Build a Team That Ships with AI

Leading AI-Using Teams

When Your Team Moves Faster Than Your Policies

In early 2023, a mid-sized marketing agency in Chicago called Brafton quietly watched something unexpected happen. Individual writers on their content team had started using ChatGPT to draft first versions of blog posts and client briefs. Nobody had asked them to. Nobody had told them not to. Within six weeks, output per writer had jumped by roughly 40%. Then a client noticed. One article contained a confident statistic about retail foot traffic that was completely fabricated, hallucinated by the AI and never fact-checked by the writer. The client pulled the contract. Brafton's leadership had a decision to make: punish the behavior, ban the tools, or build a framework. They chose the framework.

What Brafton experienced is now playing out in thousands of organizations. Employees are adopting AI tools faster than managers can set expectations. A writer discovers Claude Pro cuts research time in half. A sales rep starts using ChatGPT Plus to draft follow-up emails. An HR coordinator uses Copilot to summarize 200-page policy documents in minutes. These are real productivity wins, but they're happening without standards, without accountability, and often without the manager even knowing. The result is a patchwork of individual habits that create inconsistent quality, hidden risks, and missed opportunities to scale what's actually working.

Brafton's eventual response was to create a simple internal protocol: AI-drafted content required a visible review tag in their project management system, fact-checkable claims needed source links, and writers were required to disclose which sections were AI-assisted before submission. Output quality stabilized. The lost client eventually returned. More importantly, managers had visibility into how AI was being used, and could coach around it. That shift, from passive observer to active leader of AI-using work, is exactly what this lesson is about.

The Principle That Brafton Learned the Hard Way

The lesson from Brafton isn't really about AI. It's about the oldest problem in management: when a new tool or method enters your team, the manager who ignores it doesn't stop it from spreading, they just lose the ability to shape how it spreads. Think about how email changed office communication in the 1990s, or how smartphones changed field sales teams in the 2010s. In both cases, the managers who thrived weren't the ones who mastered the technology themselves. They were the ones who established norms quickly: what's appropriate, what's not, what good looks like, and who's accountable for what.

With AI tools, the stakes are slightly higher because the output looks finished. A ChatGPT-drafted proposal looks polished. A Copilot-summarized report reads smoothly. Unlike a rough spreadsheet or a quick Slack message, AI output doesn't visually signal that it needs review. That's the trap. Employees mistake fluency for accuracy, and managers mistake volume for quality. The principle that separates good AI-using teams from risky ones is deceptively simple: every piece of AI-assisted work needs a named human accountable for its accuracy and appropriateness before it goes anywhere.

This isn't about distrust of employees or distrust of AI tools. It's about recognizing that AI tools, including the best ones like Claude Pro and ChatGPT Plus, are confidently wrong in ways that are hard to spot without domain knowledge. A marketing AI doesn't know your client's brand voice. A legal summary tool doesn't know which jurisdiction's law applies to your specific contract. A recruiting AI doesn't know that your company has a standing diversity commitment that changes how you screen résumés. The human accountable for the output is the person who holds that context, and that person needs to be identified before the work leaves the team.

A Hospital System That Got It Right

In a very different industry, Vanderbilt University Medical Center's administrative teams began using Microsoft Copilot in late 2023 to help non-clinical staff manage documentation, meeting summaries, and internal communications. The HR and operations managers overseeing these teams faced a specific challenge: healthcare has strict rules about patient data, and staff were nervous about accidentally including protected health information (PHI) in AI tool inputs. Rather than banning Copilot, the managers created a simple rule they called the "no names, no numbers" policy, no patient names, no record numbers, no dates of birth in any AI tool prompt, ever. Everything else was fair game.

The result was that administrative staff could use Copilot to draft HR policy summaries, create onboarding checklists, and generate meeting agendas, all without touching anything regulated. Productivity in the administrative division improved measurably, compliance incidents stayed at zero, and staff reported feeling more confident rather than more anxious about AI adoption. The managers didn't need to understand how Copilot worked technically. They needed to understand their team's specific risk profile and translate that into one clear, memorable rule. That's the job of a manager in an AI-using team: not to be the AI expert, but to be the context expert.

