Leadership · 5 minute read
How to Choose Your First AI Agent
Choose a first agent by scoring candidates on volume, rule clarity, measurable baseline, contained blast radius, and data accessibility, then pick the highest score whose business owner is genuinely available. The goal is production evidence within a quarter, not the most impressive use case.
The first agent matters more than its business case, because it determines whether the organization believes the program. A first deployment that reaches production with measured results creates the conditions for everything after it; one that stalls in a pilot teaches everyone that AI does not work here. This guide gives a scoring method, the tempting candidates to avoid, and what good looks like.
What are the five criteria?
| Criterion | Score high when | Score low when |
|---|---|---|
| Volume | Thousands of instances a month | Dozens; value cannot cover the build |
| Rule clarity | A written procedure exists and people agree on it | Judgment varies by who does it |
| Measurable baseline | Cost, time, and error rates already tracked | Nobody has measured it |
| Contained blast radius | Errors are visible quickly and reversible | Errors reach customers or regulators first |
| Data accessibility | Inputs available through systems the company controls | Data in email, paper, or third-party systems |
Score each candidate one to five on each criterion and rank. The AI use-case scoring framework guide provides a fuller weighting model.
What is the sixth, unscored criterion?
Owner availability. The business owner of the process needs to be genuinely available for the specification work, the exception design, and the supervised deployment period. A process that scores well on every criterion but whose owner is mid-restructuring, on the road four days a week, or personally sceptical will fail anyway.
This criterion decides more first deployments than the other five combined, and it is the one most often ignored because it feels political. Ask directly, and choose the second-ranked process with an engaged owner over the top-ranked one without.
Which tempting candidates fail?
The executive's favorite idea. Frequently something intriguing with no baseline and no volume. It gets funded, produces a demo, and stalls. If the idea is genuinely good it will still be good in six months, with evidence behind the program.
The hardest problem. Chosen to prove that AI can do something impressive. It combines ambiguity, poor data, and high consequence, which is the exact profile of a failed first deployment.
The most visible customer journey. Attractive because the benefit would be obvious, dangerous because early errors are public and the organization has no experience running an agent yet. Come back to it third.
Anything with inaccessible data. Six months disappear into integration before anyone sees an outcome.
The why AI pilots fail guide covers the wider pattern.
What does a strong first candidate look like?
Usually unglamorous: invoice matching, document intake and completeness checking, ticket triage and routing, appointment scheduling, exception handling in a supply or service process, or internal helpdesk resolution. High volume, written rules, measured today, errors caught quickly, data in systems the company runs.
The fact that these are not exciting is the point. The first agent's job is to prove that the company can specify, build, evaluate, deploy, supervise, and operate an agent. The exciting applications become possible because that capability exists.
What should the first deployment produce?
- Production operation under supervision with real volume.
- A measured improvement against the baseline.
- An evaluation pass rate and the set behind it.
- A named owner who presents the evidence.
- Reusable integrations that the second agent inherits.
- An operating rhythm that outlasts the project.
The AI agent lifecycle explained for executives piece describes the stages this passes through.
How long should it take?
About a quarter from decision to supervised production, for a well-chosen process with an available owner. If the estimate is much longer, something is wrong: the process is too complex, the data is not accessible, or the platform is being built simultaneously without that being acknowledged in the plan. The last case is common and legitimate, but it should be stated so nobody thinks agent delivery is inherently slow.
What about the second agent?
Choose it while the first is in supervised deployment, not after. The second should reuse the first's integrations where possible, which is what makes it faster and cheaper, and it should stretch the organization slightly: a little more consequence, a little more autonomy, or a different function. Companies that wait until the first is fully complete lose momentum and rebuild context; companies that start five at once have no capacity to supervise any of them properly.
What should executives ask before committing?
- What is this process's current cost, cycle time, and error rate, and who measured them?
- Is the procedure written down, and do experienced people agree on it?
- What happens if the agent gets it wrong, and who notices?
- Can we reach the data through systems we control?
- Is the owner actually available for the next quarter?
- Will the integrations we build be reusable?
How can FISTA Solutions help?
FISTA Solutions scores candidate processes with clients, establishes the baseline before any build, and delivers the first production AI agent under supervision with reusable integrations and an evaluation set the company owns, through its AI enablement practice. Since 2017, FISTA has delivered 150+ projects for 50+ companies across 12+ countries; clients report efficiency gains of up to 47% on automated processes.
To score your candidates and pick a first agent that produces evidence, talk to FISTA on WhatsApp, or read how to prioritize AI use cases.
Share-ready article cover
Download the generated social format.
Clear answers
Questions raised by this field note.
Straightforward guidance for evaluating scope, fit, and the next step.
01What makes a good first AI agent?
High volume, rules that can be written down, a baseline someone has already measured, errors that are visible and reversible, and data the company can actually reach. Add an available business owner, and the deployment can produce real evidence within a quarter, which is what the first agent is for.
02Which first AI agent candidates usually fail?
The executive's favorite idea with no baseline; the hardest problem in the company, chosen to prove value; the most visible customer journey, where early errors are most costly; and anything depending on data the company cannot access or processes nobody has written down.
03How long should the first AI agent take?
Roughly a quarter from decision to supervised production if the process is well chosen and the owner is available. Longer usually means the process was too complex, the data was not accessible, or the platform was being built at the same time without acknowledging it.
04Should the first agent be internal or customer-facing?
Internal, in most cases. Internal processes let the company learn supervision, evaluation, and operations with contained consequences. Customer-facing agents are a better second or third deployment, once the organization has evidence that it can run one safely.
05What does success look like for a first AI agent?
Production operation under supervision with a measured improvement against baseline, an evaluation pass rate, a named owner, reusable integrations, and an organization that now knows how to do this. Impressiveness is not a success criterion.
Continue exploring
Related capabilities
Start with the hard problem
Need the outcome owned, not merely analyzed?
Tell us where delivery is constrained. We’ll map the fastest credible path from intent to verified production.