You have a few wins. How do you spread them?
Whether you're the CEO trying to spread early wins, the AI champion who built them, or the operator trying to move both — this page is for you.
Phase 3 is the make-or-break stage. A handful of people on your team are doing real work with AI. You can see it. The rest of the company can't, won't, or doesn't yet know how. The gap between your power users and everyone else is widening — and you're starting to suspect that gap is about to become your biggest organizational problem.
The hardest part: your AI champion has no formal title for the role, no real authority to spread their wins, and no time inside their existing job to do it. That's not a personal failure. It's a structural one.
Most companies that reach this phase fall back to Phase 2 because they treat the early wins as the destination instead of the starting line. The pilot succeeded — great, now what? Without a system for harvesting what works and pushing it across the org, the wins stay local. The power users either get frustrated and leave, or they quietly stop sharing.
This page is what to do about that.
You might be here if…
- You can name your power users. There are two or three people on your team who are visibly faster, sharper, or more productive than they were a year ago — and you know AI is part of why. The rest of the company hasn't caught up.
- At least one team has a documented AI workflow that works. Not just "someone uses ChatGPT." An actual repeatable process: a prompt, a way the output gets reviewed, a checkpoint that catches errors. It works. It's run more than a few times. It's local to one team.
- You're starting to talk about what AI means for hiring. Backfilling a role and quietly asking whether you actually need it. Or asking whether the role description should change. Or whether the next hire should be different from the last one. The conversations are tentative; nobody's said it out loud yet.
If most of these don't sound like you, the page you actually want is probably another phase — see the four phases →
Why companies stall here
The structural problem in Phase 3 is that the people who get the most value from AI are usually the people the company is least equipped to learn from.
Your AI power users are typically not your most senior people. They're often mid-level operators who got curious, spent their weekends figuring it out, and quietly built workflows that are dramatically more productive than the official process. They're usually not in management. They don't have authority to redesign anything beyond their own desk. And their wins don't propagate because there's no formal mechanism for them to.
Meanwhile, the people who could spread their wins — managers, function leads, executives — are usually not the power users. They're a step removed. They can see the wins exist but they don't have the lived experience of building them. So they don't know what to ask for, what to fund, or what to structurally change.
The peer-reviewed evidence on this is striking
A 2023 NBER study by Brynjolfsson, Li, and Raymond — published in the Quarterly Journal of Economics in May 2025 — tracked 5,179 customer support agents at a large software company through a staggered rollout of a generative-AI assistant. They found a 14% productivity gain on average — but the distribution was the real story: 34% productivity gains for novice workers, near-zero for the most experienced.
AI raises the floor, not the ceiling. Your most experienced people benefit least. Which means the right Phase 3 strategy isn't "ask your best people to use AI more" — it's "use AI to bring everyone else up to your best people's level."
The other piece of research that matters here is from Harvard Business School and BCG: 758 BCG consultants ran 18 realistic tasks, half with GPT-4 and half without. Inside the AI's capability frontier, the consultants with AI completed 12.2% more tasks, 25.1% faster, with 40% higher quality outputs. Outside the AI's capability frontier — tasks that looked similar but were actually beyond what the model could handle — the AI-using consultants performed 19 percentage points worse than those without AI. The researchers called this the "jagged frontier."
That's the second Phase 3 unlock. The job isn't to use AI more or less. The job is teaching your team to recognize when AI helps and when it hurts. That's the heart of what an AI translator does inside an organization, and it's a skill that doesn't develop on its own.
How to move from one win to ten
Five moves, ordered from free / this week to planning / this quarter. The first two cost nothing and require no permission — anyone reading this can act on them by Friday. The last two are real organizational changes that need leadership commitment.
- 01 free / this week
Identify your power users — by name
Every Phase 3 company has 2-5 employees doing things with AI that the rest of the company would call magic. You probably already know who they are. If you don't, ask three managers "who on your team is using AI in ways that surprise you?" The answers come back fast.
Once you have the list, do one of two things: (a) ask each person to write a 1-page note on what they're doing and how, then circulate it; or (b) put 30 minutes on the calendar with each one to watch them work. The goal isn't to extract their playbook to standardize — it's to surface what's actually happening so you can fund more of it.
Free resource: Power User Identification Worksheet (coming soon) - 02 free / this week
Run a 90-minute department audit
Pick one department — the one where AI usage is most visible. Sit with the team lead for 90 minutes and answer three questions: (1) Where is AI saving time today? (2) Where would AI save time but isn't being used? (3) Where has AI been tried and failed?