Three Teams, Three Approaches: What Works and What Doesn't

Click to enlarge

A Sales Manager Who Stopped Ignoring It

Marcus leads a seven-person B2B sales team at a logistics software company in Atlanta. For most of 2023, he noticed his top rep was consistently sending more follow-up emails than anyone else on the team, and closing at a higher rate. When he finally asked, the rep admitted she'd been using ChatGPT Plus to draft personalized follow-ups after every call, feeding it her call notes and asking it to generate a concise, warm email that referenced the prospect's specific pain points. She was spending about four minutes per email instead of fifteen. Marcus's first instinct was concern: was this authentic? Would clients feel manipulated if they knew?

Instead of reacting with a policy ban, Marcus did something smarter. He asked his top rep to run a 30-minute session showing the rest of the team exactly how she used ChatGPT for follow-ups. He then created a shared folder of their best-performing email templates, human-written and AI-drafted, and asked the team to tag which was which. Within a month, three more reps had adopted similar workflows. Average follow-up email time dropped from 14 minutes to 6 minutes across the team. Marcus didn't become an AI expert. He became a connector, someone who spotted good practice, made it visible, and gave the team permission to adopt it systematically. That's a leadership skill, not a technical one.

Click to enlarge

What Managing AI-Using Work Actually Looks Like Day to Day

Managing an AI-using team doesn't require a new management philosophy. It requires applying familiar management instincts to a new category of work output. When you review a report your team produced, you already ask: Is this accurate? Is this appropriate for the audience? Does it reflect our standards? You need to ask those same questions when the report was drafted with Copilot or Claude Pro, and you need to make sure your team knows you're asking them. The biggest mistake managers make is treating AI-assisted work as somehow self-certifying, as if the polish of the output means the review step was already done.

Practically, this means building AI review into the workflows you already manage. If your team uses a project management tool like Asana, Monday.com, or Trello, add a simple status label: "AI-assisted, under review" and "AI-assisted, approved." If you run weekly check-ins, add one standing question: "Did anyone use AI to produce something significant this week? How did it go?" This normalizes disclosure without making it feel like surveillance. It also surfaces problems early, before a client sees a hallucinated statistic or a prospect receives an email with the wrong company name.

The most effective managers of AI-using teams also build what you might call a small wins library, a running shared document or folder where the team captures AI prompts and workflows that produced genuinely good results. One good prompt for writing a client proposal introduction is worth sharing. One workflow that cuts meeting prep from 45 minutes to 10 minutes is worth documenting. Over time, this library becomes a team asset that accelerates onboarding, reduces inconsistency, and gives you concrete evidence of AI's actual impact on your team's output, which matters enormously when you're making the case to senior leadership for resources or recognition.

Key Lessons from Part 1

1. AI adoption on your team is probably already happening, the question is whether it's happening with or without your guidance.
2. The Brafton case shows that productivity gains from AI are real, but they collapse quickly without a human review step built into the workflow.
3. Vanderbilt's 'no names, no numbers' rule demonstrates that effective AI governance doesn't need to be complex, one clear, memorable rule can prevent the highest-risk behaviors.
4. Marcus's sales team story proves that the best AI practices often already exist on your team. Your job is to spot them, name them, and make them standard.
5. The comparison table shows that the manager's approach, not the AI tool itself, determines whether outcomes are positive or problematic.
6. Every piece of AI-assisted work needs a named human accountable for its accuracy and appropriateness before it leaves the team.
7. The 'Permission + Standard' formula resolves the most common management failure: either ignoring AI use entirely or banning it, both of which produce worse outcomes than structured adoption.
8. Practical AI management means adding lightweight signals to workflows you already run, project tags, check-in questions, and a shared wins library, not building a new bureaucracy.

When AI Becomes Part of the Team Dynamic

Klarna's managers faced a challenge that has nothing to do with software configuration or model selection. Their challenge was human: how do you redefine contribution when a tool can do 60% of what used to define someone's job? The agents who thrived were the ones whose managers gave them a clear new mandate, handle escalations, build customer relationships, catch AI errors. The ones who struggled had managers who hadn't yet figured out how to answer the question: 'So what exactly do you need from me now?' That question, left unanswered, quietly destroys morale.

