Buying guides
Automate work that is frequent, bounded, measurable and actionable
- Choose the first task for an AI agent with four questions. Does it come up often, does it come in a few known kinds, can someone check the result, and does it end in a concrete action?
- Fin used these four tests to pick the work for an agent that now handles its engineering alarms, and reports that the agent completes 41% of the weekly fixes.
- Ask the questions about one step of a job, not the whole job. Whole jobs rarely pass the second and third.
- Run the eight-question check below on your candidate task. The first no tells you what to fix before anyone builds.
We think the first task you hand to an AI agent should pass four tests: it is frequent, bounded, measurable and actionable. A first agent whose work can be checked builds the trust and the records that harder jobs need, so we would start with the dullest task that passes all four, not the job that hurts most.
The four words come from Fin, the company behind the Fin customer service agent. On 15 October, Oisin Hurley of Fin explained when a skill should become an agent, meaning when an AI helper that an engineer starts by hand should become an agent that acts on its own. This post turns his engineering test into one any owner can run. The SI company check scores a whole company. This filter scores one task.
How Fin chose its first agent
Fin's engineers had a Claude skill that could look into alarms about failed background jobs and suggest a fix. An engineer still had to start it, check its work and act on it. Before building more, Fin set a bar. An agent was justified only if it could safely take an alarm to a measurable outcome and cut the work left for engineers.
Why Fin's alarms suited an agent
Frequent
80%
of the pages reaching the team's backup on-call engineer, with about 50 new alarms a week
Bounded
3
kinds of failure behind most alarms: a brief outage, a bug or bad data, or a job that died without saying why
Measurable
33
distinct past scenarios, cut down from 63 real alarms, to test the agent's decisions against known answers
Actionable
3
endings for every alarm: the work retried safely, a fix opened for review, or a short hand-off to an engineer
All four figures are Fin's own, from its 15 October post.
The first version failed. One prompt had to gather evidence, decide what mattered, find the cause, recommend a fix and explain it all, so when it was wrong nobody could tell which part had failed. Its long reports were hard to verify, and engineers largely ignored them.
For the second version, Fin started by asking the people who did the work what a useful outcome was. Code now supplies the facts, and the model makes two narrow calls. Fin launched the agent in August and reports that it now completes 41% of weekly alarm fixes, counted by code changes that engineers merged, up from 4% five weeks before the post. Fin says it is still early.
The four questions in plain words
Frequent means the task comes up every week, in numbers you can count from last month's records. Testing and checking an agent takes work, and a task that comes up twice a month is unlikely to repay it. Fin's alarms often arrived faster than people could clear them.
Bounded means the people who do the task today can list its kinds of case on one page. A refund desk that sees five request types passes. A desk that answers anything a customer writes does not, and a person should keep it.
Measurable means someone can look at a finished case and say whether it was done right, and you have past cases to test an agent on before it touches live work. Fin replayed past alarms against its code as it stood before each human fix, so the agent could not see the answer. If nobody can say what right looks like, grade past cases first.
Actionable means each case ends in one of a few concrete actions, such as approve, reply, update a record or pass to a person, and a wrong action can be caught before it costs much. Fin's first version only suggested fixes, so engineers still did the work.
An assistant may be enough
Fin asked two more questions before building. The first is whether running the skill by hand would deliver 80% of the gain. If a person with an AI assistant already gets most of the benefit, stop there. Anthropic's guide to building agents gives the same advice, to find the simplest solution that works, which may mean no agent at all.
The second question is about upkeep. Models change often, and each change means new tests, monitoring and safeguards, so letting the agent act alone has to be worth that cost every time.
Ask the questions about a step, not a job
In our view, a whole job rarely passes the bounded and measurable questions, while one step of it often does. Fin's second version made the same split.
The same four questions, asked of a job and of one step
| Answer all customer email | Sort refund requests by type | |
|---|---|---|
| Frequent | Yes | Yes |
| Bounded | Any topic, in any tone (No) | A short list of request types (Yes) |
| Measurable | A good reply is partly taste (Partly) | Each sort is right or wrong (Yes) |
| Actionable | A wrong reply reaches the customer (Partly) | File each request in the right queue (Yes) |
Answer all customer email
- Frequent
- Yes
- Bounded
- Any topic, in any tone (No)
- Measurable
- A good reply is partly taste (Partly)
- Actionable
- A wrong reply reaches the customer (Partly)
Sort refund requests by type
- Frequent
- Yes
- Bounded
- A short list of request types (Yes)
- Measurable
- Each sort is right or wrong (Yes)
- Actionable
- File each request in the right queue (Yes)
An illustration of the method, not a measured case.
Where the budgets and the engineers are going
HCLTech, a technology services firm with more than 230,000 people, reported on 12 October that its advanced AI revenue, a line it defines itself, grew 78% in a year in constant currency, while its total revenue grew 2.6%.
We read that as clients moving budget they already spend toward AI work, more than finding new budgets. For an owner, that points the first agent at work the business already does and pays for every week.
CNBC reported on 15 October that US job postings for forward deployed engineers, who work directly with companies to turn AI experiments into systems they use, passed 7,300 in the first nine months of 2026, up from fewer than 200 in 2023, according to Lightcast. Consulting firms were among the biggest recruiters, led by Deloitte with 872 postings and Accenture with 688. Postings are not hires, and CNBC notes that evidence of financial returns from these projects is still thin.
When the engineers who deploy agents are scarce, the cheapest first project is the one that needs the least of their time, and that is a task that passes all four questions. An owner can do the first part without them: choose the task, count last month's cases and name who checks the result.
The self-check
Answer for one task as it runs today. The eight questions come in four groups, in the order of the four tests, and the first group with a no decides your result. Frequency comes first because without volume nothing pays back. A short list of case types is what makes checking possible, and checking comes before letting an agent act.
Should we hand this task to an agent?
Count a yes only if it is already true today. A plan counts as a no.
Frequent
1Does the task come up every week, in numbers you could count from last month's records?
2Does it take a visible share of someone's week, or pile up faster than people can clear it?
Bounded
3Could the people who do it today list its main kinds of case on one page?
4Does each case start from a clear trigger, such as an email, a form or an alert?
Measurable
5Could someone look at a finished case and say whether it was done right?
6Do you have past cases, with what was done, to test an agent on before it touches live work?
Actionable
7Does each case end in one of a few concrete actions, such as approve, reply, update a record or pass to a person?
8Could a wrong action be caught or undone before it costs much?
8 questions. Answer yes only for what is true today.
Where to start
Pick three tasks your team does every week and run the check on each. Start with the one that passes all four and had the most cases last month.
Two causes we set out in why AI pilots die map onto these tests. Nobody defined what accurate means, which fails the measurable test, and failures had nowhere to go, which fails the actionable one.
extendfuture's AI process automation method takes tasks in the same order: a baseline of what the task costs today, one workflow automated with exceptions routed to people, then the numbers side by side.
Sources
- Oisin Hurley, Fin, When should skills become Agents?, 15 October 2026
- HCLTech, HCLTech delivers strong Q2 led by record deal bookings of $3.3 billion, 12 October 2026
- Barbara Booth, CNBC, AWS, Anthropic race to deploy army of specialized engineers key to proving AI spending will pay off, 15 October 2026
- Anthropic, Building effective agents, 19 December 2024