Your next agent's best home already exists. Five tests to find agent-ready work
The agents that reach production fastest slot into a decision an existing process already makes. A La Madre framework to find those places in an afternoon.
Listen to this article · 6 min
AI-generated narration of the full article.
Most enterprise agent programs start with a list of ideas. Someone runs a workshop, teams propose use cases, and the backlog fills with “an agent that answers questions about X.” A few months later, the pilots that made it to production are rarely the ones that looked best in the workshop.
Our view, built over the past month of coverage, is that the agents that reach production fastest have something in common: they slot into a decision that an existing process already makes. We called this thesis “the best enterprise agents disappear into existing processes” in our September synthesis. This article turns it into a test you can run.
The evidence so far
Look at where agents are landing:
- In security, Exabeam’s October release puts an assistant inside the analyst’s investigation screen, gathering context and running searches, while the analyst keeps the decision. We covered it in our piece on the agentic SOC.
- In knowledge work, Progress put its multistep retrieval agent inside Microsoft Teams channels, where questions already get asked, as we discussed in our analysis of agentic RAG in Teams.
- In drug discovery, Hansa is embedding Cradle’s design model inside the design and test cycle its scientists already run, examined in our case analysis of AI inside scientific workflows.
- Earlier, Google placed vulnerability agents inside code review, in our analysis of agents as security controls, and Snowflake built accounts payable around deterministic checks with AI for the exceptions, in our InvoiceIQ case.
None of these is a new chat window. Each is a new participant in an old process.
Five tests for agent-ready work
Score each candidate from 0 to 2 on each test. It takes an afternoon with the process owner.
- 1. A decision already exists, with an owner
- 2. Inputs are digital and reachable with identity
- 3. A deterministic check sits next to it
- 4. A person already approves at that step
- 5. The outcome is measured today
1. A decision already exists, with an owner. The best candidates improve a decision someone is already accountable for: approve this invoice, triage this alert, choose this candidate molecule. If nobody owns the decision today, the agent will not create ownership; it will create ambiguity.
2. Inputs are digital and reachable with identity. The data the decision needs is already in systems the agent can read, under an identity you can scope and review. If the inputs live in email attachments and people’s heads, digitize first.
3. A deterministic check sits next to it. Something verifiable surrounds the step: a three-way match, a policy rule, a test suite, an assay. The agent proposes; the check confirms or rejects. This is what keeps error rates bounded without reviewing everything by hand.
4. A person already approves at that step. Human review is cheapest where it already happens. An agent that prepares the case for an existing approver adds speed without adding a new control. An agent that needs a new approval step adds a control nobody budgeted.
5. The outcome is measured today. Cycle time, error rate, backlog, cost per case. If you cannot measure the process before the agent, you cannot show the agent helped, and the pilot will be judged on anecdotes.
Reading the score
Eight to ten. Build it. Start in “prepare and propose” mode, measure against the baseline, and widen autonomy only where the deterministic check is strong.
Six or seven. Usually one test is weak. Fix that test, often identity or measurement, then build.
Five or less. The process is not ready for an agent. That is still a finding: the work to make it ready, clear ownership, digital inputs, a check, a metric, often pays off before any AI is involved.
A worked example: manual journal entries
Take a SOX-relevant control, such as reviewing manual journal entries above a threshold. The decision has an owner, the inputs live in the ERP, there are rules for what a valid entry looks like, a reviewer already signs off and auditors already test the control. That scores near ten. An agent that pre-reviews each entry, flags anomalies with evidence and leaves the sign-off with the reviewer fits the process the auditors already understand. Compare that with “an agent that answers finance questions,” which fails tests 1, 3 and 5.
What this framework is not
It is not a ranking of value. A low-scoring idea can still be important; it just needs process work first. And it does not say agents should never run autonomously. It says autonomy should be earned where a deterministic check and a measured baseline make mistakes visible.
What to do this month
- Take your agent backlog and score every item on the five tests.
- Kill or park anything at five or below, and log the process gap it revealed.
- Pick the top three and agree baseline metrics with the process owners before building.
- Design for prepare-and-propose first. Plan the path to more autonomy, test by test.
The agent that changes your business probably will not have a chat window. It will have a place in a process that already works, and it will make that process faster.