Four Checks Before an AI Agent Pilot Gets 2027 Budget
Budget season puts a stack of artificial intelligence agent pilots in front of whoever approves spending, and most arrive with a good demo. The question for a 2027 budget is which pilots survive.
I use four checks before I back an agent pilot. Each one ends with something the sponsor can put on the table in the budget meeting. If one is missing, I wait.
Workflow fit: bring the runbook
Could a capable new hire do this work from a two-page runbook? If not, the agent will improvise, and improvising across thousands of customer records is how a pilot loses trust.
The test cuts the other way too. If the runbook holds no judgment calls, you don't need an agent. You need automation, which is cheaper and more predictable. Gartner analyst Anushree Verma has said that "many use cases positioned as agentic today don't require agentic implementations," and advised using agents "when decisions are needed" and automation for routine workflows. Two model developers' agent-building guides agree, and one says to write an agent's instructions from existing operating procedures, support scripts, and policy documents.
Take a hypothetical contact center that wants an agent to decide refund exceptions. If the policy is whatever two supervisors agreed on last week, writing it down is the pilot's real first deliverable. If that takes a month of meetings, the budget request is early.
Data fit: bring a sample of the records
An agent can't judge better than the records it reads. Gartner reported in February 2025 that 63 percent of organizations either don't have or aren't sure they have the right data management practices for AI, based on a July 2024 survey of 1,203 data management leaders. Its definition is the useful part: the data has to represent the real use case, errors and outliers included. A tidy demo dataset fails that test by design.
So before funding, pull a sample of the records the agent will read, say 50, and count how many carry every field the decision depends on, filled in and current. Picture a hypothetical sales team that wants an agent to route inbound leads by territory and industry. If a big share of the sampled accounts has no industry value, the agent will still route those leads. It just won't route them well, and nobody will notice until the wrong rep calls.
Oversight fit: bring a name and a working pause
Every agent produces exceptions: cases it can't settle, flags as uncertain, or gets wrong. The check is whether one named person reads them on a set day and can stop the agent without filing a ticket with whoever built it. A team doesn't count. When the answer to "who reads these?" is "the team," nobody does.
NIST's AI Risk Management Framework, which is voluntary guidance, asks that roles for managing AI risk be documented and clear, and that responsibility be assigned to "supersede, disengage, or deactivate" a system performing in ways inconsistent with its intended use. One of the developer guides describes agents pausing for human feedback at checkpoints, and it warns that their autonomy brings "the potential for compounding errors."
Consider a hypothetical agent that drafts and sends renewal reminders. Some replies say the account already canceled. If nobody owns those replies, the reminders keep going to customers who've already left.
Bring the owner's name, the review day, and a pause someone has tested.
Payback fit: bring the number, the baseline and the date
Gartner pinned the cancellations on "escalating costs, unclear business value, or inadequate risk controls." Two of those are payback problems, and the third is the oversight check. So name one countable number before launch, record its baseline and set a date inside the quarter to read it. Then subtract the hours the oversight check just added. A pilot that saves reps an hour a week and costs a supervisor the same hour hasn't paid back anything.
For a hypothetical agent that drafts follow-ups on open quotes, the number might be the share of quotes that get a follow-up within two business days, measured for a few weeks before launch and again after, net of the time reps spend fixing drafts.
The fair objection is that AI value is often indirect and arrives late. Gartner made that case in 2024, saying generative AI needs more tolerance for indirect, future value than for immediate return, and noting that many CFOs haven't been comfortable investing today for indirect value in the future. That's a fair point about a program. It's a weak one about a pilot. A pilot's job is to earn the next budget, and it can only do that with a number a quarter can move.
When a pilot misses a check, I don't kill it. I fund the missing piece first. Writing the runbook, cleaning the fields a decision depends on, naming an owner and baselining the number all cost far less than a canceled pilot, and each makes the eventual agent better. Sometimes the runbook shows the job only needed automation. That's a good outcome too.
A pilot that can put all four things on the table, the runbook, the record sample, the owner with a working pause and the baselined number, is one I'd fund. One that can't is asking for money to learn what it should have known before the meeting.
Matt Kundo is founder of Matt Kundo Digital Marketing.