The principle that emerges from Klarna's experience isn't about AI capability. It's about role clarity under change. When AI enters a workflow, every person on the team silently asks the same question: does this make me more valuable or less? Your job as a manager is to answer that question loudly and specifically, before people answer it themselves with the wrong conclusion. The teams that performed best at Klarna weren't the ones with the most sophisticated AI setup. They were the ones where managers had explicitly redesigned what human excellence looked like in an AI-assisted environment.

The Trust Calibration Problem

At a mid-sized marketing agency in Chicago, a content team started using Claude Pro to draft first versions of client campaign briefs. Within six weeks, the team lead noticed something troubling: junior copywriters had stopped pushing back on AI drafts entirely. They were editing for tone and grammar but accepting the strategic framing wholesale, even when it was generic, off-brand for the client, or factually imprecise about the industry. The AI sounded confident, the output looked polished, and the junior staff didn't yet have the professional experience to spot what was subtly wrong. Trust had become uncritical acceptance.

This is the trust calibration problem. AI tools produce output that looks authoritative. The formatting is clean, the language is fluent, the structure is logical. For someone early in their career, or even a seasoned professional in an unfamiliar domain, that polish creates a halo effect. The output feels more credible than it is. The Chicago agency's team lead solved this by introducing what she called a 'challenge round': before any AI-drafted brief went to review, the copywriter had to identify at least two assumptions the AI made and write one sentence explaining why each was defensible or questionable. This wasn't about distrusting AI. It was about building the habit of professional judgment alongside it.

The deeper lesson here is about what AI does to skill development on your team. When a tool handles the first draft, the analytical work that used to happen during drafting, the research, the framing, the judgment calls, can quietly disappear from the workflow. Your team gets faster but potentially shallower. Strong managers build deliberate friction back into AI-assisted processes, not to slow things down, but to ensure the thinking still happens. The goal isn't to make AI harder to use. It's to make sure the humans using it stay sharp.

Click to enlarge

How a Hospital Network Solved the Consistency Problem

Mayo Clinic's administrative teams began piloting Microsoft Copilot in 2023 for summarizing patient intake documentation and drafting internal communications. The early results showed a familiar pattern: output quality varied enormously depending on who was using the tool and how. One department head was getting crisp, accurate summaries. Another was getting verbose, occasionally inaccurate ones that required more editing than the original work would have taken. Both teams had access to the same tool. The difference was entirely in how they were prompting it, and nobody had standardized that.

Mayo's solution was practical and non-technical. They created a shared document, essentially a prompt library, that the team lead maintained and updated weekly. It contained fifteen approved prompts for their most common tasks, each with a brief note explaining what the prompt was designed to produce and what to watch out for in the output. New staff were onboarded to the library before they used Copilot independently. Within two months, output consistency improved significantly and the time spent editing AI drafts dropped by roughly a third. The insight was straightforward: standardizing how people talk to AI tools is a management task, not a technical one.

Comparing Management Approaches to AI Adoption

Click to enlarge

What a Sales Director Learned the Hard Way

A regional sales director at a commercial real estate firm rolled out ChatGPT Plus to her team of twelve account managers in early 2024. The pitch to the team was simple: use it to draft outreach emails and property summaries faster. Within a month, outreach volume had doubled. Response rates had dropped by 22%. The emails were longer, more generic, and, her clients told her directly, felt less personal than the handwritten notes her top reps had been famous for. The tool had optimized for volume. The director had forgotten to tell the team that volume was never the point.

She course-corrected by doing something most managers skip: she sat with her top two performers and watched them use the tool in real time. What she found was that her best reps were using AI for research and structure but rewriting the opening and closing lines themselves, the parts where their personal knowledge of the client came through. She turned this into a team standard: AI handles the body, the rep owns the first and last line. Response rates recovered within six weeks. The lesson wasn't that AI hurt her sales team. It was that she had deployed it without specifying what 'good' looked like in an AI-assisted workflow.

