How Free AI Tools Become Expensive Mistakes

Every AI tool has a price, and the free ones are paid for in something other than dollars. Michael Moore has spent his career on both sides of that kind of trade, first as an engineer building products and then as the lawyer protecting them, and on this episode he put it in one sentence.

“And whenever the tool is free, that is, whenever there’s no dollar cost, your data and your personal information is usually the payment, because nothing is really free.”

Michael Moore, at 8:24

Michael is the general counsel and company secretary of Glean, the enterprise AI company Arvind Jain founded in 2019 (Glean). When we recorded, his title was vice president and head of legal, and he described Glean as “an enterprise search and agentic AI company.” Before Glean he was chief privacy officer of Lacework, the cloud security company, which is where he was the first time he came on the show, on episode 41 in April 2024. This time he tells Sasha and me what he expects from any AI product, why the free tools turn out to be the expensive ones, what legal teams have to learn next, and why legal and security belong on each other’s speed dial.

Where we sit, so you can weigh what follows: YSecurity helps startups get through the security and AI reviews enterprise buyers run, from the first questionnaire to ISO 42001 certification. So we spend our days on the seller’s side of the table Michael describes, getting founders ready for buyers who think the way he does. Sasha and I hosted. We recorded in late May 2025, about a week after Glean held its first user conference, and the episode aired on July 15, 2025.

From engineer to lawyer

Sasha opened with the question his career raises: “At what point in time did you start thinking about tapping into the legal side of business ventures?” His answer starts with building. “So from my own just personal journey, I would say, as an engineer, I was very involved in building and developing new products.” And then, “I’ve always been a builder in whatever area I’ve worked in, be it technology or law.” As he built products, he kept asking how to protect them, and that question pulled him toward law.

“And that actually really brought up an interest in me in how do I blend the creativity of engineering with the protection of that creativity through legal.”

Michael Moore, at 1:18

His LinkedIn traces the path in order. He spent four summers as an applications co-op at Analog Devices from 1994, then joined Cypress Semiconductor as an applications engineer in 1998, working on data communications and programmable logic. In 2002 he became a patent agent at Cypress, preparing and prosecuting the company’s patents. Law firm work followed at Morgan Lewis & Bockius in Palo Alto from 2007, then in-house intellectual property roles at Symantec from 2009 and at Rambus from 2011, where he became vice president of IP and deputy general counsel. At Pure Storage, from 2016, he added privacy and sales compliance to products and patents, and his LinkedIn lists The Recorder’s Legal Department of the Year awards for the legal team there in 2017 and 2019. At Lacework, from 2021, he ran privacy and IP, became chief privacy officer, and led the privacy program through Fortinet’s acquisition of the company in August 2024 (LinkedIn). He studied law at Santa Clara University School of Law, and his LinkedIn headline lists the IAPP’s credentials in AI governance and privacy.

Every step on that list is a version of the same job: understand what engineers build well enough to protect it.

One problem, three protections

The questions he asked as an engineer line up with the three main kinds of intellectual property. “How do we protect the code via copyright? How do we protect engineering designs via patents? How do we protect our marketing and brands via trademarks?” To Michael they aren’t three separate fields.

“They’re all facets of the same problem we’re trying to solve.”

Michael Moore, at 1:41

The coder writing software, the engineer building chips or mechanical designs, the creator of copyrighted works of art, the musician: “all of these are different aspects of human creativity.” His conclusion is the thread through the rest of the episode. “And the creativity involved needs to be protected and rewarded.” The work he enjoys, he said, is helping creators “expand the market for their creativity and protect their creativity.”

For a startup, all three show up in the first year, usually before anyone thinks of them as legal questions. Does the company own the code its contractors wrote? Is the core invention worth a patent? Is the name on the website yours to use? None of it is security in the narrow sense, and all of it is what an acquirer’s lawyers, or an enterprise buyer’s, will eventually ask about.

