Lucas Hathaway

The benchmark is an honest assessment against a named framework, either the NIST AI Risk Management Framework or CRI's Financial Services AI Risk Management Framework, on the reasoning that institutions need to know where they stand and regulators already point at these frameworks.

• The four foundations are AI policies written for how a team actually uses the tools rather than copied from a template, a repeatable AI risk assessment that runs inside the existing risk process, vendor security reviews extended with AI items alongside SOC 2, and incident-response playbooks updated for AI failures.
• Lucas Hathaway argues the work is additive and the evidence should be kept as it happens, not rebuilt the week before an exam.

Lucas Hathaway names the AI failure modes an incident playbook should cover, including a model acting on its own or giving a confidently wrong answer that reaches a member.

• They describe a free, one-session NIST AI RMF assessment that gives the institution a written benchmark and roadmap.
• Dozens of these sessions have been run, and nobody has passed yet—a reason to benchmark early.

Read original

A vendor is reviewed and its SOC 2 filed, then a point release switches on a summarizer or scoring model without notice.

• Lucas Hathaway argues that SOC 2 was never going to catch this: training data, bias testing, and human oversight are not trust services criteria.
• They recommend starting with two scoping questions for each vendor: Does the product touch customer data, and does it make or influence a decision about a person?
• The answers determine the depth of the review.
• Next, send critical and high vendors a one-paragraph discovery email asking what AI is in the product today, and use the replies to build the review queue.
• Add a renewal clause requiring notice before a new AI feature is enabled on the institution's tenant.

The checklist section for examiners is a dated record of what the institution asked, what the vendor said, and what it decided.

• Lucas Hathaway ties it to examiner checks under Govern 6.1 and 6.2.
• They also pass along two practices from teams that have done this well: send the discovery email first, and get the tenant-notice clause into the next renewal so a surprise becomes a scheduled review.

Read original

AI features appear through releases in vendor tools the institution already bought.

• They recommend running three discovery sources together over about two weeks, because any one source alone misses most of it: technical discovery of AI domain traffic, browser extensions on managed endpoints, and OAuth grants in the Microsoft 365 or Google tenant; vendor management, with one question per critical vendor about whether the product has an AI component; and short, explicitly non-punitive department surveys sent through department heads.
• The survey is the only source that catches personal-account use.
• Add an AI component field to the vendor inventory so the exercise does not have to be repeated.

A credit union security officer felt good about having one sanctioned tool, Copilot, until meeting recorders, summarizers, and assistants appeared through vendor tools the institution already owned. Lucas Hathaway notes that meeting recorders reliably appear in OAuth grants.

Read original

They recommend setting the models to block rather than warn: people click through warnings but cannot click through a block, stopping a spreadsheet of member data from ever reaching a chatbot. Pair the control with a documented human review procedure specifying what a person must review versus what AI decides alone, particularly near lending, fraud, and member communications, and with training on sanctioned tools.

At one institution Lucas Hathaway worked with, DLP was enabled and set to block Copilot.

• They describe that single setting as doing the most work of anything in the AI program.
• Most shadow AI, they argue, is a usability problem before it is a compliance problem, which is why training on sanctioned tools is part of the same recommendation.

Read original

They argue that approving a category such as "AI assistants" is not enforceable, while a named list of three product names is.

• The same approach runs through the vendor posts: the examiner-facing artifact is a dated record of the questions asked, the vendor's answers, and the institution's decisions, kept as the work happens.
• Their rule of thumb is to have the policy, risk assessment, and vendor review on paper before anyone asks for them.

On the allowlist point, Lucas Hathaway reports watching a category-wide approval get walked back in front of an examiner, concluding that an overly permissive approach fails later. The alternative is a short list of specific product names plus a paragraph of procedure, which they describe as roughly a very good week of work rather than new software.

Lucas Hathaway's posts present category-wide approval of AI assistants as a control that can fail under examiner scrutiny, and naming specific tools as the fix. The recommendation reflects a correction drawn from observed cases, not a reversal of the goal of approving tools deliberately and documenting decisions.

Read original · Original 2 · Original 3

It answers the first practical question before getting into vendor-level or control-level details.

Hathaway says institutions most often ask not what a framework says, but where to start.

