useful and harder to evaluate, because a mistake in an early step can carry through and reach the final output looking like solid work.
Its performance can also shift over time. Data, models and user questions all change, so a successful test at launch doesn't guarantee quality six months later.
Four MRV principles that apply to AI
- Baseline. MRV compares against a starting point. With AI, measure how long the process takes and how good it is before automating it.
- Explicit methodology. What counts as correct is defined in advance. For AI, that means quality criteria and a set of test cases agreed with experts.
- Data traceability. Every figure should lead back to its source. So should every answer an agent gives.
- Independent, recurring verification. The people who build a system shouldn't be the only ones reviewing it, and the review has to be repeated over time.
What an agent audit looks at
An audit answers three questions: what did the agent do, with which data and to what quality? In practice, that includes:
- Activity logs. Queries, steps, sources and answers are stored so they can be reviewed.
- Quality evaluation. Experts review a sample of outputs against agreed criteria.
- Bias and errors. You check whether it fails more often with certain groups, topics or languages, and how serious those failures are.
- Limits and escalation. You confirm it recognizes what it doesn't know and hands over to a person when it should.
- Cost and governance. You review what it costs to run, who is accountable for it and whether it meets donor and regulatory requirements.
Where to start
You don't need a large program to begin. Pick an agent or workflow already in use, turn on logging for conversations and actions, and ask two experts to review a sample every month against written criteria. That simple routine gives you something many organizations lack: evidence of how their system actually behaves.
In our traceability, MRV and verification work, we apply the same discipline to AI agents as to field data. If your organization already uses AI, here's a useful test: could you show a donor what the system did last month?