If you're leading AI transformation, 2027 planning is probably on your desk right now. It's also a harder planning cycle than last year's: AI costs more than most of us budgeted for, agents are running inside the business, and boards want proof that the investment is working.

We recently held an open office hours for leaders who are in the middle of this planning with our Head of AI, Michael Domanic. Here are all the big decisions we covered and the advice Michael gave.

1. My board wants to see real progress next year. How do I create reasonable ROI goals I can prove?

In past years, most progress questions centered on adoption and proficiency, which are relatively easy to answer. In 2027, boards will want to see where employees have the enablement to build automations for their own work and where work has actually been reinvented. In other words, you'll need to bring receipts.

The most sophisticated boards will go further and ask how you're redefining the way you do business, including the metrics that tell you whether the business is successful. That makes measurement the first decision in your plan, because everything else ties back to it.

Decision 1: What will we measure, and at what levels?

Adoption on its own won’t satisfy a board anymore, but you still need it, because these three levels together demonstrate a clear line from people using AI, to the workflows it changed, to the business results those changes produced.

So plan to have baselines and a methodology of measurement for the following:

• Adoption and proficiency: what baseline usage and skill look like across the workforce, and how you'll track them
• Workflows: for each workflow you bring automations or agents to, whether the goal is saving time, increasing impact, or driving a specific outcome
• Business outcomes: revenue, margin, headcount or revenue per employee

Decision 2: Which ROI can we actually prove?

Some AI value can be measured precisely and some can't, and your board talk track should be clear about which is which:

• If you already measure the workflow, you can report hard ROI – the before-and-after numbers once AI has had time to take hold.
• If you're seeing key business metrics move as AI rolls out across a function, you can report directional ROI – a correlation between your AI work and the improvement.
• If the gain is mostly productivity, like output per person rising 25% or not needing to hire as many people in a part of the business – you're reporting assumed ROI.

Directional and assumed ROI are still real value. They just don't come with a smoking-gun equation, so present them that way.

2. How much should I budget, and who should own it?

Decision 1: Central or distributed?

Central. There should be one enterprise AI budget for 2027, not costs scattered across functions.

Decision 2: How big should it be?

For a 1,000-person organization, we put a defensible budget at about $6.6 million, which is higher than most leaders expect. You can see my breakdown on this here. Scale it up or down for your headcount. (You can also use our AI budget worksheet.)

Decision 3: What goes in it?

• Platform licenses: the relatively predictable cost of giving people access to AI tools.
• Individual inference: 2–5% of salary per knowledge worker and 7–10% per engineer, based on what we saw in 2026. This is your biggest and least predictable line. Most organizations underinvested here this year, as more people got connected to tools, tools moved to consumption pricing, and agents were tied to critical business outcomes.
• Agent inference: centrally built agents run more inference than an individual's personal agent, so budget for them separately.
• A head of AI: the organizations that have made real progress have one. Transformation doesn't happen when a couple dozen people take it on as a side gig.
• Enablement: a coaching platform for day-to-day support, workshops, change management, a champions program and hackathons. We think champions should be compensated, even if it's through incentives rather than pay.

If you're tempted to trim, just don't trim enablement. It's still the most underinvested part of AI transformation, and it's what protects every other line. Many organizations that skimped on it in 2026 are now asking whether their AI spend produced any real movement, and the answer is either "we don't know" or "no."

3. Some teams are flying with AI and others have barely started. How do I rightsize the difference?

Every organization has this split. Token leaderboards aren't the answer, though highlighting people whose AI use is driving results is fine. What you do next depends on which end of the curve you're looking at.

Version 1: If a team is lagging

Plan to invest more time and resources with their managers for team enablement, and check whether that part of the organization has champion representation.

If it's more on an individual basis rather than a team, your Head of AI should have a direct, supportive conversation with them. When four of five people in a role are getting value from AI, the unlock for the fifth is typically just asking them what would help them find their use cases.

Version 2: If a team is far ahead

If someone is using 10x the tokens of their peers, assume they're doing something right rather than wrong. Have a conversation about what they’re doing and how you might actually be able to scale it.

When one of our salespeople was consuming far more than the rest of the team, it turned out she was doing something everyone on the team should have been doing, and we found ways to optimize the spend along the way.

4. Which workflows should we redesign, and who should do it?

The organizations that gained the most momentum in 2026 moved beyond individual AI use to redesigning whole workflows. Build a repeatable model for this into your plan, so progress doesn't depend on whoever happens to be enthusiastic.

Decision 1: Should this workflow exist at all?

Imagine you were building the business around AI from scratch.

• If the workflow can be worked around, cut it rather than automating it.
• If it would still be needed, look at how AI changes it and who it affects, then redesign it.

Decision 2: Who's doing the work?

Two groups in conjunction: 

1. The people who run the workflow day to day and know it inside out
2. Your builders (your head of AI, or champions with build skills).

5. How do I actually get agents built?

Outside of heavily regulated environments, everyone should be building agents, to some degree and at the right scale – organizations that switch on agent building without a structure tend to find a mess six months later.

Decision 1: Is the agent personal, team-level, or cross-functional?

Match who builds to who the agent affects:

• If it only affects the builder, anyone can build it. Every employee should be able to create personal agents, like a chief-of-staff agent or one that handles a small recurring workflow.
• If it affects others on the same or adjacent teams, that's where champions come in.
• If it's cross-functional or high-impact, a smaller group of champions builds it with your head of AI and function leaders.

