In June, the day Claude Fable 5 dropped, I published a LinkedIn post saying I was speechless. The model impressed me less than what it exposed about the humans using it. The post collected 397 reactions and 185 comments, and the comment section proved the point better than the post did: most of the pushback came from people describing vague goals and blaming the tool for vague results.
For two years the standard excuse in every B2B team was some version of "it hallucinates" or "it does not get our business." That excuse just died. On open-ended agentic work, the benchmark I quoted in the post, performance went from 26% to 76% in a single model generation. Give a modern agent a clear goal and it runs for days. Give it a fuzzy one and it mirrors your confusion back at you with impressive fluency. This article is the long version of that argument, written from inside Growth Cab, the GTM advisory I founded, where we deploy AI agents for business workflows every week: prospect research, outbound, content, CRM hygiene, reporting.
in
“Claude Fable 5 just dropped and I'm speechless... Not because of the model. Because of what it exposed about humans.”
Why AI Agents for Business Fail in Most Teams
When a company tells me their AI pilot went nowhere, I ask one question: show me the exact instruction you gave it. Nine times out of ten the instruction reads like a wish. "Improve our outbound." "Write posts that sound like us." "Find good leads." No definition of good, no constraints, no example of a great outcome, no way for the agent to check its own work.
Then the output comes back mediocre and the team concludes the technology is overhyped. The uncomfortable truth is that the same brief handed to a talented new hire would produce the same mediocre result. The difference is that the new hire would come back and ask you twenty clarifying questions. The agent politely fills the gaps with averages, and average is exactly what you get.
The Mirror Test
The most useful mental model I can offer: the AI is a mirror. Bad output means blurry thinking. If the agent's draft feels generic, your positioning is probably generic. If its lead list misses the mark, your ICP definition probably lives in your head and nowhere else. Every disappointing output is a diagnostic report on your own clarity, delivered in seconds instead of a quarter.
This is also why the "secret prompt" economy is mostly a scam. People collect magic prompts to skip the thinking, then wonder why the copied prompt performs for its author and dies in their hands. The author's prompt encoded their strategy. You copied the words and left the strategy behind. Plenty of the people saving prompt libraries could not explain their own quarter to a new hire, and no template fixes that.
We ran this test on ourselves before running it on clients. Last quarter we audited every agent workflow inside Growth Cab and rewrote the briefs where the output had been disappointing. In every single case the rewrite made the old model version perform better too. The clarity was the upgrade. The model was just waiting for it.
What a Clear Goal Actually Looks Like
Here is the standard we use at Growth Cab before any agent touches a revenue workflow. A goal is clear when it has four parts. One, a definition of done: "a list of 50 companies that raised in the last 90 days, with the revenue leader identified and a verified email." Two, the constraints: regions, segments, tools it may use, things it must never do. Three, one worked example of excellent output, because an example transfers more strategy than a page of adjectives. Four, a self-check: the questions the agent must answer about its own work before showing it to you.
Write those four parts once and something surprising happens: the document outlives the agent. It becomes onboarding material, QA checklist and strategy memo at the same time. Teams that do this get compounding returns from every model upgrade, because their clarity is stored in a reusable form. Teams that never wrote it down experience each upgrade as noise.
Where the Agents Are Genuinely Not There Yet
Honesty section, because this keyword attracts a lot of hype. There are still places where AI agents for business use cases underperform a mediocre human. Long-running work where the context changes week after week. Political judgment inside an organization. Anything where the ground truth lives in conversations nobody transcribed. And high-stakes external communication still needs a human sign-off; we never let an agent send a cold email or a proposal unreviewed, and our clients sign off on sequences before they go out.
None of those weaknesses excuse the vague-goal problem. They just define the border of it. Inside the border, from research to drafting to enrichment to reporting, the machine is there now and the bottleneck is the brief.
Your Move
Take the workflow you most want to hand to an agent. Before opening any AI tool, write the four parts: done, constraints, one excellent example, self-check. If you cannot finish that document, you have located the real problem, and it was never the model. If you can, hand it to an agent this week and judge the output against your own definition of done rather than a vibe.
I share one AI revenue play every day in The Revenue AI Brief, the daily newsletter for B2B operators, including the exact agent briefs we use for prospect research and outbound at Growth Cab. Drop your work email below and you will get tomorrow's edition. And if you think I am overselling the mirror argument, tell me on LinkedIn. I answer everything.