A common language

Sasha saw the advantage in that mix: “It puts you in a very interesting position, because you can speak both legal and engineering languages.” Michael agreed and named the problem. “So, for example, lawyers are well known for speaking legalese.” Engineers have their own jargon, “which can be off-putting to folks who are not of technical background.” His job, as he describes it, is translation, understanding “the tech and the business and the market and the concerns of users and any concerns of vendors, suppliers” and bringing them all “to a common platform we can all understand and communicate.”

“Because good communication and clarity really helps unblock many of the issues that slow down deals.”

Michael Moore, at 3:17

Sasha put it more simply: “Communication is the key. Whether it’s life, business, work, communication is the key.”

From where we sit, this is where a lot of startup deals actually stall. A buyer’s questionnaire asks about data subjects, subprocessors and retention, an engineer answers in services, buckets and queues, and the deal waits while somebody translates. The founder who can say what the product does with customer data, in the buyer’s words and with the evidence attached, gets through review faster. We wrote the playbook for that in what a DDQ is and how to answer one.

Both sides of the table

Michael’s next point came from his own procurement work. “I’ve been both on the buy side and the sell side of enterprise software.” He has done “a lot of buy side procurement work, vendor work,” and “a lot of enterprise sales side also.” On the buy side he learned what a buyer is protecting: “the privacy and security of their data” and “the confidentiality of their business.”

“So when you’re selling a product to an enterprise, you need to be able to put yourself clearly in the buyer’s shoes and understand what is important to the buyer.”

Michael Moore, at 4:04

Then the three questions he asks as the seller: “How can I make my product better for the buyer? How can I make my deal and transaction process better for the buyer? How do I reduce friction wherever possible, and get to a point where it’s easy and clean and smooth for us all to transact together?” And the whole job in three words: “Just unblocking the process.”

A year later he put a number on one piece of that friction. In the second part of a Law.com series on legal AI, he wrote that the teams that deploy AI well start with high-volume, structured work like privacy and security questionnaires. They build agents that pull out the questions, query a curated bank of approved answers and draft fully cited responses for a person to review. By his account, that turns a questionnaire of more than 100 questions, hours of work, into roughly a 20-minute workflow.

Why Glean

I asked why he picked Glean. The first reason was people, team members “who are just awesome,” some of them folks he had worked with before. The second was the product: “I’ve personally used it.” The third is the one that matters for everyone else.

“And the key thing is it puts AI in the hands of employees.”

Michael Moore, at 5:01

He has seen the opposite approach, companies where “AI was sort of this sacred thing and guarded by a very limited number of people who wanted to keep control on it.” In his words, “It was almost like this was the new special toy, and some companies didn’t want regular users, regular staff, to have access to this toy.” Glean’s approach, he said, is “to make the tool available to all the users, and just make the jobs of everyday users so much more fun, productive and enjoyable.”

Then he described the problem from the inside. Before Glean he spent time at “a very large multinational company,” a very good company by his account, “But it was very, very siloed, and trying to find anything was almost impossible.” The cost is time and morale. “When you’re spending 20 or 30 percent of your time just trying to find the information you need to do your job, that tends to reduce productivity.” And when people “have to do sort of drudge work just to find the basic information they need to do the job and be productive, that is a demotivator.”

The security side of putting AI in everyone’s hands is access. A tool that searches everything a company knows is safe when each person sees only what they could already open, which is why permissions come first on the minimum bar Michael sets below. Glean’s own security page goes a step further for agents, saying an agent “should not automatically access everything the user can” (Glean). Checking exactly that, before and after an AI rollout, is what our AI access control audits do.

The proudest day

Sasha asked about the proudest day so far. Michael picked Glean:GO, “our first ever user conference, which was a huge success.” It had happened “up in San Francisco about a week or so ago. Huge success, sellout.” Glean’s announcement put the conference at the Fort Mason Center on May 20, 2025 (Glean), which dates our recording to late May.

He worked with the marketing team on the content, the PR and the planning, “and the whole thing [ran] like clockwork.” And he worked with the engineers on the other side of the launch, “to protect the fantastic innovations they launched at the Glean:GO event.”

“And it’s personally very rewarding to me to see engineers, both seasoned and also early career engineers, get recognition for their amazing innovation through filing patents and other means to protect the IP.”

