Better uses for agents?
I know everyone says I'm supposed to, but I'm not connecting agents to my digital life just yet. I suspect many of you aren't yet either.
We all want to be Tony Stark with Jarvis. If a model’s actions were traceable, repeatable, predictable, and accountable (thanks Joel) I’d wire one into everything tomorrow.
They aren’t, so I haven’t.
(I’m guessing that I’m in bigger company than the “agentic everything” headlines suggest. Surveys say most large enterprises now have agents in production. Look closer and most of those are assistants following prescribed paths, with a human finishing the job.)
There are two surveys on this issue (links below): in Sinequa’s, half of large enterprises said they had agents in production, yet only 29% described their most advanced deployment as a true agent.1 In Zapier’s, 71% had agents in production, but only 30% said those agents complete most tasks without human intervention.2
That doesn’t make me a sceptic. I use these tools daily and my observation is simply that we may be pointing them at the wrong jobs.
The question I now ask of any use is: if there is always at least a bit of wrong stuff in GenAI outputs, who absorbs those errors?
When Copilot or ChatGPT drafts three replies and I pick one, I absorb it, at a cost in effort and risk that I decide I can afford. When that same model sends the reply for me (ie as an agent), someone else has to do the work, and I find out later.
It seems to me the single question “who absorbs the errors” sorts the good uses from the worrying ones.
So, for example:
- Don’t automate the process.
Generate the edge cases the process forgot. - Don’t book my trip.
Sketch five itineraries I’d never have thought of. - Don’t run customer service.
Play the angry customer so teams can rehearse. - Don’t decide the case.
Explain the policy in the citizen’s own words.
Then there’s the part I really don’t understand. We spent years trying to stamp out hallucinations. In a game character, a brainstorm or a rehearsal, making things up is the whole point. We’ve been focusing on suppressing the thing these models are really really good at.
The transformer was invented for translation3 (Google Translate itself now runs a Gemini model). The best uses still are translation: jargon to plain English, policy to person, idea to prototype.
Are you exploring using these tools in a way where being wrong is cheap, or even useful? Happy to chat if you think it might be helpful.
Notes:
-
Sinequa, The State of Enterprise Agentic AI in 2026 (PDF) ↩
-
Zapier, AI agents survey 2026 ↩
-
Google blog post from a contributor to “Attention Is All You Need”: Transformer: a novel neural network architecture for language understanding ↩
What do you think?