Building a Review Process That Actually Works

Most managers approach AI output review the same way they'd approach proofreading, scan for obvious errors, fix the language, move on. This misses the most important failure modes. AI tools don't typically produce output that's obviously wrong. They produce output that's plausible but subtly off: a market size figure that's two years out of date, a client recommendation that doesn't account for a constraint the AI wasn't told about, a tone that's technically professional but reads as cold to someone who knows the recipient. These errors pass a surface-level review every time.

An effective review process for AI-assisted work operates at two levels. The first is factual verification: any specific claim, number, name, or date that the AI generated should be independently confirmed before it leaves your team. This sounds obvious but is routinely skipped under time pressure. The second level is judgment review: does this output reflect the actual context, relationships, and constraints that a human in this role would know? This is where AI consistently falls short, and where your team's expertise is genuinely irreplaceable. Build both checkpoints into your workflow explicitly, they won't happen automatically.

One practical structure that works well across industries is a two-minute 'AI output card', a brief mental checklist your team applies before finalizing any AI-assisted work. It has four questions: Is every specific fact in here verified? Does this reflect what we actually know about this client, situation, or audience? Would a senior person on this team be comfortable putting their name on it? And, critically, does anything here feel too easy, too smooth, or too generic? That last question is the most useful. When AI output feels frictionless, it's often because it's been averaged across thousands of similar examples and lost the specific edge your situation requires.

Click to enlarge

Key Lessons from This Section

1. Role clarity is a performance driver, not just an HR concern, when AI changes a workflow, redefine human contribution immediately and explicitly.
2. Trust calibration is your responsibility: train your team to challenge AI output professionally, not accept it because it looks polished.
3. Skill atrophy is a real risk, deliberately build moments of human analyzis back into AI-assisted workflows to keep your team sharp.
4. Standardizing prompts is a management task, not a technical one, a shared prompt library maintained by the team lead dramatically improves output consistency.
5. Deploy AI with a definition of 'good', without it, your team will optimize for the wrong metric (usually volume or speed).
6. Watch your best performers use AI tools and systematize what they do, peer-derived norms spread faster and stick longer than top-down policy.
7. Effective AI output review operates at two levels: factual verification and judgment review, surface-level proofreading misses the most important failure modes.

Building a Team Culture Where AI Makes Everyone Better

When Klarna announced in 2024 that its AI assistant was doing the work of 700 customer service agents, the headlines focused on displacement. But the quieter story inside the company was more interesting. The human agents who remained weren't doing less, they were handling the complex, emotionally charged cases that AI couldn't navigate. Escalations. Fraud disputes. Customers in genuine distress. Klarna's managers had to learn something fast: how do you lead a team where the definition of 'good work' has fundamentally shifted? The answer wasn't a policy memo. It was a cultural reset about what humans are actually for.

The principle Klarna's team leaders extracted from that pressure was blunt: when AI handles the routine, humans must own the exceptional. That sounds obvious until you're managing a team of five people who spent three years getting good at routine work. Their identity, their confidence, their performance metrics, all of it was built on tasks AI now does in seconds. The manager's job became psychological as much as operational. Helping people see that their value had moved upstream, not disappeared, required constant, specific, evidence-based reinforcement. Not pep talks. Real examples of humans catching AI errors, making judgment calls, and building client trust in ways no model could replicate.

This is the core cultural challenge for managers leading AI-using teams: you are not just adopting a new tool, you are redefining what competence looks like on your team. The teams that get this right don't just become more productive, they become more confident, because every person understands exactly what they bring that the AI cannot. That clarity is your job to create.

What High-Trust AI Teams Actually Look Like

At Unilever, HR teams began using AI tools to screen CVs and draft first-round interview summaries. Early adoption was uneven, some recruiters embraced it, others quietly ignored it. The turning point came when one regional HR manager, Priya Nair's team in Southeast Asia, started running weekly 15-minute 'AI check-ins.' Not training sessions. Conversations. What did the AI get wrong this week? Where did you override it and why? What prompt worked better than expected? Within three months, her team's output had doubled and error rates on offer letters dropped by 40%. The rest of Unilever's HR function eventually borrowed the format.