Michael Moore, at 7:13

He called it “a way of acknowledging and recognizing the very hard work that they’ve done to bring such an amazing product to market in quick time.” Decades after the applications engineer started asking how to protect what he built, he protects what other engineers build.

The minimum bar for an AI product

Sasha asked the procurement question every AI founder is living with: “Is there a minimum bar that every AI product should be hitting, or should have, in order to go through that procurement process quite easily?” Michael answered from the buyer’s chair.

“Any AI product that we are selling, I look at it from the perspective of, if I were the buyer, what would I be asking for?”

Michael Moore, at 7:48

His list, in his order. First, “very strong [controls or] permissions around the user or customer data put into the AI tool.” Second, “The key thing is protecting the personal data information of data subjects, of individuals, protecting confidential information, protecting trade secret information.” Third, “very clear guardrails against training using that user’s data without their knowledge or consent.” He also asks what kind of communication he would expect, which matters as much as the controls, because a buyer can only rely on what the seller writes down.

That list is close to the first page of the AI section our clients get from enterprise buyers. “Do you train on our data?” is question one in the AI security questions your first enterprise buyer will ask, and an answer a buyer can trust has three parts: a sentence in your terms, provider settings that match it, and access controls you can show. A management system certified to ISO 42001 is how a company shows the whole list is run on purpose, and Glean’s security page lists ISO 42001 next to SOC 2 Type II and ISO 27001 (Glean). Our ISO 42001 work builds that system for startups.

Michael made the same argument to law students at Santa Clara University in March 2026, calling ISO 42001 the “how” that turns principles and risk frameworks into practice an auditor can check.

Nothing is really free

The third item on his list is the one this page is named for, and the clip at the top of the page.

“I would also expect very clear guardrails against training using that user’s data without their knowledge or consent.”

Michael Moore, at 8:13

Then the market reality: “There’s a lot of cheap or free AI tools out there on the market.” When there’s no dollar cost, the user’s data is the payment, and on the consumer side that data can be very personal. What gets used to train the product can include “the consumer’s personal information, habits, their preferences, even their appearance, or sometimes their thoughts and their emotions, depending on what they put into the tool.” He doesn’t hide his view of it: “That’s something I personally dislike.”

He said it as a parent, too.

“As a parent, I think that’s a scary concept, when the data of the individual users is being used to train the product.”

Michael Moore, at 8:52

His answer is education, “particularly younger users,” so they learn to “protect their personal information, protect their online activities, and avoid that being used as training material without their knowledge or consent.”

For a founder selling AI, his rule cuts both ways. It’s the question your buyer’s legal team will ask you, and it’s the question to ask every model provider and tool your product depends on, because the terms on a free or consumer plan are often not the terms on a business plan. Read the terms for the plan you actually use, and make your settings match the sentence you send the buyer.

Enterprise data, enterprise tools

Then he moved from consumers to companies: “I know some businesses are trying to use consumer-grade tools for business applications, and that’s something I think is not a wise approach.” The risk he named first was to intellectual property, privacy and confidentiality.

“If you’re using enterprise data, you should be using enterprise tools.”

Michael Moore, at 9:23

His example was one most of the industry remembers: “We may remember in 2023, there was a Samsung ChatGPT code leak that was widely publicized.” In April 2023, Samsung employees put sensitive internal data into ChatGPT, and the company restricted generative AI tools on its devices soon after, concluding that data on an outside server is hard to retrieve and delete (TechCrunch). KED Global reported that the ban covered Samsung’s device division, after an engineer uploaded internal source code (KED Global). Michael’s reading is about how it happened.

“That’s an example of where individual employees bypassed company controls to go use consumer tools to put their company confidential information in, and with resulting negative PR for the company and probably negative consequences for the employees.”

Michael Moore, at 9:40

What he pushes for at Glean is the other column, “a secure platform that is intended for business use, with all the security and privacy structure and permission controls in place to satisfy the most discerning and security-aware business users.”