• The recommended first step is an honest benchmark against the NIST AI Risk Management Framework or CRI's Financial Services AI Risk Management Framework.
• Next come four foundation pieces: AI policies written for actual team use, a repeatable AI risk assessment within the existing risk process, vendor security reviews with AI items added, and incident response playbooks updated for AI failures.
• The post urges institutions to add AI to the existing program and keep evidence as work happens rather than rebuild before an exam.
• It also offers a free NIST AI RMF assessment and notes that dozens have been done and nobody has passed yet.

For brands planning a campaign

For a fintech compliance software company targeting credit unions, this post supports a co-branded webinar built around a benchmark-first message. Open with the gap between wanting to start and passing an assessment, then show how software supports the four foundation pieces.

Newcomers can see the specific gap and the first actions to take.

Hathaway opens with a risk leader's panel comment: vendors with an AI element are not being adequately assessed.

• The post explains why: a vendor was reviewed two years ago and its SOC 2 filed, then a point release turned on a summarizer or scoring model without notice.
• SOC 2 was never going to catch it because training data, bias testing, and human oversight are not trust services criteria.
• The recommended process starts with two questions for each vendor: Does it touch customer data?
• Does it make or influence a decision about a person?
• Practical steps include sending critical and high vendors a one-paragraph discovery email asking what AI is in the product today, adding a renewal clause requiring notice before a new AI feature is enabled, and keeping a dated record of what was asked, what the vendor said, and what was decided.
• The post ties that record to examiner expectations under Govern 6.1 and 6.2.

For brands planning a campaign

This is the strongest anchor for a co-branded webinar on AI vendor risk assessment. The SOC 2 gap and two qualifying questions can shape the webinar, with the checklist as the takeaway asset.

It offers a concrete discovery method, not just a general warning.

Hathaway describes a credit union security officer who felt good about having one sanctioned AI tool, Copilot.

• Then meeting recorders, summarizers, and assistants appeared in vendor tools the institution already owned, with no procurement and no new domain in the logs.
• The post argues that shadow AI is harder to find than shadow IT because the AI was not purchased separately.
• It recommends running three discovery sources together over about two weeks: technical discovery through AI domain traffic, browser extensions on managed endpoints, and OAuth grants in Microsoft 365 or Google; vendor management, with one question per critical vendor about an AI component; and short, non-punitive department surveys to catch personal account use.
• The post recommends adding an AI component field to the vendor inventory.

For brands planning a campaign

Use this post to frame AI governance as an inventory problem before it becomes a policy problem. Its three discovery sources can anchor a webinar segment or practical worksheet for credit unions assessing current exposure.

It explains what to turn on first and why a block is stronger than a warning.

Hathaway points to prebuilt GLBA data loss prevention models in Microsoft 365 that look for Social Security numbers, card numbers, and driver's license numbers.

• Setting the models to block rather than warn stops a spreadsheet of member data from reaching a chatbot.
• One institution had this enabled and blocking for Copilot.
• Hathaway says it did more work than anything else in their AI program because people click through warnings but cannot click through a block.
• The post pairs that control with a named tool allowlist of three product names rather than a category-wide approval, a written human review procedure for lending, fraud, and member communications, and training on sanctioned tools.

For brands planning a campaign

Use this post for practical content or a webinar module on quick wins in AI governance. For a compliance software vendor, it’s a natural way to discuss configuration reviews, policy templates, and human review workflows alongside existing Microsoft 365 controls.

From there, they name four foundations: AI policies and procedures written for how a team actually uses the tools; a repeatable AI risk assessment that runs inside the existing risk process; vendor security reviews extended with AI items alongside SOC 2; and incident-response playbooks updated for AI failures such as a model acting on its own or giving a confidently wrong answer that reaches a member.

• The approach is additive: add AI to the program already run and keep evidence as work happens rather than rebuilding before an exam.
• They offer a free NIST AI RMF assessment as a working session that produces a written benchmark and roadmap.

Benchmark against a named framework, then build four repeatable foundations inside the existing risk program.

For brands planning a campaign

A co-branded webinar on AI vendor risk assessment can open with this benchmark-first message, then position vendor review as one of the four foundations rather than a standalone product. The named frameworks give credit union compliance and risk staff a familiar starting point.

Examples: Source 1

The problem is not discipline: a SOC 2 may have been filed two years ago, but a point release can turn on a summarizer or scoring model without notice, and SOC 2 criteria do not cover training data, bias testing, or human oversight.