The check-in format worked because it did three things simultaneously. It normalized talking about AI failures without blame. It created a shared vocabulary for what good AI-assisted work looked like. And it positioned every team member as an expert, because their lived experience with the tool was the most valuable data in the room. The manager wasn't the AI expert. The team collectively was. That inversion of the usual knowledge hierarchy is something smart managers learn to create deliberately.

The broader lesson here is about psychological safety. Teams that feel safe admitting 'the AI gave me something wrong and I nearly sent it to a client' will catch more errors than teams where AI mistakes feel like personal failures. Your job is to make the near-miss stories celebrated, not hidden. One manager in a London consulting firm started every Monday standup with 'AI save of the week', a 60-second story about a human catching an AI error. It became the most popular part of the meeting.

Different Roles, Different Relationships With AI

At a mid-sized US school district, curriculum coordinators started using Google Gemini to draft lesson plan templates and differentiation strategies for teachers. The coordinator who got the best results wasn't the most tech-comfortable one, it was a veteran of 22 years named Marcus, who had deep subject matter expertise and used AI as a drafting assistant rather than an answer machine. He knew exactly where the AI's suggestions were generic and needed specificity. His less experienced colleagues, lacking that subject knowledge, sometimes couldn't tell the difference between a good AI suggestion and a plausible-sounding bad one.

Marcus's experience illustrates something managers often miss: AI amplifies existing expertise. It doesn't substitute for it. A junior team member using AI may produce output that looks polished but contains subtle errors only a domain expert would catch. This means your most experienced people are still your most valuable quality filters, but only if you structure the work so their judgment gets applied. Routing AI-generated outputs through senior review isn't bureaucracy. It's quality control for a new kind of workflow.

Putting It Into Practice This Week

Three practical moves work faster than any formal training program. First, make AI use visible. Ask your team to share one AI-assisted output in your next team meeting, not to evaluate it, but to normalize the conversation. The simple act of showing work created with AI reduces the secrecy and anxiety that builds up when people are unsure whether using AI is encouraged or frowned upon. Your visibility as a manager who takes this seriously sets the tone more than any policy.

Second, create a team prompt library. A shared Google Doc or Notion page where anyone can add a prompt that worked well is a low-effort, high-return investment. Within a month, you'll have a custom resource built from your team's actual work, far more useful than generic AI guides. Assign someone to own it, but make contribution voluntary. The prompts people share voluntarily are the ones that actually get used.

Third, update how you define good work. If your team's performance metrics still measure outputs that AI now produces in seconds, your metrics are measuring the wrong things. Work with your team to identify two or three human-specific contributions that matter most in your context, client relationship depth, judgment quality, creative direction, error-catching, and find ways to make those visible in how you give feedback and recognition. The teams that thrive with AI are led by managers who updated their definition of excellence before the team had to ask.

Click to enlarge

Key Lessons From This Section

1. When AI handles routine work, humans must own the exceptional, your job is to make that transition explicit and supported.
2. Psychological safety around AI mistakes is a performance variable, not a soft skill. Teams that share errors catch more of them.
3. Weekly AI check-ins, short, conversational, focused on real work, outperform formal training programs for building team capability.
4. AI amplifies existing expertise. Junior staff using AI without strong domain knowledge are a quality risk. Structure your review process accordingly.
5. A shared prompt library built from your team's actual work is more valuable than any generic AI guide.
6. Your performance metrics need to catch up with what AI has changed. Measure human judgment, not just output volume.
7. Picking one tool and going deep beats adopting five tools and staying shallow, especially in the first 60 days.

Takeaways

• Leading an AI-using team is a cultural challenge as much as an operational one, clarity about human versus AI roles is the foundation.
• The best managers of AI-augmented teams model good AI habits publicly and create space for the team to learn collectively.
• Human value moves upstream when AI handles the routine, your job is to help your team see and own that shift.
• Small, consistent practices, check-ins, shared logs, visible prompt libraries, compound into significant team capability over months.
• Updating your definition of excellent work is the most important leadership move you can make as AI becomes part of your team's daily workflow.