Natalia Quintero, who leads Every’s consulting practice, has spoken with more than 100 companies about AI adoption, including the New York Times and the hedge fund Walleye Capital. In June, she and Mike Taylor—now Every’s head of evals practice—hosted a live webinar for 400 executives on how organizations can adopt and implement AI without stalling out. More questions came in than they could answer live, so afterward they wrote up 33 of them—covering strategy, winning over skeptics, tool selection, governance, and how to restructure a company for an AI-native future. That document went only to attendees at the time; we’re making it public for the first time here.—Kate Lee
When you put a room full of executives in front of two AI consultants and let them ask anything, you learn about where organizations are stuck with this technology. Here are all the questions that came up on our live session on June 2—from how to get skeptics on board to keeping context fresh.
Every Consulting works with tech and finance teams to implement AI across their workflows, from strategy to training to building the actual tools.
Strategy and organizational transformation
Q: How do you get teams to think bigger about change? Not just getting the work done faster, but rethinking how the work is done?
A: Give your most AI-curious staff permission to build ambitious projects. It’s easier to raise the ceiling than the floor. It’s easier for one AI-fluent person to create 10 good skills and do 10 times the work than it is to get 10 people to each create one skill or double their output. Once that’s in place, set a vision, designate AI champions, prove out high-value skills, then scale.
Q: What does a good vision for AI adoption look like?
A: AI is a tool, not a strategy. A good vision applies AI to your company’s strategic roadmap and allows you to tackle parts of it that were impossible before.
For example: customer service. There’s effectively infinite demand for it, and it’s too expensive to staff the phones enough to get wait times down to a few minutes. Even then, the person that your customer talks to might not have the context the best customer representative has. AI is a great leveler here. It makes it cheaper to serve every customer quickly, it gives every representative full context by allowing them to search for it live, and it can pre-emptively answer questions before anyone gets on the phone. That’s what a good vision looks like: ambitious and finally affordable because of AI.
Q: Do you need to convince skeptics? How do you get them over the hump of seeing value in adopting AI?
A: Rather than focusing efforts on convincing skeptics, highlight AI champions. When AI champions, often skeptics’ own peers, demonstrate how they are getting value from AI tools, others start doing it out of practicality. Pair that with demos of new AI tools or skills that are relevant to your teams.
I’d also encourage side projects. People with side projects are much further ahead on AI because they can take bigger risks than they can at work, by using tools that aren’t on the approved vendor list, and by pushing the limits of what works. Previously, a manager might feel uncomfortable about that. Now, I’d be disappointed if an employee didn’t have a side project, because it means they don’t have the space to experiment, play, and take bigger risks.
Q: Is there a tension between using AI for productivity and using it to pursue bigger ambitions—and how do you think about the balance?
There are two kinds of productivity: doing what you already do faster, and doing things you otherwise wouldn’t be doing. That second one is vision. You can’t only focus on productivity, but once you solve the productivity problem and automate what you already do, you’ll start looking for new things to do. Prioritize existing pain points for immediate value and place strategic bets on bigger products.
Show it once, and it runs the process from then on
Walk a Bot through a workflow the way you’d walk a new hire through it: the steps, the exceptions, the part where you check the invoice against the PO. It remembers, then runs the whole thing on its own, in its own browser, on its own computer in the cloud. Onboarding, reconciliation, outbound, release notes: pick the process nobody on your team wants to own and hand it over. Grok Bot, from the team behind SpaceXAI.
AI fluency, training, and ways of working
Q: What is AI fluency? What are the levels of AI?
A: We think of them as these eight levels. Most people are in the first couple of levels—they use ChatGPT, or they co-work with AI using Gemini in Docs or Claude alongside Excel. Level three, agentic, is where knowledge workers are moving toward. Many are leveraging Codex, Claude Cowork, and Claude Code and using these tools more agentically. In order to do that, they need to adopt skills, and they need to learn how to work with agents and run multiple tasks at a time.
Above that, it’s largely experimental. Engineers are way past the early stages of AI adoption—sometimes running multiple calls at once, and getting into orchestration (around Level 8), where a manager AI manages other AIs.
Q: What is the most efficient way for teams or a company to use AI together?
A: By creating shared skills, shared examples, and shared standards. The company should not end up with 1,000 people maintaining 1,000 private prompt folders. Start by turning the best recurring workflows into reusable skills, then make those visible, editable, owned by someone, and accessible to everyone.
Q: What works best when teaching non-technical users how to use AI?
A: Do not teach AI in the abstract. Give people tools that they can use their real work and have them build something useful in the session. The aha moment usually comes when they see AI produce an artifact they recognize, like a briefing, memo, spreadsheet, review checklist, customer response, or project plan.
Q: What are good early wins or simple skills?
A: The best early wins are boring and frequent: doing meeting preparation, call summaries, document review, first drafts of marketing content, research briefs, triaging your inbox, quality control checklists, and turning messy inputs into structured tables. If the task happens every week, has a known output, and a human can review it quickly, it is a good candidate for automation with AI.