• The method starts with two questions per vendor: does it touch customer data, and does it make or influence a decision about a person?
• Those answers set review depth.
• Lucas Hathaway recommends sending a discovery email to critical and high vendors asking what AI is in the product today, using that list as the review queue, and adding a renewal clause requiring notice before a new AI feature is enabled on the institution's tenant.
• A dated record of what was asked, what the vendor said, and what was decided supports examiner checks under Govern 6.1 and 6.2.

Two-question triage, vendor discovery email, renewal notice clause, dated decision record.

Lucas Hathaway frames the gap as a change-detection problem in existing vendor relationships, not as a failure to perform initial reviews.

For brands planning a campaign

This is the strongest source for the proposed webinar. A co-branded session can walk through the two scoping questions, the discovery email, and the tenant-notice clause, showing how a fintech compliance product operationalizes each step.

Examples: Source 1

A credit union had one sanctioned tool, Copilot, then meeting recorders, summarizers, and assistants appeared through vendor releases.

• Lucas Hathaway recommends running three discovery sources together over about two weeks: technical discovery of AI domain traffic, browser extensions on managed endpoints, and OAuth grants in Microsoft 365 or Google tenants; vendor management asking one question per critical vendor about AI components; and non-punitive department surveys of five questions or fewer sent through department heads.
• The survey is the only source that catches personal-account use.
• Lucas Hathaway also recommends adding an AI-component field to the vendor inventory so the exercise does not have to be repeated.

Run technical, vendor-management, and department-survey discovery in parallel; add an AI-component field to vendor inventory.

For brands planning a campaign

A campaign targeting credit unions can use this as a practical lead magnet or webinar segment, since the three-source method is concrete and vendor-neutral. A compliance software company can map each source to a workflow without suggesting the tools automate discovery.

Examples: Source 1

For institutions in Microsoft 365, prebuilt GLBA data-loss-prevention models already detect Social Security numbers, card numbers, and driver's license numbers.

• Setting DLP to block rather than warn stops a spreadsheet of member data from reaching a chatbot; people click through warnings but cannot click through a block.
• They pair that control with a named tool allowlist, because 'approved AI assistants' is not enforceable while a list of three product names is; a documented human-review procedure, especially near lending, fraud, and member communications; and training on sanctioned tools.
• They describe most shadow AI as a usability problem before it becomes a compliance problem.

Configure existing DLP to block, name specific approved tools, document human review, and train on sanctioned tools.

Lucas Hathaway presents a category-wide approval of AI assistants as a control that can fail under examiner scrutiny; the specific list is the recommended alternative.

For brands planning a campaign

A co-branded webinar can use this as a concrete, low-cost starting point for credit unions and then show where a vendor risk product adds visibility and evidence. The named-allowlist point supports a practical takeaway that does not depend on buying new software.

Examples: Source 1

They convert repeated questions into reusable artifacts: a two-question vendor triage, a discovery email, a renewal notice clause, a three-source shadow-AI discovery method, DLP settings, and a named tool allowlist.

• The posts consistently point readers toward what they can put on paper before someone asks for it, and one post offers a free NIST AI RMF assessment as a working session that produces a written benchmark and roadmap.
• The emphasis is on doing the work inside the institution's existing program rather than creating a separate AI program.

Package repeatable questions into checklists, offer a working session with a written output, and keep the artifacts tied to existing risk and vendor processes.

For brands planning a campaign

This pattern fits a co-branded webinar that leaves attendees with checklists and a benchmark, not just a product demo. A fintech compliance software company can sponsor the resource while Lucas Hathaway supplies the practical method and credit union context.

Examples: Source 1 · Source 2 · Source 3 · Source 4

They offer no-budget first steps: pull the application list from the identity provider, sort OAuth grants by scope and filter card spend for SaaS. Both write for security and IT leaders seeking practical governance steps rather than framework theory.

Nir Sofer's posts span ITSM, software management and leadership structure, including an argument for a CEO-CFO-CISO power triad and survey findings on software visibility. Lucas Hathaway focuses on AI governance for financial institutions: vendor AI assessment, credit union exam evidence and named AI risk frameworks.

Read original · Original 2

Jeff Willinger's Center of Excellence case study and Lucas Hathaway's foundation-building posts favor phased, documented programs.