The rest of the Samsung story arrived three years later. In June 2026, Samsung Electronics said its device division would bring external generative AI services into everyday work, after testing the candidates with about 2,500 employees, and paired the rollout with a security framework meant to deliver both “expanded AI utilization” and “risk control” (CIO.com). The ban turned out to be a bridge to approved tools, which is the pattern in our shadow AI guide: give people a safe path that’s easier than the unsafe one.

Finding the consumer tools already in use comes first, and Michael has written about doing it without a survey. In February 2026 he described how Glean built its own AI inventory by joining records it already had, from finance and procurement, identity provider assignments and existing governance documents, then using search and AI to stitch them together.

That’s the same doorway our shadow AI detection work starts from: find every AI tool touching company data, keep the good ones, and give them a paved road.

Legal teams will need to embrace AI

I asked him to look five years out. He answered for legal teams, and he was candid that some of them are hesitant, “sometimes in regulated industries, sometimes they’re in markets that do not prefer to use the newest tools.”

“But regardless of that, legal teams will need to learn and embrace AI.”

Michael Moore, at 10:39

The pressure comes from whoever pays the bills. “So outside counsel law firms generally bill by the hour, and those hourly rates are expensive, and they’re generally only going up.” Then, dryly: “I’ve never seen them go down.” In-house legal teams are the law firms’ customers, and Michael expects them to “demand that their outside law firms use more efficient tools,” and to make sure “the billable hours are used as effectively and as efficiently as possible.”

The argument reaches past law. “Any professional services role, so think of any roles that traditionally have been hourly based, that involve information or knowledge processing, could greatly benefit from a tool that makes access to information easier.”

In 2026 he published what he had learned from more than 100 conversations with general counsel, chief legal officers and legal operations leaders. He expected accuracy to be the main barrier, and wrote that he was wrong: most legal teams are blocked by organizational readiness, not by technology.

A month later, at a session in San Francisco with David Ambler of Paul Hastings, he made the billable-hours point from the other side: legal’s core problem is capacity rather than tooling, and much of the routine, repeatable work is moving in-house with AI.

Grounded in the source

The second half of his five-year answer is about trust in what an AI writes. The tool that helps knowledge workers is one that “makes the integrity of the information clearer by citing the source documents in every case.” For lawyers that part isn’t optional.

“being sure that the content that your system or LLM is creating is grounded in fact and truth, and being able to validate the source document very quickly and easily will help improve confidence and acceptance by users”

Michael Moore, at 11:50

He meant “users who really need to be sure that what they’re creating or stating is true,” and lawyers are the clearest case. Legal teams have learned it in public, from court filings that cited cases nobody could find, which is why “How do we know a citation is real?” is one of the legal questions in our list of what enterprise buyers ask AI vendors.

For a startup building AI features, citations are also the cheapest trust you can ship. A buyer who can click from an answer to its source doesn’t have to take the model’s word for anything, and a reviewer who can check an answer in seconds actually checks it.

What a CEO should look for

I asked him to picture me as the chief executive of a big company, or a founder, who knows AI is how to stay competitive and wants people focused on “the meaningful tasks that bring them the most joy and the delight.” My question: “How can I, as a leader in this imaginary space, best support a legal team through those challenges that you were talking about?”

“For CEOs and founders, if you’re picking a tool, make sure you’re picking a tool that has security and privacy built in from the get-go, has very strong access control and permissions for any LLM content generated, that is tied very clearly to the source documents and that can be validated by a human very easily.”

Michael Moore, at 12:46

Then the fifth item, the one most security checklists leave out: “And frankly, it’s something that your employees enjoy using.” Adoption is the point. “There’s lots of tools out there that are kind of clunky and maybe somewhat effective, but if people don’t adopt or use them, it doesn’t quite get the results you need.” For a company trying to become AI-first, which he called “a buzzword or a hot topic,” his advice is to “start by making the tool easy to use, easy to enjoy, easy to access.”

Turn the list around and it’s the buyer’s checklist for your product. Each line maps to evidence you can hand over: a note on how security and privacy are built in, a permissions test, an answer with its citations, the review step, and adoption numbers that show people use it. The scorecard below runs the list against your own product.

Your lawyer on speed dial

Sasha asked how security and compliance teams can support legal through the move to an AI-first company. Michael started with prevention: “I do think legal and security should and must partner closely to prevent issues from occurring.” That takes “a combination of technical controls and internal process controls,” and then detection and mitigation, “particularly in a ransomware type scenario.” In his words, “You want to prevent it from happening, but if it does happen, you want to detect it the instant it happens and mitigate it very, very quickly before it spreads.”

Then the rules that make it urgent. He pointed to “some of the regulations in place requiring companies to report security incidents in the US or in Europe,” and said they call for “a very, very tight coupling of the security org and the legal org.”

“I mean, you should always have your CISO on speed dial. The CISO should always have their lawyer on speed dial.”

Michael Moore, at 14:47

Two of those clocks are short. Under the SEC’s 2023 rules, a public company’s Form 8-K disclosure of a material incident is generally due “four business days after a registrant determines that a cybersecurity incident is material” (SEC). Under GDPR, a controller notifies its supervisory authority of a personal data breach, where feasible, “not later than 72 hours after having become aware of it” (GDPR Article 33). Neither clock waits for the two teams to be introduced.

His closing image is the one to remember: “So if the stuff hits the fan, you’re ready to go. You know each other, you have a plan, you have a playbook, and you’re ready to roll.”

He had written about the ransomware case before. In March 2024, with Lea Kissner and Merritt Baer, then Lacework’s chief information security officer and field chief information security officer, he published an article for in-house counsel at the Association of Corporate Counsel on ransomware as a threat to reputation, privacy and compliance. It walks through double extortion, the privacy laws a leak can trigger, and why paying a ransom can expose the victim to US Treasury sanctions enforcement, and it lands on prevention: “an ounce of prevention is worth more than a pound of cure” (ACC).

Most startups meet this the day it happens, and it’s far cheaper to do beforehand. Write one page that says who calls whom, which clocks apply to you, which outside counsel you’d call and what you’d tell customers, then run it once with both teams in the room. Our 24/7 monitoring and response team works from that page when something fires, and the blast radius playbook covers the containment side.

An introvert who decided to speak

Sasha asked the question we ask everyone: “If you could meet younger self, would you? And if you would, would you give yourself an advice?” Michael started with who he is. “So, look, I’m an introvert by nature. Many of us who are of technical or engineering backgrounds often are.” Early in his career he “wasn’t as active in networking or getting out and speaking and publishing, and just wasn’t as comfortable getting in front of people.” Then he made a decision: “look, I’ve got to do this. I’ve got to get out in front of people and speak, and be comfortable speaking and putting my voice and opinions out there.”

“The sooner you can get out and speak and be noticed, publish and be recognized, get a viewpoint out there, sort of build a name for yourself, and make connections, the better.”

Michael Moore, at 15:48

It’s what he tells early career attorneys, and the reason is the business itself. “Because in this tech business, I do think your connections are everything, and the business is very much about who you know and how you are perceived by them.”

He took his own advice. The posts on this page are a sample of one year of it: a Law.com series, an IAPP article, a talk at his law school, a session with a law firm, a video series and a conference he co-chaired. His LinkedIn entry for Glean counts more than 30 speaking events and ten articles.

The safest ship is in the harbor

Sasha added the line we hear from founders all the time: “I often hear people saying that, hey, I’m afraid of making this mistake. And my response is usually, if you don’t want to make mistakes, you don’t do anything.” Michael: “Yep, don’t get out of bed in the morning.” And me:

“So the safest ship is still in the harbor, not going anywhere.”

Jon McLachlan, at 16:28

Michael took it seriously. “You have to take risks.” The skill is calibration: “you have to develop the appropriate gauge or calibration of what risk is appropriate for the stage of company that you’re at, and what are the things you don’t give up on.”

“So I would say in the business [we’re in], security and privacy is absolutely key.”

Michael Moore, at 17:08

Everything else can bend, “like pricing or terms or things like that.” What can’t: “you must, as a fundamental element, preserve your customer trust, protect your customer data, make sure they can trust you as a provider of services to their company.” His test is simple. “If you can do all of those, you’re probably on the right path.”

