I’ve said it before and I’ll say it again, Microsoft are always hiding the good stuff away. 

I blogged previously about how you can make the answers that you get from Copilot more personalised by giving it information to remember about you. This blog is about how you can tune the amount of effort Copilot puts in to make its answers truly ‘yours’. 

So let’s see what’s now hiding under the dropdown just quietly hanging out in the top left. This has changed even since I began this blog series: 

What does the Auto dropdown in Microsoft Copilot do?

What happens when you click on the ‘Auto’ dropdown is you get these different response options:

• Auto 
• Quick response
• Think deeper 
• And an ‘experimental’ Advanced reasoning

Here you can choose how much ‘thinking effort’ Copilot will put into answering your questions. (You’ve also got your model picker there – you can find out more about those in R is for Researcher where I talk about the model council option).  

Auto is supposed to ‘decide’ for itself how much effort to put into answering for you. Quick response does what it says on the tin, and think deeper is supposed to, well, ‘think’ deeper. (Although I think that’s a pretty misleading term…anthropomorphising dangers and all that.) 

I wanted to know how the different ‘effort modes’ would compare in terms of quality of response.

The experiment: same question, three settings

I asked Copilot the same question three times, once on each effort setting: Quick response, Auto and Think deeper. (I left out the experimental one because it made me think of science labs where things explode…) 

The question was deliberately ordinary, with no ‘correct’ answer, so that I could see Copilot’s shades of grey between each mode:

“Give me an onboarding plan for a junior software developer.”

Quick response and Auto: the twins

Quick and Auto must have been buddies, both came back with almost the same answer. Both gave a 90-day plan in four phases, with tidy green ticks, a mentor, and weekly check-ins. They even opened and closed on the same idea: onboarding is about confidence, not speed.

Both were okay, although pretty generic. 

Quick said the new starter should have their first piece of code merged by the end of week one. Auto said weeks three to four. That’s a fortnight’s gap in expectations for a new starter, without any real reference to our actual product and what they’d be building in real life. 

I’d also expected the Quick response to be, well, quick. Or at least, quicker to read.  It wasn’t noticeably faster than Auto, and it wasn’t any shorter either. Its ‘quick’ party trick was grouping the success measures into a neat little 30/60/90-day table, which I liked. Easy to scan, easy to steal for one-to-one meetings:

Think deeper: the one that looked

Think deeper took longer. And it looked at how we really do things around here. 

It searched files, meetings, emails and chats, and came back knowing who we are. It named Empowering.Cloud and our ChangePilot product, and it found our internal software development guidelines. It then built the onboarding plan around them instead of offering generic advice. 

It also covered the grown-up stuff the others skipped. That included sorting access before day one, not giving a junior unnecessary access to production systems or customer data, and thinking about customer impact and how to roll back a change on every task. It also told me which bits came from our internal guidance and which came from external best practice.

(It also claimed “one thing I’ve seen work well with junior developers…” Seen? Where?? Copilot, you’ve never met a junior developer in your life.)

Who do I ask?

What I’d hoped it would do, and it didn’t, was tell my new starter who to go to. It suggested “a line manager” and “a technical buddy”, but no names. I’d have thought it would be able to guess given the amount of data it had trawled through. 

So I asked it directly. It correctly suggested our Technical Product Owner as the go-to for day-to-day coding help, on the basis that he owns our development guidelines. It also suggested other colleagues for product and customer context, and sensibly warned against one person becoming the only source of knowledge.

Then I noticed who was missing: one of our co-founders, who is absolutely key for product positioning. Not a mention. Copilot can only join the dots that exist in your work data. If someone’s influence lives in desk-side chats and untranscribed calls rather than documents and meetings, as far as Copilot is concerned, it doesn’t exist.

Copilot’s map of who does what is only as good as the trail your colleagues leave. That’s definitely a blog for another day…

Side by side

	Quick response	Auto	Think deeper
The plan	90 days, 4 phases	90 days, 4 phases	90 days, plus pre-boarding
Speed	Not much faster than Auto	About the same as Quick	Slowest
First code merged	End of week 1	Weeks 3–4	Tiny change in week 1, full workflow by end of month 1
Asked me questions first?	No	No	No
Used my work data?	No	No	Yes: files, meetings, emails, chats
Knew our company and product?	No	No	Yes
Named who to ask?	No	No	Only when asked, and it missed a co-founder
Security and data access	A line on “security principles”	A line on “security basics”	Built into every stage
Showed its sources?	No	No	Yes, internal and external
Best for	A generic starting point (with a handy summary table)	A generic starting point	A plan you’d actually use

What I took from it

The difference wasn’t how much each setting wrote. Quick and Auto had access to exactly the same work data as Think deeper. They just didn’t go and look.

In my test, turning the dial up didn’t just make Copilot think harder. It changed whether Copilot bothered to find out about my world at all. The lower settings gave me an onboarding plan. Think deeper gave me our onboarding plan.

One caveat: none of the three asked me a single clarifying question before answering. That still seems to be my job. (See Q is for Questions.)

TLDR

There’s a dropdown in Copilot Chat that lets you choose how hard it thinks: Auto (it decides for you), Quick response, and Think deeper (plus an Experimental “Advanced reasoning” I left alone). 

In my test, turning the dial up didn’t just make Copilot think harder. It changed whether Copilot bothered to find out about my world at all.

Whichever setting you pick, check what comes back, because even Think deeper forgot one of our co-founders existed.

Have you played with the effort settings yet? Did turning the dial up change the answer you got, or just make you wait longer for it? Let me know in the comments.