Simple prompt or agent workflow? How not to overthink AI

Photobank2/iStock/Getty Images Plus

Follow ZDNET: Add america arsenic a preferred source on Google.

ZDNET's cardinal takeaways

• Gen AI occurrence is astir knowing which attack to use.
• Start with the simplest instrumentality that solves your problem.
• Sometimes you'll request a prompt, and different times an cause workflow.

"You're astir apt making AI harder than it needs to be." 

This advice from Corey Noles and Grant Harvey's latest occurrence of The Neuron podcast urges greater simplicity successful what has go a analyzable and confusing matter successful caller years -- making generative AI acceptable for the organization. 

Also: 3 astute ways concern leaders tin physique palmy AI strategies - earlier it's excessively late

With 95% of generative AI projects failing, arsenic recovered in a caller MIT study, it's clip to enactment these initiatives into perspective. 

Many radical wrong and extracurricular the exertion assemblage presumption implementing gen AI arsenic a large task that requires rigorous task absorption discipline. Experts adjacent notation to that improvement with ZDNET, telling radical not to unreserved into AI, but to deliberation cautiously astir what they are doing first. 

Also: Gartner says adhd AI agents ASAP - oregon else. Oh, and they're besides overhyped

While immoderate AI initiatives, astatine a basal level, don't necessitate specified overthinking, others, connected the different hand, bash necessitate much thoughtful planning. So, wherever does a budding AI nonrecreational gully the line?

Overall, radical "are getting it each wrong" with implementing AI, Noles and Harvey asserted. Success is simply a substance of knowing erstwhile the simplest approaches use -- utilizing a screwdriver alternatively of applying a sledgehammer for a situation that whitethorn necessitate AI. It's besides astir knowing erstwhile much blase approaches are needed to forestall AI projects from going disconnected the rails.

The prodcasters suggested 4 ways to lick a occupation with gen AI: 

1. Simply "chatting."
2. Writing a structured prompt.
3. Building a project.
4. Deploying an agent. 

Also: Gen AI disillusionment looms, according to Gartner's 2025 Hype Cycle report

"Start with the simplest instrumentality that solves your problem," Noles advised. "Don't physique an cause erstwhile a basal chat volition do. We find radical get ace excited astir each these awesome tools. But successful practice, you truly don't request to leap consecutive to gathering analyzable agents erstwhile a elemental punctual works." 

But besides "don't walk 2 hours pulling your hairsbreadth crafting the cleanable punctual and trying to one-shot a analyzable task erstwhile an cause could grip the full workflow for you," helium added. "There are decidedly times you truly bash request to trade an epic punctual oregon physique an cause workflow. The instrumentality is knowing erstwhile to usage what."

Also: 5 ways automation tin velocity up your regular workflow - and implementation is easy

Noles and Harvey supply immoderate guidelines to assistance professionals determine however overmuch effort an AI-related task needs:

• When to simply chat: "If you privation to dive heavy down immoderate rabbit spread astir the past of bunny rabbits and whether they truly devour carrots, you tin perfectly bash that successful a mean chat window," said Noles. "You tin bash an atrocious batch much determination than what astir radical would think. There is simply a full batch of worldly you tin bash successful a chat model conscionable asking it a question without having to bash immoderate fancy prompt, without having to get into each of that." 
• When to bash a structured prompt: A structured punctual is champion for a multi-part oregon a formatted output, said Harvey: "A multi-part task would beryllium you request the AI to spell retired and bash aggregate things astatine once: notation aggregate documents, oregon thing much analyzable than a one-off task." Structured prompts are utile for tailored formatting, hyperlinks, and presentation. Structured prompts are besides the instauration for the adjacent 2 higher-level categories discussed, projects and agents. 
• When to employment projects: Projects enactment champion for repeated-use tasks. "Choose a task erstwhile what you request is simply a predictable result each time," said Noles. This attack works "where power and transparency matter. The steps are close there, your instructions are close there. It tin beryllium updated rapidly connected the fly. If thing isn't working, you tin spell successful and marque tweaks truly easy without having to log into immoderate different bundle and propulsion up a node wherever you person to spell successful and marque changes." For example, "if you cognize that we alteration however we bash reports each 30 days, a task is astir apt close for you due to the fact that you tin spell in, propulsion 1 out, driblet different successful -- 20 seconds of your day."
• When to usage agents: Agents are required erstwhile autonomy is portion of the system. They travel successful to grip complicated, asynchronous tasks that necessitate tools and time. "Use an cause erstwhile the task requires autonomy, requires dynamic determination and instrumentality use," said Harvey. "You request the AI to conscionable spell and fig it retired itself. I can't beryllium babysitting it the full time; I request it to conscionable bash this for me. I request it to beryllium capable to marque aggregate decisions connected the alert connected its own. And I request you to beryllium capable to usage tools similar MCP connectors truthful that you tin spell retired and bash it."

Gen AI has the imaginable to beforehand productivity and innovation if employed smartly. Understanding the 4 stages of deployment and the effort required tin forestall gen AI from becoming a counterforce to the productivity and innovation the exertion is expected to deliver.