Latest / Excel at AI in Finance | CFO, FP&A, AI Tools, Claude, Microsoft Copilot / Microsoft Copilot Agents Are Already On. Does Your CFO Have Rules for Them?
Transcript
- 0:00If your company set up a Microsoft Copilot spending policy this fall, there's a setting
- 0:04inside it called auto-apply new services.
- 0:07It's on by default,
- 0:09so every new Copilot agent Microsoft ships gets added to that policy automatically.
- 0:14This includes the ones that hasn't shipped
- 0:16yet, and that's happening at companies that already have written AI policies
- 0:20and ones that don't.
- 0:22Ernst & Young asked 202 big US
- 0:24companies about AI this fall.
- 0:2698% of them say they have a formal AI policy,
- 0:28but 47% have skipped it at least
- 0:31once to push an urgent rollout through.
- 0:34So today I want to talk about AI tools moving faster than
- 0:36the rules your company wrote for them.
- 0:39And I'll tell you what I'd want written down for a finance team before the next close.
- 0:43Welcome to Excel at AI in Finance, the show for
- 0:47finance pros who want to get home sooner and be
- 0:49the ones their company can't afford to lose.
- 0:51I'm Mike, I run F9 Finance, and I'm a senior finance leader working in the trenches.
- 0:56Each week, I share the AI tools, workflows,
- 0:59and mindsets that get you where you want to be.
- 1:02All right, let's get into it.
- 1:04Now, I've seen this happen in my own career.
- 1:07AI development is moving so fast,
- 1:09features keep popping in left and right.
- 1:11One of the first things that came to mind
- 1:13as I started thinking through this was the first time I opened Copilot after an update.
- 1:18Previously, you opened Copilot and there was just the chat window and you would just chat
- 1:22with it.
- 1:23And then one day, all of a sudden, Cowork appeared.
- 1:25The button was there.
- 1:26I had started using Claude.
- 1:28I knew what Cowork was.
- 1:29And I'm like, wow, Microsoft Copilot Cowork.
- 1:31That's cool. Click over to it and start playing around with it.
- 1:36Because it was in Copilot and because Copilot was approved, it never really crossed my
- 1:41mind if this was or wasn't on the AI policy because Copilot was our approved tool.
- 1:44Well, a couple of days later,
- 1:46we all get an email that Copilot Cowork
- 1:49actually costs credits every single time you run it and isn't included in our licenses.
- 1:54Now, nobody was mad at anybody.
- 1:55It was just, it was just a miss on everybody's part, but it snuck past
- 1:59the IT reviews and the IT policy.
- 2:01At that point, they locked it down.
- 2:03You had to, you know, put a business case together, request it.
- 2:06But it was a miss on that side that for a couple
- 2:09of days, it snuck through because it showed up
- 2:11so fast and it was inside an approved tool.
- 2:15Now, after this, when I first got access
- 2:18to Claude to formally use inside,
- 2:20you know, a company's secure data setup,
- 2:22there were all of the connectors available,
- 2:24including a connection to Microsoft.
- 2:26And after the Microsoft experience where,
- 2:30you know, Cowork seemed like it was usable,
- 2:32but wasn't quite ready to use, I wasn't sure if
- 2:35I could start using those connectors. They were available. They looked ready to use.
- 2:39But after that learning experience,
- 2:41it gave me the pause.
- 2:43The challenge is not everybody's going to pause.
- 2:46The challenge is we're not going to be able
- 2:48to catch every single feature coming through because so many
- 2:51of them are being auto-applied.
- 2:53The catch is that new models are
- 2:55coming through and the new models
- 2:57might become the default
- 2:59and start running existing processes.
- 3:01There's just really
- 3:03no way to keep up with all of that and make sure that every single thing is caught.
- 3:07And there's certainly no way that our AI
- 3:09policies that require multiple rounds
- 3:11of committee reviews
- 3:13and sign-offs and all of that are going to be able to keep up with our policies.
- 3:17And that's what I want to talk about in today's episode.
- 3:20This episode is sponsored by Octopus AI.
- 3:22Every quarter, somebody asks why a
- 3:24line came in over budget.
- 3:26Half the time, you already explained it, and the answer is sitting
- 3:29in an old deck somewhere.
- 3:30Octopus saves those explanations and the decisions that came with
- 3:34them right next to the numbers.
- 3:36It brings business input into your rolling forecast for your team to review and approve,
- 3:40so there's less chasing people for the same answers.
- 3:42You'll find a demo link in the show notes.
- 3:45So much of AI policy conversation is still
- 3:47about whether people should be using a specific tool at all.
- 3:50Now, that part's mostly settled.
- 3:52Teams are using AI.
- 3:53We've got that down.
- 3:55What changed this
- 3:56year is the tools can act now.
- 3:58They have more connectors, they open workbooks, they make edits,
- 4:01they can run on schedules.
- 4:02So first I want to walk through what's switching on and why the
- 4:06policy you have doesn't cover it.
