Now
01
Survey and interviews both open
Interviews began first and continue on their own schedule. The survey is collecting alongside them. Weekly practitioner polls run as an open thread and are treated as conversation, never cited as research.
Research · In the Field
State of AI in Cloud Operations
What cloud teams will actually let AI do, in their own words.
Vendors have plenty to say about what AI should be doing in your cloud. Almost nobody has asked practitioners what they would actually permit it to do today, and what would have to change for that answer to move. That is the question this study exists to answer.
There are two ways to take part, and they are independent of each other. The survey takes 5 to 7 minutes. The in-depth interviews run 30 to 45 minutes and are already underway. Either way you get the report before it goes public, and the full question set publishes alongside the findings.
5-7 minTo complete
150-250Target qualified responses
15-20In-depth practitioner interviews
Oct 2026Report publishes
Why We're Running It
The interesting number is the one nobody has published.
There is no shortage of research on AI adoption in engineering. Almost all of it measures the same thing: how many people are using an assistant, and how happy they are with it. That question has been answered.
The question that has not been answered is the one that decides whether any of this reaches production. What would a cloud team let an AI actually do in a live environment? Read the account and report back? Recommend a change a human executes? Execute with approval on every action? Act on its own inside a defined blast radius? Where a team sits on that ladder, and what would move it up a rung, is the thing that determines what gets built and what gets adopted over the next two years.
We are building in this space. We would rather know the answer than assume it, and we would rather publish it than keep it.
What We're Asking
Eight questions the report will answer with numbers.
The instrument runs 5 to 7 minutes across screeners, current AI usage, trust, findings and operations, and adoption requirements. These are the questions the answers roll up to.
1
What would cloud teams let an AI assistant do in their environment today, from nothing at all to autonomous action?
2
What is the single biggest barrier to letting AI act on cloud infrastructure: credentials, accuracy, blast radius, compliance, or visibility?
3
Which safeguards would most increase confidence, and by how much do they differ by role and company size?
4
Where do teams want AI help: inside the LLM they already use, embedded in existing cloud tools, in the provider console, or in a separate platform?
5
Which work already starts in an LLM rather than the original tool, and which work would if the access and safeguards were there?
6
Where does AI-assisted work happen today, and where do practitioners expect it to happen 12 to 18 months from now?
7
How many findings and recommendations are sitting unresolved right now, and what happens to them?
8
What does an organization actually require before an AI tool can touch production?
How We're Running It
OpsCanvas is running this study. Here is how you can check our work.
This is vendor-funded research, and pretending otherwise would be the fastest way to make it worthless. What separates useful vendor research from a lead magnet is not who paid for it. It is whether the method is published and holds up. So here is ours, committed to in advance.
- The full question set publishes with the reportEvery question, in the order asked, with its answer options. Wording stays frozen for the entire field window so the numbers stay comparable. You can see exactly what was asked before you decide what the answers mean.
- Every chart carries its sample sizeNo percentage appears without the n behind it. Any cut with fewer than 50 responses is labeled directional and written about that way. At 150 qualified responses a headline stat carries roughly a plus or minus 8 percent margin at 95 percent confidence, and roughly plus or minus 6 percent at 250. We will say so on the page.
- No sponsors and no paid placementNobody paid to appear in this research, to shape a question, or to be named in the findings. OpsCanvas funds it and OpsCanvas publishes it, including the results that are inconvenient for us.
- Survey plus interviews, not survey aloneThe interview arm runs as its own track and started before the survey opened: 15 to 20 recorded conversations with practitioners, 30 to 45 minutes each. The numbers tell you what teams believe. The interviews tell you why, and they are where the report's argument comes from.
- Aggregate reporting onlyIndividual responses are never published, shared, or sold. Quotes appear anonymized to role and company type, for example "Chief Architect, mid-size logistics company", and only where the response is a written answer the respondent chose to give.
- Your email is not the productWe use it to send you the report and to send the thank-you card if you are in the first 100. We do not sell it, rent it, or add you to unrelated lists. Anything beyond the report is a separate box you check yourself.
Timeline
Where the study is right now.
Both arms are collecting now. If you want to be counted, this is the part that has a deadline.
Late Sep
02
Field window closes
The survey closes once we reach 250 qualified responses or the window ends, whichever comes first.
