When Should Your AI Work Fast, and When Should It Think Harder?

Six · 6 min read · May 12, 2026

When Should Your AI Work Fast, and When Should It Think Harder?

Say you run a small consulting firm. The Monday team meeting just ended and your next call starts in ten minutes. Before then, you want the notes turned into a recap people can act on: what was decided, who owns each next step and what's still open.

That's not the moment to ask an AI to spend half an hour weighing every possible reading of the conversation.

The same meeting left two more pieces of work on your desk. A client changed the scope of a project after you sent the first proposal, so the proposal needs revising. And the team spent twenty minutes on a bigger question: should the firm start taking on healthcare clients?

That one deserves a different kind of attention. The sources may conflict. A missing assumption can change the answer. And a polished recommendation can still send the firm toward the wrong bet.

Some AI models can be told how long to think before answering. Developers using Claude 3.7 Sonnet through the API could cap how long it thinks, which Anthropic said lets you "trade off speed (and cost) for quality of answer." Google's Gemini 2.5 Flash came with a thinking budget developers can set to zero when cost and speed matter most.

So for each of this week's three jobs, the question is how much thinking it needs before someone at the firm can use the result.

Start with the work

Teams often pick one setting and use it everywhere. Fast for everything, because people are busy. Or maximum effort for everything, because the work feels important once it's been handed to an AI. Either way, nobody looks at the work itself.

Look at each piece of work through four questions instead. They'll give you a reason for the setting you pick.

Three assignments, three choices

The recap: work fast

The notes are messy, but the decisions were made in the room. The job is to give everyone a clean record while they still remember the conversation: decisions, owners, dates and open questions.

Start with a lower-effort setting, if the model you're using offers one. Give it the agenda, the notes and the format you want. Then whoever ran the meeting scans the names, the commitments and anything that sounds more certain than the conversation was. If people can read it, fix a mistake and act on their next step, the recap did its job.

The proposal: enough effort to protect the relationship

Revising the proposal means more than rewriting a page. Some of the original still fits. Some of the new scope changes the price or the delivery plan, and the client needs to understand those trade-offs before anyone sends it.

The deadline still matters, but so does the cost of a confident mistake. A proposal that promises the wrong timeline keeps creating work for whoever manages the account long after the draft went out.

Ask for more effort here, because the work means comparing the original brief, the new request and the reasons behind each change. Set the review conditions before it starts: the draft should show what changed, what stayed the same, what's still open and who has to approve a new price or date.

Then run a lower-effort version against the same conditions and see which one the person who owns the client relationship can check and stand behind.

The healthcare question: think harder, then widen the review

This one is harder still. Nobody at the firm knows how much demand there is, and the sources disagree. There could also be legal or operational limits that no model can settle by itself.

Give the work room to trace the evidence, separate what's known from what's guessed and say where the case falls apart. But extra reasoning doesn't make the decision for you.

Make the request sharper than "give us a recommendation." Ask for the options it considered, the evidence behind each one, the missing information that could change the answer and the decision the document is meant to support. Then take it to the people who'd live with the result, including whoever would run the work, whoever watches the money and whoever would answer for the risk.

This is where a slower, more expensive run may be worth it. It may also tell you the right next step is to gather more evidence before deciding anything. That's useful work too.

Run the comparison like a manager

You don't need a model bake-off. Pick one assignment that comes back on a schedule and has a real owner. The Monday recap works. Keep the instructions, sources, model and finish line the same. Run it once at a lower effort setting and once at a higher one, using settings the model actually supports, then review both with the person who uses the result.

There's a reason to test instead of assuming more is better. Researchers built tasks where letting reasoning models think longer made them less accurate, including counting problems padded with irrelevant details and deduction problems with constraints to keep track of. They still wrote that longer reasoning "remains promising" for making models more capable, and recommended testing at different reasoning lengths to catch failures like these. Nobody can tell you ahead of time which way it goes for your Monday recap. Running both versions tells you.

Write down what changed:

If the extra effort didn't earn its place, keep the faster path. If it did, save more effort for this kind of work. If neither version meets the finish line, change the brief, the sources or the review before you turn the setting up again.

Where Supanova fits

Say this week's work went to Supanova atoms. The recap, the proposal and the healthcare recommendation would all end up in the same place: finished work comes back for your review. The four questions are how you decide what that review has to cover, from a quick scan of names and dates to a meeting with everyone who'd own the healthcare decision.

You also choose how much your workforce checks in, from approving every step to working more independently. If most of your week looks like the Monday recap, letting atoms work more independently and reviewing what they finish is a sensible trade. In a week when the healthcare research matters most, switching to approving every step is worth the time. A wrong guess about demand early in the research would carry through everything after it, and you'd see that guess before a recommendation rests on it.

Pick one recurring assignment and run the comparison on the model you already use. Keep the winner only if the person who uses the work can tell you why it was better. When you want a workforce to hand that work to, start at supanova.team.

Sources

AI DelegationAI work managementAI reasoningAI reasoning effortAI Quality