Picking an AI platform touches your data, your security posture, and your culture. Here's how operators are thinking through platform choice, data safety, and team adoption.
This piece is adapted from an Operators Guild Focus Session on AI adoption, led by Jennifer Ybarra and Aram Fischer. The conversation drew operators from companies of every size, from two-person consultancies to some of the largest tech companies in the world, all comparing notes on the same questions: which platform, how deep to integrate, and how to get teams to actually use it.
Focus Sessions are small-group, member-only conversations where operators compare notes on decisions in flight, pressure-test tradeoffs, and surface the operational realities that rarely show up in polished frameworks.
If you want access to sessions like this, including the recordings and the community behind them, you can apply to join OG.
Early on, the big AI companies positioned adoption like signing up for Canva. Log in, poke around, figure it out. That framing has failed almost everywhere it's been tried.
Choosing an AI platform is an enterprise software decision. It interacts with your data, your documents, your tools, and your people. It implicates security, privacy, and labor. The more capable these systems get, the more true that becomes. Treat the choice the way you'd treat choosing a CRM, and invest deeply in the platform you commit to.
Depth of use matters more than headcount
A common question teams ask is at what team size should we consolidate onto one platform?
The better question is how deep your AI use goes.
A two-person team connecting AI to a CRM is making a bigger decision than a fifty-person team using chat for grammar checks. Once AI can read and change things in your systems of record, you want one platform you've vetted, secured, and invested in. You won't want five different tools touching your CRM.
A related warning on connectors: MCP servers vary widely in quality, and weak ones can expose your data in ways you won't see coming. Vet them before you connect them. In some cases, a traditional API integration is the more secure choice.
The harness matters as much as the model
Picture two teams building a race car. One chases the most powerful engine and ignores everything else. The other accepts a slightly less powerful engine and obsesses over suspension, tires, and aerodynamics. In AI, the second team wins.
The customization layer around a model, sometimes called the harness or scaffolding, has an outsized effect on results. Projects, custom instructions, integrations, and purpose-built models all shape whether outputs are right the first time. That matters for quality, and it matters for adoption. Half your staff has no interest in learning prompt engineering. The less work they have to do to get an acceptable output, the more they'll use the tool.
It's also an efficiency lever. As subsidized pricing gives way to metering and overages, a well-designed harness gets you to the same output with fewer tokens.
The atomic unit of value is the workflow
Nobody adopts AI generically. People adopt AI for a specific job.
Abstract training sessions on "how AI works" lose people quickly. What works is finding the person who sends a customer newsletter every week and showing them how AI turns three customized versions into half the work of one. Build the use case around the person's actual job, and they'll understand AI in a way no workshop can teach them.
Practical implications:
- Keep generic AI 101 training short
- Split people by function fast, since comms and data teams use these tools completely differently
- Anchor every rollout to a workflow someone already owns
Keep human judgment at the beginning and the end
Weaving AI into a workflow raises the stakes of getting it wrong. An agent with write privileges that hallucinates can do real damage, and the vast majority of agents in production today fail, usually because teams layered on too much automation with too little human judgment.
These are probabilistic systems. They're extremely good estimators, and treating them like calculators is how teams get burned. Test heavily before automating anything with downside risk, keep humans at the start and end of every workflow, and iterate in small steps so you can see where things break.
Know where your data goes
AI is exposing bad data practices that companies have gotten away with for years. A few principles from the discussion:
- People are using AI whether you sanction it or not, and shadow AI thrives when you haven't given teams a safe, approved option
- Expect zero privacy from free models
- The less you control the servers, the less you control the data, which is why many large companies run open source models on their own infrastructure for sensitive work
- There is no user privilege with an LLM, and bringing an AI notetaker into a privileged conversation can put attorney-client privilege at risk
- Bias is endemic to these systems, and in contexts like hiring it can quietly produce discriminatory outcomes
Before putting data into any third-party system, ask why it needs to be there at all and what happens once it is. Once it's in, it's in.
Disclose when it matters
Legal compliance varies by industry, so consult your own counsel. As a matter of best practice, the rubric is straightforward:
- Anyone interacting with a chatbot should know it's a chatbot
- Any AI manipulation of someone's likeness requires their consent
- Major decisions made or heavily shaped by AI deserve disclosure, along with a path to a human
- Ghostwritten content is different. If a principal reviews and signs their name to it, that sign-off is the accountability
More transparency earns more trust, with customers and with your own team.
Address resistance before it hardens
Resistance to AI clusters into a few categories: data privacy and security, bias, climate impact, workers' rights, and, for anyone close to the arts, copyright. These concerns are strongly held, and ignoring them turns skeptics into refusers.
Addressing them upfront, with genuinely good answers, does the opposite. It turns skeptics into curious users, curious users into dedicated ones, and dedicated ones into internal champions.
Then get people to a quick win as fast as possible. Alleviating a real pain point, like turning an hour of manual data entry into a two-minute task, is the moment adoption actually starts. Imagination kicks in from there.
Four conditions for adoption
The human side of adoption comes down to four conditions, a framework Jen developed through her work helping teams navigate change at The Good Human Group:
- Connection. People understand how the tool relates to their work and the mission
- Competency. People get the reps and the room to build skill without judgment
- Choice. People have real options within clear parameters, including the ability to raise their hand and try something different
- Consent. People know what they're signing up for and what the tool does with their work
When adoption stalls, the gap is usually in one of these, and it sits further upstream than most rollout plans ever look. You can have flawless process and still fail if people don't trust what's happening.
One more reality check: AI exacerbates whatever already exists in your culture. Fractious workplaces get more fractious. Poorly designed workflows get every flaw exposed. Culture is the core of adoption, for AI and for everything else.
Join the conversations operators need
This session was one example of the work happening inside OG every day.
If you want access to sessions like this one, including the recording, the discussion, and the operators comparing notes on what's working across their companies, consider becoming a member.
OG is where senior operators go to learn from peers, pressure-test decisions, and trade the practical insights that help teams and companies move faster.
If you're ready to be part of a community built for operators, you can apply to join OG today.