89% of Executives Say AI Is Making Their People Faster. Only 6% Can Find the ROI. The Missing Piece Is the Team.

Everyone Got Faster. Almost Nobody Got Value.

At VB Transform 2026 in July, Atlassian shared a number that should stop every operator cold. Its annual State of Teams Report surveyed 12,000 global knowledge workers and interviewed roughly 200 Fortune 1000 executives. 89% of those executives said individuals in their companies are speeding up with AI. Only 6% could point to a specific example of clear ROI.

Read that gap again. Nearly nine in ten leaders can feel the acceleration. Fewer than one in ten can find where it pays off.

The instinct is to blame the model. That instinct is wrong. The models are fine. The problem is the unit of adoption. Companies deployed AI to individuals — a subscription per person, a chat window per desk — and then wondered why the organization didn't move.

Dr. Molly Sands, who leads Atlassian's Teamwork Lab, put it plainly: most companies are approaching AI adoption backwards, optimizing how individuals use AI instead of how teams work together. That single reframe explains the 6%. It also explains why the same week OpenAI announced Presence, a platform for running governed agents inside a business under company-defined policies and escalation rules. The whole industry is arriving at the same conclusion at once: individual AI is a dead end. The value is at the team level.

Why Speeding Up Individuals Makes Things Worse

Here is the counterintuitive part. Making every individual faster, on its own, does not make the business faster. In many cases it makes it slower.

Sands described what happens when you accelerate people who are pointed in slightly different directions: they "very quickly start to crash into each other." A faster salesperson generates leads the ops process can't absorb. A faster analyst produces reports nobody asked for. A faster developer ships code that fails review. Local speed, global friction.

The Atlassian data confirms it. Roughly 14% of teams had turned AI usage into real value. That means a single company can hold a handful of high-performing teams surrounded by teams seeing no return at all. The difference wasn't which model they used or how clever their prompts were. It was whether the team had a shared way of working with AI, or whether every person was quietly running their own private experiment.

That last part is the real cost. When each worker develops their own prompts, their own agents, and their own assumptions, they create a new layer of unspoken knowledge that never translates into organizational performance. You didn't buy a capability. You bought 40 disconnected capabilities that don't talk to each other.

What Bad Looks Like

A 30-person company buys everyone an AI subscription. Adoption is celebrated. Six months later:

• The marketing lead has a favorite prompt library nobody else can see.
• Two ops people built spreadsheet macros with AI that only they understand.
• The founder uses it to draft emails and assumes everyone is getting the same lift.
• Nobody can answer the question "where did AI actually save us money?"

Every one of those individuals got faster. The company got a pile of shadow processes and a leader who can't find the ROI. This is the 6% problem in miniature.

What Good Looks Like

The same company deploys AI as a coworker that belongs to the team, not to a person. The AI coworker has access to the shared context — the goals, the past decisions, the workflows — and it runs the same way whether the marketing lead or the founder is talking to it. When it hits something ambiguous, it escalates to a named human. Its work is visible to everyone who needs to see it.

Now the question "where did AI save us money?" has an answer, because the work happened in one place, on shared context, against a defined workflow. That is the 14%. And it is reachable on purpose, not by luck.

The Three Things Winning Teams Have

Atlassian found the leading teams shared three characteristics: context, workflows, and culture. These aren't soft ideas. They map directly to how you have to build if you want AI to produce value instead of noise.

Context. The winning teams built what Atlassian calls a context graph — goals, decisions, and organizational knowledge captured in shared digital records instead of trapped in individual memory. This is the single most important finding in the report. AI is only as useful as the context it can reach. An AI tool that lives in one person's chat history has access to one person's context. An AI coworker plugged into the team's shared knowledge operates with the full picture.

Workflows. The teams pulling ahead redesigned entire end-to-end processes rather than accelerating isolated tasks. This is the difference between "AI helps me write this email faster" and "AI runs the follow-up sequence, logs the outcome, and flags the deals that need a human." One speeds up a keystroke. The other owns a piece of the operation.

Culture. The fastest-moving teams worked under leaders who explicitly encouraged experimentation and made it clear that some experiments would fail. Nobody figures out where the human-AI boundary sits without trying things that don't work. That's not recklessness. That's how you calibrate.

Notice what all three have in common: they are properties of a system, not of a person. You cannot buy context, workflows, and culture by handing out subscriptions. You have to build the layer that holds them.

Individual AI Tools vs. Team AI Coworkers

This is the distinction that determines whether you land in the 6% or the 14%.

An AI tool does what you tell it, in the moment, then forgets. It has no persistent context, no shared workflow, no identity beyond the current chat. It is genuinely useful for the individual holding it — and structurally incapable of producing organizational value, because organizational value requires shared state.

An AI coworker is different by design. It persists. It remembers the goals and decisions that matter. It runs the same workflow every time. It shows up in the channels the team already uses. It has a defined role and a defined set of things it escalates rather than guessing. It is, functionally, a member of the team — which is exactly why it can produce value at the team level.

The tell is in the Atlassian recommendation itself. They found that teams which adopted explicit "AI working agreements" — deciding what they'd use AI for, what they'd deliberately avoid, which agents they'd share, and what common skills would keep everyone on the same context — used AI more, moved faster, and produced higher-quality work.

That is intent engineering. Encoding what the organization wants the AI to do, what it should refuse, and where it should stop and ask. When that intent lives in a shared system instead of 40 private chat windows, the 6% becomes the 14%. Our own take on why the objective matters more than the model is in why 74% of businesses see no ROI from AI, and the tool-versus-coworker line is drawn in full in AI coworker vs. AI tool.

The Governance Point Everyone Is Arriving At

The OpenAI Presence launch the same week is worth sitting with, because it confirms the direction even though it's built for a different buyer. Presence packages policies, system connections, evaluations, guardrails, and escalation rules so agents behave reliably in production as conditions change. Each deployment starts with a defined job. The agent gets only the information and systems it needs. It can take approved actions and escalate to a human under company-defined policy.

Strip away the branding and that is a description of a team member with a role, permissions, and an escalation path. It is the anti-individual model. It is deployed to the organization, governed by the organization, and evaluated against the organization's standards.

The catch: Presence is not self-serve. It runs on OpenAI's models at the core, and deployments are led by OpenAI Forward Deployed Engineers and select systems integrators. That's a fine fit for a Fortune 1000 company with an integration budget. It's the wrong shape for the SMB that makes up most of the economy — the business that needs team-level AI without an embedded engineering team and without betting the operation on a single model vendor.

That's the whole reason model-agnostic infrastructure matters. When the winning move is to build shared context and durable workflows, the last thing you want is for that context to be locked inside one provider's app. We wrote about that trap directly in why your AI coworker shouldn't live in someone else's app.

How to Move Your AI From Individual Hack to Team Advantage

You don't need a Fortune 1000 budget to land in the 14%. You need to change the unit of deployment. Here's the concrete path.

1. Stop counting seats. Start counting roles. Instead of "who has an AI subscription," ask "what job do we want an AI coworker to own." A role has a defined scope, a defined output, and a defined escalation point. A seat has none of those.

2. Build the shared context first. Write down the goals, the recurring decisions, and the workflows the AI coworker needs to operate on. This is your context graph. It doesn't have to be fancy — it has to be shared and durable, not stuck in one person's head.

3. Redesign one whole workflow, not one task. Pick a process that runs end to end — lead follow-up, invoice reconciliation, permit status updates — and hand the coworker the whole thing, with a clear human checkpoint. Accelerating a single task inside a broken process just breaks it faster.

4. Write the AI working agreement. Decide, as a team, what the coworker will do, what it will deliberately not do, and when it stops and asks a human. Atlassian's data says this one practice separates the teams that win from the teams that spin.

5. Give it a home you control. Put the coworker on infrastructure that keeps your context and lets you change models as the market moves. Shared context is your most valuable asset. Don't rent it inside someone else's subscription.

Do those five things and the ROI question stops being a mystery. The work is happening in one place, on shared context, against a defined workflow, with a human in the loop where judgment matters. That's a system you can measure — and improve.

FAQ

Q: Why does giving everyone an AI subscription not produce ROI? A: Because organizational value requires shared state, and individual subscriptions don't have any. Each person builds private prompts, private agents, and private assumptions that never connect. You get 40 disconnected capabilities instead of one system. Atlassian found only 6% of executives could point to clear ROI despite 89% seeing individuals speed up — that gap is the cost of deploying to individuals instead of to the team.

Q: What is a "context graph" and do I need one? A: It's shared organizational knowledge — goals, decisions, workflows — captured in durable records instead of living in individual memory. It's the single strongest predictor Atlassian found for teams that turn AI into value. You don't need enterprise software to start. You need the context written down somewhere your AI coworker can actually reach it, and kept current.

Q: What's the difference between an AI tool and an AI coworker? A: A tool does what you tell it right now, then forgets. A coworker persists, remembers the context that matters, runs a defined workflow, holds a role, and escalates to a human when it hits something outside its scope. Tools help individuals. Coworkers produce value at the team level because they operate on shared state.

Q: Isn't OpenAI Presence exactly this? Why not just use that? A: Presence is a genuine team-level, governed agent platform — and it validates the whole direction. But it's not self-serve, its deployments are led by OpenAI engineers and integrators, and it runs OpenAI models at the core. That fits a Fortune 1000 company with an integration budget. Most businesses need team-level AI without embedded engineers and without locking their shared context to one model vendor.

Q: We're a small business. Is team-level AI even realistic for us? A: It's more realistic for you than for a large enterprise, because you have less legacy process to untangle. You don't need a big budget — you need to deploy AI as a shared coworker with defined roles and durable context instead of handing out individual subscriptions. The five-step path above is built for exactly that.

Q: How do we actually measure AI ROI once it's at the team level? A: When the work happens in one place, on shared context, against a defined workflow, ROI becomes measurable: time on the workflow before and after, output quality, escalation rate, cost per completed job. The reason 94% of executives can't find ROI is that individual AI usage leaves no measurable trail. Team-level deployment creates one.

Start Running a Real Team of AI Coworkers

The 6% and the 14% aren't separated by a better model. They're separated by whether AI lives in one person's chat window or in a shared system with context, workflows, and a defined role. If you're ready to stop using AI tools and start running a real team of AI coworkers, Associates AI Teammates gives you a 14-day free trial with no credit card required. Start your free trial at associatesai.team.