Early Oct
03
Analysis
Cuts by role and company size wherever the sample supports it. Anything thinner than 50 responses gets labeled directional rather than quietly rounded up.
Mid Oct
04
Report published
Everyone who took part receives it by email a week before public release, ungated, as promised at the point of asking.
Two Ways In
Take the survey, sit for an interview, or do both.
The two arms are independent. Plenty of people will only ever take the survey. Some will only sit for an interview, and several of the interviews already booked came before the survey opened. Doing both is welcome and is the most useful thing a practitioner can give this research, but neither one requires the other.
Everyone who takes part, either way, receives the finished report by email a week before it goes public, with no form to fill in a second time.
The first 100 completed, qualified responses each receive a $25 gift card as a thank-you. Qualified means the response passes the screeners, includes a work email, and reads as a real answer rather than a click-through. Cards go out in weekly batches, so it will not land the same afternoon.
This is a defined first-come thank-you, not a drawing and not a prize. If you are among the first 100 qualified responses, the card is yours.
The Interview Arm
The survey gives us the numbers. Interviews give us the reasoning.
Thirty to forty-five minutes with Brian, recorded with your consent, covering what you would actually let an AI do in your environment and why. No script to grind through, no pitch at the end. These conversations are where the report's argument comes from, and they are the part practitioners tell us they enjoy. You do not need to have taken the survey first.
Who's Behind It
Where we sit, so you can weigh what we publish.
The research is run by Brian Kathman, CEO of OpsCanvas, with in-depth interviews conducted personally. OpsCanvas builds Oscar, an AI cloud engineering agent that connects the AI tools engineers already use to their live cloud environment, with read-only defaults and human approval on every change. You can read what we are building on the Oscar Ops page.
That is the commercial interest, stated plainly rather than buried. It is also the reason the trust questions in this survey are written the way they are: the answers change what we build. If practitioners tell us they will not let an agent act under any safeguard we can offer, that is a finding we need more than a flattering one, and it will appear in the report either way.
Nothing about participating puts you on a sales list. The report is the report.
Before You Start
Reasonable questions about a vendor running a survey.
How long does it actually take?
5 to 7 minutes. Roughly 20 questions, most of them multiple choice, with one optional open text question at the end. Logic skips the sections that do not apply to you.
Do I have to give my email?
A work email is required to qualify, because it is how we send you the report and how we keep the sample honest. It is used for the report and the thank-you card. Everything else is opt-in on the last screen.
Will my individual answers ever be published?
No. Results are reported in aggregate. Written answers may be quoted anonymized to role and company type, and never with your name, your employer, or anything that identifies your environment.
Is this research or is this marketing?
Both, and we would rather say so than perform neutrality we do not have. OpsCanvas funds it, publishes it, and will run a campaign around it. What makes it research rather than a brochure is that the question set is published, the sample sizes are reported, no result is suppressed for being inconvenient, and nobody paid to be in it.
Who is the survey for?
Practitioners and leaders who are responsible for cloud environments: cloud and platform engineers, architects, DevOps and SRE, FinOps and cloud cost, security, and engineering or infrastructure leadership. If you do not own or operate cloud infrastructure, the screeners will route you out quickly rather than waste your time.
Can I see the results before they are public?
Yes. Everyone who takes part gets the finished report by email a week ahead of public release.
I would rather talk than click boxes.
Then talk. The interviews run 30 to 45 minutes, are recorded with your consent, and are the part of this research that produces the sentences worth quoting. They are a separate track from the survey and do not require it, so you can go straight there. Check the box on the survey's last screen if you are already in it, or use the volunteer button above if you are not.
Someone has already invited me to an interview. Do I need to take the survey too?
No. The two arms are independent and the interviews started first. Doing both is genuinely the most useful thing you can give this research, since it puts your numbers next to your reasoning, but the interview stands on its own and so does the survey.
Both Arms Open
Five to seven minutes, and you get the report first.
The trust question in cloud operations has not been measured. It is about to be, and your answer is one of the ones that decides what the number says. Five to seven minutes in the survey, or thirty in your own words.
Everyone who takes part receives the report before public release. First 100 qualified survey respondents receive a $25 thank-you card.