It’s the right frame for a startup, and the opposite of the safe harbor. Ship, take the commercial risks your stage calls for, and decide ahead of time which few things you’ll never trade. When those few things are security and privacy, a buyer’s legal team can tell.

Start their day in your tool

Sasha closed the loop on trust: “Trust is everything. People’s relationships are built on trust.” And on experience: “And user experience is something that often is not considered as the top priority, but often should be.” Michael agreed. “User experience can make or break a product, make or break a company.”

“You want them to start their day every day in your tool.”

Michael Moore, at 18:28

It’s the fifth item on his list for chief executives again, and it’s a security point as much as a product one. A safe tool nobody wants to use sends people back to the free one. Ted Harrington made the same case on episode 80: the best security is the kind the user barely notices.

What happened after the tape

The episode aired on July 15, 2025. On June 10, between the recording and the release, Glean raised a $150 million Series F led by Wellington Management at a $7.2 billion valuation, and said it had passed $100 million in annual recurring revenue in its last fiscal year (Glean). Its story page now lists more than $350 million (Glean).

Michael’s title is now general counsel and company secretary, and his year in public reads like the episode, written down. In February 2026 came the AI inventory, in March the Santa Clara talk, and in April the two-part Law.com series and an IAPP article with Michael Bishop of Resmed on moving from shadow IT guesswork to data protection impact assessments a regulator can follow. May brought the Paul Hastings session. In September he sat down with Curtis Salinas, head of business technology at Harvey, for Glean’s Unprompted series, and co-chaired The L Suite’s LegalOps x GC Conference in San Francisco with Cindy Kaneshiro of Nutanix.

The law moved, too. On February 10, 2026, Judge Jed Rakoff of the Southern District of New York ruled in United States v. Heppner that documents a defendant generated with a consumer AI tool weren’t protected by attorney-client privilege or the work-product doctrine, pointing to terms that let the provider train on prompts and disclose user data. The court left open whether an enterprise product that keeps user data out of training and carries contractual confidentiality could come out differently (Proskauer). It’s one court and one set of facts, and it reads like Michael’s line from the tape in legal form: enterprise data belongs in enterprise tools. Our list of what law firms ask legal AI vendors starts from that ruling.

Samsung’s story found its second half in June 2026, with the device division bringing approved outside AI tools to work, as told above. And when Glean held its second user conference, security had its own stage. One session walks through Glean’s platform for enterprise AI, with context, an AI gateway, usage controls and data protection working together, and a short one covers the usage controls that set guardrails and manage access and spend.

Michael’s minimum bar hasn’t changed in the year and a half since we recorded, and the buyers we work with ask about every line of it. If you sell AI to enterprises, write down what happens to their data, make your settings match the sentence, show where every answer came from, and put your lawyer and your security lead on each other’s speed dial. Nothing is really free, and the companies that act like it are the ones buyers trust.

Did you catch it? A six-question quiz

Does your AI product clear the minimum bar?

Michael answers every AI procurement question as the buyer would. This scorecard asks eight of those questions about your own product. It takes about two minutes.

How many hours do security questionnaires take?

Michael’s Law.com piece describes teams that turn a 100-question questionnaire into roughly a 20-minute workflow. This calculator compares that pace with answering by hand, using your own numbers.

From Analog Devices to Glean

Everything on this page sits on one timeline, from Michael’s first engineering job to his year of writing about legal AI. Filter it by thread.

What to do this week

Five things, in the order Michael would probably check them as a buyer.

1. Read your AI terms the way a buyer will. Find the sentence about training on customer data, in your own terms and in each model provider’s terms for the plan you use. If it isn’t there, write it, and make the settings match.
2. Test permissions as a user who shouldn’t see something. Ask your product about a document that user can’t open in the source system, and confirm it never shows up in an answer.
3. Put a source behind every answer. If your product writes content, cite the documents it came from, so a person can check it in seconds.
4. Find the consumer AI tools your own team uses for company data. Move that work to enterprise accounts, and make the approved path the easy one.
5. Put legal and security on one playbook. One page with who calls whom, the reporting clocks that apply to you and the outside counsel you’d call, then one practice run before you need it.