Decision 2: Where do the champions sit? 

Last year, champions mostly helped with individual enablement. Now they’re creator enablers who give the rest of the workforce a path to building. 

They should also manage agents once they're live: monitoring output, spotting agents that aren't meeting expectations, and shutting them down.

6. How do I decide which agents to prioritize?

Pre-launch decision: blast radius and reversibility

Ask what happens if the agent goes off the rails, and answer two questions:

• Blast radius: how much damage could it do?
• Reversibility: how quickly could we undo it?

Based on your answers, here’s what you’re looking at:

• Low blast radius, high reversibility: Start here. A chief-of-staff agent that drafts a daily digest into your email drafts folder is a good example. If it gets something wrong, you'll notice and correct it, and there's nothing to undo.
• High blast radius, low reversibility: Your riskiest builds. An agent sending customer outreach or marketing messages sits here, because a wrong or offensive message is hard to call back once it's out.

High-risk agents aren't off-limits. Some will be valuable enough to justify the risk, so weigh what could go wrong against the metric you're trying to move, and decide from there.

The same lens answers the question of semi-autonomous vs. fully autonomous agents. The more an agent can change in the world, the stronger the governance it needs before you take the human out of the loop.

Post-launch decision: keep it or kill it?

Plan for sprawl. It's easy to turn around two months after enabling agent building and find 50 agents still running inference whose output nobody uses. For every live agent, ask:

• Is it tied to a critical need of the business?
• Is anyone using its output?

If the answer to either is no, shut it down. Make this a standing responsibility for your champions.

7. How much governance do I actually need, and who decides?

As little as possible, set up as early as possible. The bigger problem we see in enterprise AI right now is over-governance. You need enough to meet your responsibilities, but governance shouldn't get in the way of the impact AI can have on the business.

Decision 1: What does your environment actually require?

Govern for the environment you actually operate in.

It’s good practice for everyone to use enterprise-grade tools that don't train on your data. If you’re in a highly regulated industry, make sure you're fulfilling those specific regulatory obligations. But if you're not, don't borrow someone else's rules for the sake of it. We see a lot of standard SaaS companies governing as if they were in healthcare or financial services.

When you decide to pursue AI transformation, you're also deciding to increase your organization's surface area of risk. Every governance debate comes down to one question: is there greater risk in doing or in not doing?

Decision 2: Who makes up the AI governance council?

Your council should be made up of the people leading the transformation:

• Your head of AI, usually as chair
• Your CISO or whoever owns security
• Your general counsel
• Your head of HR, because AI affects people
• Your CTO or CIO, depending on their role in the transformation
• Your CEO, either as a regular participant or with direct line of sight into the council's conversations

Your CEO breaks ties. When CEOs sit on the sidelines, security, legal, HR, and AI leaders can debate indefinitely without a decision. Rotating senior champions through the council gives them visibility into how governance decisions affect AI rollouts. 

Meet at least monthly to review new capabilities, use cases and agents, and keep the structure flexible enough to keep up with how fast things change.

8. My people are anxious about AI. What do I tell them?

The truth, in writing. 

Publish an AI manifesto (use our template): a clear, actionable and honest set of principles explaining why AI is happening in your organization and what it means for everyone.

We frame it with what we call cut and create: e.g. “We're cutting old ways of working, old vendors and anything that no longer makes sense, and creating new opportunities and ways of working. None of us individually controls this transformation, but we believe everyone deserves a seat at the table, so give every employee a real pathway to participate.”

Don't promise that AI won't change anyone's work, because it isn't true. Lines like "AI won't take your job, someone using AI will" aren't any better, and you lose credibility when you use them.

On doomerism, pay attention to the public conversation but take a lot of it with a grain of salt, because much of it is sensationalized. That doesn't mean dismissing your employees' concerns. It's a disorienting time, and there's a lot nobody knows yet. Have frank conversations, then bring the focus back to what AI can actually change in your business.

9. Should we build this ourselves or buy it?

It depends on what "this" is.

• If it captures insight (reports, dashboards or agents that tell you about your business), lean toward building. These used to cost hundreds of thousands or even millions of dollars in consulting work, and your team can now often build them.
• If it's software, lean toward buying. We probably overestimated earlier this year how much agents would replace software. A 10,000-person company probably shouldn't vibe-code its own CRM to replace Salesforce, because building good software is still hard, and enterprise SaaS comes with the vendor's team maintaining it.

If you're still considering building software, answer three questions first:

1. What will the inference cost be to build and run it?
2. Who will maintain it over time?
3. How quickly will bugs get fixed?

10. What should I leave room for in 2027?

It's hard to commit a strategy for something that changes this fast, so expect to revisit your plan as you go. A few things we expect to see more of next year:

1. Persistent personal agents that stay with each employee over time. They could make every employee more productive, which would push up per-person inference and enablement costs. So treat that budget line as a floor.
2. Computer use. It only recently became reliable, and it opens up workflows that were hard to automate because the tools didn’t connect. Because these agents act directly in your systems, they go through the blast radius and reversibility check. (more on that here)
3. Agents that find or build other automations. They could speed up workflow redesign, and they could speed up sprawl just as fast. That makes the keep-or-kill process more important.

What to do if you only do one thing

Set your impact measurements and baselines before you finalize anything else.

Your budget, your workflow priorities, and your agent risk trade-offs all depend on knowing what you're trying to move and where it stands today. Without that, you won't be able to show the board where its money went.