Q: How do you encourage reuse instead of everyone inventing their own prompts?
A: Make reuse a visible management habit. Have demo days, office hours, a shared skill repository, and clear owners for the best workflows.
Q: Is there value in building a centralized hub for AI knowledge inside a company, or does information move too fast for that to work?
A: I don’t think one centralized AI knowledge hub really works—every document starts going out of date the moment it’s published, and there’s no good way to identify what needs updating. People are also at very different stages: Someone might know everything about skills but be behind on MCPs. Articles are valuable precisely because they’re timely––they were news when they were published. To turn them into a hub, you’d need another layer on top to review the content and keep it up to date, which is very resource-intensive and probably wouldn’t satisfy people anyway, since everyone has different opinions on what matters right now. A hard problem to solve, and not a huge payoff.
Q: What are you seeing in human resources, marketing, legal, and other non-IT teams?
A: The use cases are often very strong because these teams do a lot of language-heavy, context-heavy work. HR can use AI for onboarding, policy drafts, employee communications, and benefits questions. Marketing can use it for briefs, segmentation, creative variations, and research. Legal and compliance can use it for first-pass review and issue spotting, but with stricter human review.
Tools, platforms, and technical infrastructure
Q: What about organizations that can’t build frontier-level AI? How do you handle implementation if you need to buy?
A: In organizations where you can’t build your own frontier tools, you can still create skills and connect into the software you already have. Even if you can’t build your own model, you’re still better off customizing how you use AI rather than waiting for a third-party vendor to add AI features for you because these vendors are often months behind.
Q: How do you manage the number of platforms available? Do you recommend HR leaders go to Claude Code? How do you handle constant switching?
A: We tend to recommend that you select one platform. Switching is expensive—you have to renegotiate contracts and train staff, and there are practical barriers, such as the fact that skills written for Codex might not work as well in Claude Code, and vice versa.
As long as the model is from Anthropic or OpenAI, you’re fine. Or go with a third party like Cursor or Copilot that has access to both, which can lower switching costs. This issue is not fully resolved, because engineers in particular want the best model available and will adapt their processes to get it.
Q: How can AI review designs in Figma?
A: There’s a Figma MCP. Codex also has a built-in browser and annotations that are really good—Codex can control the Figma app directly. That’s typically how we’d do it.
Q: Is it worth constantly switching between AI tools to stay on the cutting edge?
I don’t recommend constant tool-switching for novices. Being a month or two behind the frontier is absolutely fine. Unless you’re an AI engineer or very deep into this, you won’t get much edge from switching—and anything genuinely good gets copied by the other labs within eight to 12 weeks.
Q: What do you know about implementing AI today that wasn’t obvious a year ago?
NQ: To use AI effectively, you still need to provide clear instructions and guidance. Clear thinkers, strong writers, and subject matter experts with an eye for excellence are best positioned to get value from it.
MT: How fast things would move, and it keeps accelerating. Also, that AI would still not be great at writing: I thought writing was nearly solved and code would take longer, and it’s been the other way around. I also assumed everyone would be managing lots of AI agents. Instead, the predominant use case is one AI agent everyone uses that’s good at routing different tasks.
Governance, security, and risk
Q: How does AI governance differ in a heavily regulated environment where regulators want transparency and explainability?
A: AI actually gives you more transparency. If you’re implementing through Claude, activity logs are saved through the enterprise plan. You can use AI to analyze those logs and generate reports. So you can provide more transparency this way, not less.
Q: How can business teams pursue AI automation without making IT feel bypassed?
A: Make IT a design partner, not the department of “‘no.”’ The business team should own the workflow and the definition of success; IT should own access controls, system architecture, and security review. Different tools require different levels of scrutiny. An AI assistant helping someone write is a very different proposition from a system that can access company databases. Treating every use case as equally risky can make it harder to move forward. Using AI is about understanding the potential upside and downside risks, and IT is already a resource for that. It’s an evolving conversation to explain business outcomes and collectively decide which protocols make sense for the business.
Q: Do AI pitfalls vary by industry?
A: There are universal pitfalls. AI adoption doesn’t succeed when it’s not clear who owns the process, executives have weak fluency in these tools, the tools are scattered, and a company does not rethink its workflows. What changes by industry are the constraints within which the AI adoption must take place. Finance and healthcare have more compliance, media has more judgment and brand risk, engineering has faster tooling cycles. But the adoption problem is basically the same: People have access to tools before they know how to incorporate them into their work.
Q: What’s the discourse around token economics? How do you weigh value versus cost?
A: For a brand-new task, something you’re innovating on, or something you’ll only do once, use the best model. The best model is often cheaper even when it’s the most expensive per token, because it saves time correcting mistakes, lowers the risk of bad outcomes, and usually gets the answer right faster without self-correcting. So almost all of your work should run on frontier models. Then, when it becomes a recurring task you do again and again, offload it to smaller models using a framework like DSPy, prompt optimization, or hill climbing to keep it working well at the smaller size.