Jeff Willinger's case study covers an enterprise AI Center of Excellence across a health and life sciences organization, with adoption, training and business-value metrics. Lucas Hathaway focuses on financial institutions, where the priorities are examiner-facing evidence, vendor AI risk and member-data controls.

Read original · Original 2

Robert Cain promotes a short, plain-language survey across five cybersecurity domains; Lucas Hathaway promotes a framework benchmark and repeatable AI governance checklists.

Robert Cain's survey addresses cybersecurity risk broadly for small and midsize businesses and is not AI-specific. Lucas Hathaway's benchmark and controls focus on AI governance for financial institutions, covering NIST AI RMF, vendor AI review and AI incident response.

Read original · Original 2

The posts give attendees a concrete visibility exercise: pull the application list from the identity provider, pull the OAuth grants sorted by scope, and pull twelve months of card spend filtered for SaaS, then treat the gap between the three lists as the shadow estate.

• One post supplies a survey figure to lead with — 718 IT professionals, with only 9.9% reporting full visibility into their software ecosystem — and reframes the problem as an AI assistant holding a delegated OAuth grant with read access to mail and files, sitting inside the identity provider where a firewall never sees it.
• For a compliance software company whose product covers software inventory, SaaS discovery, or identity-adjacent risk, this puts the product at the discovery step before any vendor review.

Lucas Hathaway's posts stop short of the financial-institution specifics a credit union compliance audience needs.

• They do not cover vendor questionnaires, the limits of a filed SOC 2, examiner expectations, or the Govern 6.1 and 6.2 framing that anchors the proposed session.
• The result: a strong discovery exercise, but less on the review process.
• The headline statistic comes from a software vendor's service-management study, which suits a visibility message more than a regulator-facing one.
• Ask for audience composition and willingness to partner before moving forward.

Examples: Source 1 · Source 2

The case study walks through a 90-day AI Center of Excellence rollout at a $1.2B health and life sciences organization with 400+ Microsoft 365 Copilot licenses: executive charter and steering committee, SharePoint and OneDrive permission cleanup, Purview sensitivity labels, role-based Copilot workshops and a prompt library, a monthly cadence between the CIO, chief revenue officer, HR, compliance, and data leaders, and usage analytics reporting proposal creation time down 28%, executive prep time down 35%, and 14 internal workflows automated.

• A compliance software company whose product supports AI policy, enablement, or adoption measurement on a Microsoft 365 estate would get a ready-made session spine with named outcomes.
• The posts also show sustained teaching across the Microsoft ecosystem — Modern Work, Copilot, Power Platform, and security — and a career spanning Microsoft MVP, consulting, and partner work.

The tradeoff is product surface and vocabulary.

• The Center of Excellence story is about internal deployment, adoption, and return on Microsoft licensing; it does not address vendor due diligence, AI notice clauses in renewals, or dated examiner records — the territory of a vendor-risk compliance product.
• A short career-change post also makes this the only comparator whose account reads partly as a personal milestone, affecting how a co-branded asset would be positioned.
• The account is openly Microsoft-partner oriented, so the webinar would lean on the Microsoft stack rather than framework-neutral governance and may read as overlapping with, rather than supporting, the compliance company's process story.
• Ask for audience composition and partnership terms before moving forward.

Examples: Source 1 · Source 2

The posts target organizations under 500 people where someone in IT, the COO, or the founder carries security on top of their actual job.

• They describe a 15-question, 1-to-5 snapshot across five domains — breach preparedness, breach deflection, incident response, breach recovery, and breach optimization — with a gap analysis tied back to the core business.
• The framing is deliberately modest: a summary read, not an audit, that gives an honest starting point without a consultant's calendar.
• For a credit union audience where the vendor-risk contact may not hold a security title, that is an approachable registration asset.
• The argument that the weakest domain, not the strongest, decides the next bad day gives a session a clear teaching point.

The material is about general security posture, not AI governance.

• It does not discuss AI features inside vendor products, AI-specific questionnaire items, or the review-depth triage the proposed webinar is built on.
• A compliance software company selling AI vendor risk would therefore be co-branding a five-domain security survey instead of an AI asset.
• The survey is presented as the work of a named external risk consultant, so the resource would credit that consultant rather than the company's own method.
• The strongest content lives in the LinkedIn post, with the link and profile directed to a comment thread.
• Ask about audience composition and commercial availability before moving forward.

Examples: Source 1 · Source 2