- 4:08Then what I would do about it this month and every month going forward.
- 4:11Just like my experience with Copilot Cowork, agents are now arriving
- 4:15as defaults, often before anyone on your team
- 4:17signs off on them. Starting with Microsoft, since so much finance work already runs
- 4:21through Excel and Teams.
- 4:22Microsoft's own admin page says
- 4:25the agent capability is enabled by default
- 4:28in every tenant with a Copilot license.
- 4:30Admins can block agents or remove them,
- 4:33but somebody has to go in and do it.
- 4:36Now to be fair, the big one, Copilot Cowork,
- 4:39now stays off until an admin turns it on.
- 4:42Cowork is the agent that does the work itself.
- 4:44Once it's on, it can edit your existing Excel, Word, and PowerPoint files in SharePoint and
- 4:48OneDrive. So can that Claude connector I mentioned. It can work in the browser
- 4:52using your own sign-ins, and it can run tasks
- 4:55on a schedule, in Microsoft's words, without a person present.
- 4:58So how does it get turned on now?
- 5:00Well, since October, anyone who doesn't
- 5:03have access can open Cowork and request it.
- 5:05That was the step that was missing the first time I went in.
- 5:08And the default setting for who's allowed to ask is all users.
- 5:11Then there's the spending policy from the top of the show.
- 5:15Auto-apply new services is on by default.
- 5:17So the agents Microsoft ships next quarter are already in it.
- 5:21And Microsoft says a policy with a limit of one credit.
- 5:24Just one credit still grants access.
- 5:26So the budget works as the permissions
- 5:28slip, even when it's tiny.
- 5:29Now Claude goes further.
- 5:31On Claude's team and enterprise plans,
- 5:34Cowork, their version of
- 5:35the same idea, is on by default,
- 5:37and an owner has to actively switch it off.
- 5:41And on November 16th, for existing enterprise customers,
- 5:44Claude's chat and Cowork become one
- 5:46screen by default, unless an admin keeps it
- 5:49off. Google's agent builder in Workspace follows whatever release setting
- 5:53your company picked.
- 5:54So it can show up on or off.
- 5:56At least the choice is made obvious.
- 5:58All of this is written down in the vendor's admin docs.
- 6:01It's not being hidden.
- 6:02But IT reads the admin docs, and the files
- 6:05these agent edit belong to finance.
- 6:07So for a lot of finance teams, the first decision about these agents comes after
- 6:11they're already on the screen.
- 6:13Nearly every big company has an AI
- 6:15policy, and half of the ones running agents haven't updated it for them.
- 6:19It's the same EY survey from the top of the show.
- 6:21They talked to 202 senior
- 6:23AI decision makers at U.S.
- 6:25public companies over a billion dollars in revenue. This was back in May and June.
- 6:29It's a small sample, and EY sells
- 6:31governance work, so keep both in mind.
- 6:33You already heard the 98% with a policy,
- 6:36and the 47% who skipped
- 6:38it to rush something out.
- 6:40The number I'd circle is a different one.
- 6:4291% of those companies are
- 6:44using AI agents in a pilot or for real.
- 6:47And 49% of them haven't
- 6:49updated their governance for agents at all.
- 6:5185% are running agents that
- 6:54act without a person involved in real time.
- 6:56So the policy was written for a person typing a question and reading the answer.
- 7:01And the tool on the desk now opens the file and makes the change.
- 7:04Then there's the timing.
- 7:07Almost all of those companies, 98%,
- 7:09review their AI at least once a year.
- 7:11In the 23 days between
- 7:13September 15th and October 7th,
- 7:15Anthropic put out four model launches.
- 7:17Microsoft's own Cowork change log
- 7:20shows eight new model options in six months.
- 7:22So a once-a-year review is
- 7:24reading last year's product.
- 7:26To be fair, policy is moving.
- 7:28ISACA surveys over 3,000
- 7:31IT and risk professionals every year.
- 7:33Companies with a formal comprehensive AI
- 7:36policy went from 15% to 38% in two years.
- 7:39That's real progress.
- 7:41It's just slower than the tools.
- 7:43And if you heard episode 13,
- 7:45this is where your auditor comes back in.
- 7:48The first thing they ask for is an inventory, and the agents have to be on it.
- 7:53So the policy is there.
- 7:55It's just answering last year's questions.
- 7:58The part the policy conversation misses is that your team made their
- 8:01own rules months ago.
- 8:02KPMG and the University of
- 8:04Melbourne surveyed over 32,000
- 8:06employees in 47 countries.
- 8:0856% said they've used AI at
- 8:11work without knowing whether it was allowed.
- 8:1357% said they use it
- 8:15without telling anyone, and some of
- 8:18them pass off what the AI wrote as their own.
- 8:21And 48% have put company information into
- 8:23public AI tools, with financial information
- 8:26one of the kinds they named.
- 8:28That survey ran before agents were common, by the way.