Q: How should companies think about data sovereignty, cost control, and dependency on one model or vendor?
A: Do not over-engineer this on day one. Pick a secure enterprise platform, get the workflows right, and keep your important instructions, skills, evals, and data layer portable where you can. If compliance or data residency rules make the off-the-shelf tools impossible, then build a harness. But custom infrastructure is expensive, so the reason to build should be a real constraint, not driven by aesthetic preference.
Measuring impact
Q: How do you measure collaboration, output quality, and value from AI adoption?
A: There is no single metric. Measure the benefits of AI adoption at the workflow level: cycle time, the quality of the output, how many errors there are, how often things need to be reworked, user satisfaction like NPS, and business impact in terms of traditional cost-benefit analysis. Also track reuse: how many people use a shared skill, how often it is improved, and whether it becomes part of the team’s normal process.
Q: How should teams evaluate AI workflow performance over time?
A: Start with a baseline before AI: how long a task took, to what standard it was being performed, at what cost, and with what common mistakes or errors. Then track whether the workflow actually improves after AI is introduced. For important workflows, keep a small eval set of real examples and rerun it whenever the prompt, model, or tool changes. Otherwise you’re just trusting vibes.
Knowledge and context
Q: How do you keep your AI’s understanding of the company accurate and up to date—and who owns that?
Everyone’s experimenting with memory and file-storage solutions. In practice, people are solving it by putting the right context into GitHub repositories, but I don’t think that’s a great solution. Microsoft has Windows IQ, and Google has its own MCPs for Docs, but this is still not well solved.
Q: What are best practices for knowledge-management debt and stale context?
A: AI exposes knowledge debt, and it does not magically fix it. We would avoid starting with a giant knowledge-base project. Start with the 10 to 20 sources that matter most for a workflow, assign owners, add freshness dates, and prune aggressively. Context is not a warehouse where you dump everything you might need––every document needs to earn its place or it could be hurting more than helping.
Who owns AI?
Q: What balance works best between a dedicated AI team and broad organization-wide enablement?
A: You need both. A central AI team should set standards, handle security and vendor decisions, maintain shared infrastructure, and help build the first few examples. But the distinct business teams need to own the workflows. If the central team owns every use case, it becomes a bottleneck, and if everyone builds alone, you get sprawl.
Q: How do you handle AI crossing the borders between functions and departments?
A: This is one of the real management questions. AI attacks the boundaries in org charts because the work itself was never as cleanly separated as the chart made it look. I would not try to solve that abstractly. For each workflow, define the business owner, technical owner, risk owner, and human approver. That is usually clearer than trying to redraw the org chart first.
Q: If you were institutionalizing AI at a 200-to-500-person organization from scratch, where would you start?
A: Start with the executives. Get leadership hands-on enough that they understand what is now possible, then pick two to three teams with painful, repetitive, high-value work. Map the workflow, build one useful skill or agent, prove it works, and use that as the internal example. The mistake is trying to “roll out AI” broadly before anyone can point to a concrete workflow that changed.
Q: Do executives usually know what the highest-leverage automations would be, or do they have blind spots?
A: They usually know the business pain, but not what’s easy or hard to do with AI tools. The best process is to start with the company’s goals, then ask where work is slow, expensive, inconsistent, or bottlenecked by scarce expertise. AI opportunities tend to show up where there is a clear output, a lot of context gathering, and human judgment at the end.
Every Consulting—how we work
Q: What is the shape of an Every consulting engagement?
A: It varies, but the pattern is usually: executive alignment, discovery calls, workflow mapping, AI champion selection, hands-on training, skill or workflow building, office hours, and handoff. Sometimes that is a half-day executive session; sometimes it is a multi-month implementation. The common thread is that we help people build tools that solve a pain point, instead of making them sit through them to listen to presentations.
Q: How does Every’s consulting inform product work and offerings?
A: Clients show us where the tools break, where people get confused, what workflows actually matter, and what kinds of skills get reused. That feeds back into our writing, our products, and our own internal systems.
Q: How are you using AI in your own consulting and client-development work?
A: We use AI across the whole process: preparing for calls, finding patterns across discovery interviews, drafting proposals, creating training materials, building client-specific skills, and testing workflows. We also have an agent, Claudie, who supports our consulting practice with day-to-day ops.
Q: How do you avoid dependence when outside AI experts or forward deployed engineers drive early wins?
A: Pair every external builder with an internal champion. The deliverable should not just be a working tool; it should include the skill files, the evals, the decision log, the operating instructions, and a colleague who can modify it. Otherwise you get a great demo and no organizational capability.
Natalia Quintero is the head of consulting at Every. You can follow her on X at @NataliaZarina and on LinkedIn. Mike Taylor is the head evals at Every and a co-author of Prompt Engineering for Generative AI (O’Reilly). To read more essays like this, subscribe to Every, and follow us on X at @every and on LinkedIn.
Everyone’s a builder now. Every All Access gets you the full membership plus the Builder Pack—$9,000+ in credits for the tools we build with.