Don't try to fix anything in this meeting. Just write the answers down. The output is a one-page snapshot per department. If you do this for three departments, you'll see patterns: the same wins, the same gaps, the same failure modes. Those patterns are your roadmap.
Free resource: Department Use-Case Catalog (coming soon) - 03 low cost / this month
Stop running pilots like science fairs. Start running them like products.
Most Phase 3 pilots fail not because the AI doesn't work, but because the pilot has no owner, no metric, no end date, and no decision waiting at the end of it. "Let's see what happens" is not a pilot — it's a slow death.
A real pilot has: a named owner who'll lose face if it fails, a single metric that defines success, an explicit time box (4-8 weeks is the sweet spot), and a pre-committed decision at the end ("we'll either roll this out to the rest of the team or kill it"). If you can't define those four things, you're not ready to pilot — you're still scoping.
Free resource: Pilot Program Template (coming soon) - 04 low cost / this month
Run a Demo Day — internal only, 60 minutes, monthly
The single highest-leverage event for spreading AI inside an SMB. Invite anyone who wants to come (no required attendance). Three or four people get 5-10 minutes each to show one thing they've built or learned. No slides, no theory, just "here's what I do, here's how I do it."
The first one is small and slightly awkward. The third one has 25 people in the room. By the sixth, your power users are competing for who has the most useful demo. Recordings get shared in Slack. People build off each other's ideas. This is how culture changes — not from a memo, from monthly proof.
Free resource: Demo Day Playbook (coming soon) - 05 planning / this quarter
Redesign one workflow — really redesign it
McKinsey's research on what separates high-performing AI organizations from the rest says it cleanly: 6% of companies are capturing disproportionate value, and the single biggest predictor is workflow redesign. Only 21% of gen-AI users have redesigned at least some workflows. Most are layering AI on top of how the work used to flow — capturing maybe 10% of the available value.
Pick one workflow. Map the current state in detail. Then ask: "if AI did the steps it can do well, what would the workflow actually look like?" Often the answer is fewer humans, fewer handoffs, faster cycle times — and a different shape entirely. This is hard work and you can't do it for every workflow at once. Pick one. Do it right. Then do another next quarter.
Free resource: Workflow Redesign Worksheet (coming soon)
When the AI champion gets a real role, the wins spread
Here's the pattern, drawn from the Phase 3 SMBs we see most often. A 320-person B2B SaaS has three power users in their customer-success operations team. One of them has built a Claude-based workflow that triages inbound support tickets, drafts first responses, and flags the ones that need human escalation. It cuts average first-response time by roughly 60% and frees the team to spend more time on the hardest cases.
The CEO knows it exists. The COO knows it exists. Neither of them uses AI themselves and neither knows how to spread the workflow to the other departments that could benefit from similar approaches — sales operations, account management, the data platform team. The power user is getting frustrated. She's built something valuable and watched it sit local for five months.
The unlock is usually small: she gets promoted into a half-time "AI enablement" role inside ops. Two days a week she stays in her CS-ops job; three days a week she runs what she calls "office hours" — basically helping anyone in any department who wants to figure out an AI workflow. Eight months later, every customer-facing function in the company has at least one documented AI workflow. The CEO will tell you, in the retrospective, that it was the single highest-ROI organizational change of the year.
The pattern is repeatable. You usually don't need to hire — you need to redesign one role to give your existing power user the time and the mandate to spread their work.
If you'd rather not run these moves alone — Phase 3 is where most of Sillewa's embedded operator work happens. Here's what that looks like →
What changes when you move to Phase 4
When the moves above land — when you have multiple departments running real AI workflows, when your power users are visible and supported, when at least one workflow has been redesigned around AI rather than just enhanced by it — you'll start running into a different set of problems.
You'll find yourself asking org-design questions: do we hire less, restructure, redirect, or reduce? You'll notice middle managers either accelerating or blocking the spread. You'll need formal AI governance because an employee with AI and bad judgment can damage your brand at scale. You'll be making harder calls than you've made all year.
That's Phase 4.
Read what to do in Scale & structureIf you found this useful, also read:
- Credits & sources — the primary research backing the claims on this page
- More about Sillewa — who we are and how we work
The other phases
Where you actually are matters. If Phase 3 isn't quite right, here's the rest: