Know what's running.
3 SIGNALS · 1 LESSON · 1 ACTION · ABOUT 5 MINUTES
We bring our judgment to AI, including our bad judgment. AI brings its own blind spots to the work, too.
Dr. Abel Sanchez shared a thoughtful LinkedIn post this week that really got me thinking about the judgment we bring to AI and how we develop it. I'm looking at what human oversight requires, what AI agents actually do, and how to tell whether their work is good enough.
SIGNAL 01
Being in the loop doesn't automatically make us right
What happened. Dr. Sanchez's LinkedIn post makes a great point: keeping a human involved doesn't automatically mean we're exercising good judgment. His question about whether a poor result came from the execution or from what we asked for prompted me to reflect on my own work with AI. I added that AI makes assumptions, too. In recent work, an AI gave me directions for menus it hadn't checked and reverted to an old date after I'd supplied the correct one.
Why it matters. I asked it to assess our mix-ups: what could I have done better, and which mistakes were its own? That helped me examine both sides. Its explanation still needed checking against what happened.
What to do. Review one disappointing result. Separate problems with the requested outcome, missing context, and mistakes in execution. Decide what evidence supports each explanation before changing your instructions.
SIGNAL 02
An AI agent can choose the next step
What happened. The term “agent” covers different designs. Anthropic's engineering guide distinguishes workflows that follow predefined paths from agents that decide how to proceed and which tools to use as they work.
Why it matters. Consider an agent handling a customer inquiry. It could look up an order, check policy, draft a reply, and update a record. Each step depends on access and permission. A mistaken assumption could affect later actions.
What to do. Define one agent's job, permitted actions, and stopping conditions. Specify when it must ask a person for help. Check that its actual permissions enforce those limits before letting it change live records.
SIGNAL 03
Define a good result before testing the AI
What happened. That same Anthropic guide emphasizes measuring performance and adding complexity when it improves outcomes. My practical application is to define an acceptable result before reviewing what the AI produces.
Why it matters. A customer reply can sound helpful while promising an unauthorized refund or exposing someone else's order details. Polished writing alone tells you little about whether the task was completed correctly.
What to do. Choose five representative cases, including an exception and a case that requires escalation. Write down the expected result and prohibited actions. Test with fictional or appropriately protected data, and inspect both the response and any changes the agent makes.
THE LESSON
Review the instruction and the execution
Suppose you ask an agent to resolve customer complaints as quickly as possible. It closes tickets faster. The dashboard improves. Customers keep calling back because their problems remain unresolved.
Start by checking the objective. Did “resolve” mean the customer received a verified answer, or did the team end up measuring how quickly a ticket was closed? Then inspect the execution. Did the agent use the current policy, identify the correct customer, and stay within its authority?
Those findings call for different fixes. An unclear objective needs a better definition of success. An outdated policy needs a corrected information source. An unauthorized action needs a permission change. Rewriting the prompt won't necessarily address all three.
I've found that comparing responses from multiple AI systems can help expose assumptions and gaps. The disagreements are often more useful than the answers themselves because they give me something specific to investigate. Agreement still needs checking against the source material or actual result.
Give the person reviewing the work enough time, evidence, and authority to question the objective as well as the output. They should be able to inspect what the agent did and pause its work when the result is unclear.
DO THIS WEEK
Run a five-case review
Choose one recurring AI-assisted task and involve the person who knows the work. Use a test environment for anything that can take action.
Set expectations. Select five cases. Include routine work, an exception, and a case where the AI should stop and ask for help. Record an acceptable result for each before running the test.
Inspect the work. Check accuracy, information access, permitted actions, and whether the result solves the intended problem. Record the corrections and review time required.
Find the cause. For each failure, examine the objective, supplied information, tool permissions, and execution. Use the evidence to choose a fix.
Check the fix. Rerun affected cases and try an additional case the change wasn't designed around. Name the person who will decide whether the task is ready for broader use.
Five cases are a starting point for finding problems. Expand the review to match the range of work and the consequences of a mistake before increasing the AI's responsibility.
WHEN YOU WANT HELP CHECKING THE WORK
Hackademic Solutions helps small and medium-sized organizations define AI responsibilities, set practical safeguards, and evaluate results. Book a free 20-minute assessment to talk through one AI-assisted workflow in your business.
See you next week.
— Chris Simpson
Hackademic Solutions
The Hackademic Briefing · Know what's running.