- 8:32Workers in finance and insurance reported some of the highest AI use in the whole study.
- 8:36Now, put that next to how these agents
- 8:39keep their records.
- 8:40When a Cowork task runs, Microsoft says it
- 8:43runs as the user who created it.
- 8:45So whatever the agent changes in that forecast file, the log shows your name.
- 8:50And when it asks for approval, one of the options is always.
- 8:53So one click can approve that
- 8:55action for every run after it.
- 8:58When this goes wrong, it can cost real money.
- 9:01IBM's breach report this
- 9:03year looked at shadow AI, meaning AI tools the company never reviewed and approved.
- 9:06It showed up in 43% of security incidents,
- 9:10up from 20% the year before.
- 9:12And those breaches?
- 9:14They averaged 5.39 million.
- 9:16And that's only companies that had a breach.
- 9:19And IBM also sells security,
- 9:21so weigh it that way.
- 9:23At those same companies, strict approval for
- 9:26AI rollouts fell from 45% to 38%.
- 9:30So the obvious answer is to ban it,
- 9:32but the KPMG data says that backfires.
- 9:34Breaking the rules was most
- 9:36common where AI was banned, 67%.
- 9:39And lowest where there was no policy at all, 33%.
- 9:43So a ban pushes people onto their personal laptops, and a policy that never mentions
- 9:47agents just leaves them guessing.
- 9:50So what should you do this month?
- 9:52First, I'd ask whoever runs your Microsoft or Claude admin two questions.
- 9:56Who can request an AI agent today?
- 10:00And does our spending policy add new agents on its own?
- 10:04That's one email and IT already knows where to look.
- 10:08Then I'd write one page about agents.
- 10:09It says which files and
- 10:11folders an agent can touch.
- 10:12It says which data
- 10:14never goes in, like payroll detail or anything with a customer's name on it.
- 10:18It says who on the team is allowed to click always on approval.
- 10:23And it says who you call to switch an agent
- 10:25off with a real specific name next to it.
- 10:29Then pair that page with a
- 10:30tool people want to use.
- 10:32One study of 564 consulting and professional
- 10:36service employees found hidden AI
- 10:39use was lower where there was a clear policy.
- 10:41And lower again where the official
- 10:44tools actually did the job needed.
- 10:45So that's two questions and one page.
- 10:49And you can get both done before the next close.
- 10:51Your AI policy was
- 10:53written for a tool you ask questions.
- 10:56And the tools on your team's
- 10:57laptop now do the work.
- 11:00All right, that's my take, but I want to be fair about where it falls short.
- 11:04Now, in many cases, the big enterprise defaults are
- 11:07more careful than I've made them sound.
- 11:09Copilot Cowork usually stays off
- 11:12until an admin turns it on, at least now.
- 11:14And on Claude Enterprise,
- 11:16auto-approving cloud sessions
- 11:18start off by default.
- 11:19A lot of the on-by-default settings
- 11:22sit on smaller plans and then request buttons for medium-sized companies.
- 11:25Now, the policy side is catching up.
- 11:29The ISACA numbers more than
- 11:31doubled in two years.
- 11:32And EY found that after a yearly review,
- 11:3564% of companies made big changes to at least a quarter of their AI systems.
- 11:39So governance works when somebody runs it.
- 11:43Heavy rules can backfire too.
- 11:45One study at a Fortune 500 company
- 11:47followed 20 teams for 28 weeks.
- 11:49The guardrails cut hallucinations
- 11:52in half and took audit trails from
- 11:5453% to 96% complete.
- 11:57But 35% of the teams worked
- 11:59around the guardrails they found confusing or out of proportion.
- 12:02That's one author and one
- 12:04company, so hold it loosely.
- 12:05There's also a case that the
- 12:08shadow use is where the value shows up first.
- 12:10MIT's research found workers at over
- 12:1390% of companies using personal AI tools,
- 12:15while only 40% had bought an official subscription.
- 12:18It's a small sample, but the idea holds.
- 12:21People find the use before the company does.
- 12:24So the lag is partly by design,
- 12:26and a slow policy that learns from
- 12:28real use may beat a fast one that bans it.
- 12:31So here's how you tell which one you've got. Ask your admin who can request an agent today.
- 12:36If they answer in a sentence,
- 12:38your policy is keeping up.
- 12:40If they have to go look it up, there's a really good chance the tools are ahead of the policy.
- 12:45And that's this week's show. If you got something out of this, do me a favor and
- 12:48rate it on Apple Podcasts, Spotify, or wherever you listen.
- 12:51It's the fastest way to help another finance pro find it.
- 12:56And if you want more from me between episodes, the AI library is where
- 12:59I drop the prompts and templates I'm using.
- 13:01I'll also send you my free weekly newsletter, Finance AI Insider,
- 13:04with tips, tricks, and hacks just like this.
- 13:07The links are in the show notes, and I'll catch
- 13:09you over there. Until then, this is Mike signing off from F9 Finance. Cheers.