How to choose your team's first agentic AI workload
Score candidate workloads on six questions and three gates to pick a first agent your team will trust, and skip the picks that stall most pilots.
Series: Where to start with AI: from your week to your org (3 articles)
- 2How to choose your team's first agentic AI workload This article
Enterprise Solutions Architect
Your team's first agent should be the boring one: a frequent, rule-heavy task with a clear owner, read-only access, and a person who checks the result. Score your candidates on six questions, run the top few through three gates, and you'll know which one to build first and which to leave for later.
What you need: the shortlist from part 1 of this series from each person, and 90 minutes with the people who actually do the work.
What changes when the workload belongs to a team
In part 1, you automated your own week. If something went wrong, only you saw it. A team workload is different in three ways that matter for picking the first one.
More people depend on the output, so mistakes are visible. The agent needs access to shared systems, which means someone has to approve its permissions. And the first pilot sets expectations for every pilot after it. A first agent that fails publicly makes the second one harder to fund.
That's why the right first pick is rarely the most impressive one. It's the one most likely to work and be trusted. AWS Prescriptive Guidance makes the same point: give each agent a clear job and boundaries, not just a cool demo .
Run it as a 90-minute session
Choosing alone is the fastest way to pick wrong. The people who do the work know where the edge cases are. Run it as a short working session:
- List (20 minutes). Everyone brings two or three repeated tasks from their own week. Put them all on one board.
- Ladder (20 minutes). Run each task through the five rungs from part 1: eliminate, simplify, automate with a script, assist, delegate. Anything that stops at the first three rungs leaves the board. Those are still worth doing, but they don't need an agent.
- Score (30 minutes). Score what's left with the six questions below.
- Gate and pick (20 minutes). Run the top three through the gates and pick one.
Score each candidate on six questions
Give each question a 1, 2, or 3. Add them up.
Question | 1 | 2 | 3 |
|---|---|---|---|
Volume: how often does the team do this? | Monthly | Weekly | Daily |
Time: how long does one instance take? | Minutes | Up to an hour | Hours |
Rules: how much is fixed logic versus judgment? | Mostly judgment | Mixed | Mostly rules |
Reach: can the team already read the inputs? | Needs new access | Some of it | All of it |
Blast radius: how bad is a wrong answer caught late? | Serious | Annoying | Trivial |
Measurable: can you state a before and after? | No | Roughly | Clearly |
These questions follow the scoping phase in the Well-Architected Generative AI Lens : feasibility, risk, data availability, and how you'll measure success.
If two candidates tie, pick the one with the smaller blast radius. A first agent that's slightly less useful and never embarrassing is worth more than the reverse.
Pass three gates
Scores rank the candidates. Gates decide whether a candidate can go first at all. A candidate has to pass all three.
- A named owner. One person reviews the output every week and decides when it's good enough. "The team" doesn't count.
- A human checkpoint. A person sees the result before anything leaves the team, reaches a customer, or changes a record.
- Access that exists today. The agent can read every input with permissions that are already approved. If getting access will take a month, the candidate waits.
Candidates that fail a gate aren't bad ideas. They're second or third projects.
In the AWS Agentic AI Security Scoping Matrix , these gates keep your first agent in the lowest-risk scopes: read-only, or acting only after a person approves.
Patterns that make good first picks
Across the teams I work with, the same few shapes keep winning:
- Prep briefs. Before a meeting or a review, an agent pulls recent notes, email threads, and what changed in the relevant systems into a one-page brief. A person reads it. If it misses something, the cost is a few minutes.
- Reconciliation and gap reports. An agent compares what should match across two systems and lists the differences. It writes a report and changes nothing.
- Investigation assistants. When someone has to chase a question across several systems, such as why an invoice is disputed, an agent gathers the records, the related correspondence, and a timeline, then drafts a summary. The analyst still makes the call. The agent removes the searching.
- Triage drafts. An agent reads incoming requests, proposes a category and an owner, and a person confirms or corrects it.
All four read from several places, produce a draft, and leave the decision with a person. That's the shape you want first.
Patterns that make poor first picks
- Customer-facing chat. It's public, the scope is wide open, and every mistake is visible. Build it later, once you've learned how your agents fail.
- "Ask anything about all our documents." There's no boundary, no baseline, and no easy way to tell if it's right. Start with one document set and one kind of question.
- Anything that writes to financial or production systems. Write access raises the bar for testing and approvals. Earn it with a read-only agent first.
- A broken process. If the process itself is the problem, an agent makes the wrong thing faster. Fix it at rung 1 or 2.
- A task nobody owns. It will fail the owner gate, and it should.
What a team build adds to a personal one
Part 1 described the basic parts of an agent job: a trigger, a model, tool access, a place for the output, a notification, and a checkpoint. A team workload adds three things:
The agent gets its own identity. It should act with its own scoped, read-only permissions, not a person's credentials. That way you can see what it did and revoke it without locking anyone out. Amazon Bedrock AgentCore includes a managed runtime with built-in identity, a gateway for exposing tools to agents, and observability for tracing each step. It works with open-source frameworks such as Strands Agents .
A small test set. Collect 20 to 30 real past cases with the answer a person would have given. Run the agent against them before the pilot starts and after every change. Amazon Bedrock evaluations can score outputs using rules, a judge model, or human reviewers. Part 3 goes deeper on this.
A feedback channel. Give the people reading the output one place to say "this was wrong" and why. Every report becomes a new test case.
What I've seen go wrong
Picking the demo instead of the workload. The most impressive idea in the room usually scores low on volume and high on blast radius. The winner is often the task everyone finds tedious.
Skipping the baseline. If nobody wrote down how long the task took before, nobody can show the pilot helped. Time three or four instances by hand before you build.
Letting the pilot grow. A first agent that starts as a prep brief and turns into "also draft the follow-up, update the record, and schedule the next meeting" during the pilot never finishes. Write the scope down and add to it after the first version is trusted.
Where to start this week
- Ask each person for two or three repeated tasks from their own week
- Hold the 90-minute session: list, ladder, score, gate
- Pick one candidate that passes all three gates
- Name the owner and write the scope in a few sentences
- Time three or four instances by hand as a baseline
- Collect 20 to 30 past cases with the right answers
- Build it read-only, with a human checkpoint and a feedback channel
Next in this series
Part 3 covers what changes when other people start depending on your agent: test sets, guardrails, tracing, cost, and what to do when the model underneath it changes.
For more, see part 1 , the Agentic AI topic page , and the Best practices topic page here on Builder Center.
Series: Where to start with AI: from your week to your org (3 articles)
- 2How to choose your team's first agentic AI workload This article
Enjoyed reading this content? Let the author know!
Your likes, comments, shares, and saves help creators reach more builders.
Loading recommendations
Loading article