AI got its start with consumers and small businesses, but the future of AI needs to get paid for by large enterprise. Investors and Big Tech are pouring tens of billions into AI with the expectation that enterprise will spend billions every year on AI products and services.

But these enterprise buyers also need to see their own ROI, and quickly. If you’re a leader making AI investments, it won’t be long before your CFO starts asking about the return. In fact, it’s already started: Microsoft CEO Satya Nadella is already warning of slower AI sales due to tighter enterprise budgets and the need for better change management.

And when the current IBM Watson commercial (running over the last two weeks during the US Open) starts with “Think scaling your AI pilots is hard?” you know there are problems in enterprise AI paradise.

Optimize, Accelerate or Transform

There are 3 ways companies will get AI returns – based on our AI strategy framework –  Optimize, Accelerate, Transform (OAT):

• Optimizing internally looks like making your day-to-day workflows more efficient with AI
• Accelerate looks like adding AI into your existing products or services to drive revenue and retention
• Transform looks like creating all new products with AI or entering new markets – or withstanding existential business risk from AI

Our POV is that every organization should be optimizing internally with AI at this point, so we’ll focus on how to measure the ROI of these efforts in this post (let us know if you want to see breakdowns on the others). These initiatives are usually the fastest to launch and get value from – but often the hardest to measure.

Here’s how you should be thinking about optimizing internally.

Optimize

When you’re optimizing with AI, you’re likely augmenting your workforce with AI and improving the efficiency of your internal workflows.

In some cases, you will be replacing humans with trained AI – and in this scenario, the business case will be clear: Direct labor savings vs the cost of AI implementation and on-going operation.

The business case for replacing people is the easiest to prove (if the AI can do the job) – but currently it doesn’t apply to very many jobs outside of roles like translation, content creation, and customer service/call centers. Today AI mostly performs tasks not jobs – so immediate staff reductions are less obvious.

Most of you will augment your knowledge workforce in one of three ways:

1. Deploying an “off-the-shelf” LLM

Licensing and deploying a chatbot like ChatGPT for Enterprise, Microsoft Copilot, or Google Gemini is your fastest way to AI ROI – just how much ROI?

Obviously, you will need to do more than buy and deploy these tools. AI is not like any previous software deployment, so the change management and workforce training challenges are real if you want meaningful adoption – which you will need to get the ROI and payback.

Your ROI should come from productivity gains, improved quality of output, and over time, lower headcount.

At Section, we have seen a 10-20% productivity gain after six months of sustained usage of ChatGPT and Claude – but we are a small company and we have a LOT of tasks well-suited to LLMs (we make a lot of content).

When I talk to enterprise buyers or attend conferences with AI buyers, a 10% productivity gain seems to be the consensus on what’s possible after deployment of an off-the-shelf LLM at work. But these productivity gains are likely to be anecdotal and self-reported, which is what makes it harder to measure.

You could also look at quality of output, but that’s even harder to prove unless you already have metrics in place you can compare to.

Measuring less headcount is your long term play. This doesn’t mean firing people (that would be replace, not augment) but rather having to hire less people into specific teams, even when their workload is increasing.

‍2. Workflow automation using AI agents with a human in the loop

This is when you create a prioritized list (10 or so) of high value workflows in your business that AI can augment or take over. These are often language intensive (because that’s how large language models work!) – like content creation, coding, legal review, and hundreds more – and should be highly repetitive and laborious for your employees – these are the processes you want to offboard to AI.

In this case, you’re getting or building a specific AI tool that can automate those workflows for a large enough team to justify the investment – and the human stays in the loop to manage quality control. For coding it might be Cursor, for writing it could be Jasper.

With this approach, your expectation should be more productivity gains over time (e.g. 20%+) and it’s going to be more obvious, faster, if it’s worth continued investment because these humans can report on the gains (or not) that they are getting.

3. Using an AI agent with no human in the loop

This option shows a lot of promise – there are a lot of proof of concepts floating around – but it’s probably a better bet in 2025.

This form of augmentation will look like deploying an AI system that operates mostly independently of human intervention or oversight. Essentially, the AI will be fully trusted to complete very specific tasks, and manage well-defined processes autonomously.

There is a lot more to discuss here, but let’s save it for another time – it’s no longer science fiction, but there is still a lot of hype and few good case studies of this being used inside of everyday knowledge work.

How to measure AI optimization

Here’s how you can put a plan in place to measure your AI ROI, with the two optimization modes possible today.

‍With off-the-shelf LLMs you should be aiming for 75% adoption if you're in a company or team of ~1,000 people within 12-18 months. If you have a smaller team like ours (about 30) that number should be 100% within 12 months. For larger organizations (1-5K), I would assume 25% adoption in year 1, and 50% by the end of year 2 would be VERY successful – with a few outliers, like Moderna, with 80% ChatGPT adoption after 12 months.

Assume a 10% productivity gain for the team and the math is simple:

(Number of users x cost of LLM seat) + cost of your change management and training plans = Total expense. Then calculate 10% of your total employee expenses and compare.

Now, we know it’s not simple. Time-savings does not always translate to productivity gains, and productivity gains does not always translate to business outcomes. Just think about email – it was initially a much more time and cost efficient way to communicate – but did that convert into real productivity gains? Or are we now spending a LOT more time handling a dramatically increased volume of emails?

To get a more accurate read on productivity gains in your team, have them log the impact of AI on their role. We created an AI ROI Calculator that generates this kind of report and converts time savings into cost savings.

Because a lot of these productivity gains will be self-reported, you should also plan to run regular surveys to understand how AI is being used and how effective it is. I recommend running one every 6 months and using the results to inform how you expand or change your deployment strategy.

This is where you’ll get a sense of the anecdotal ROI – such as job satisfaction – and how successful your training plan has been. We built an AI Proficiency Diagnostic that you can use, if you don’t want to create one internally. Just reach out.

‍If you use AI agents with a human in the loop, ROI will be much easier to calculate. Not only do you know the workflow, how much it costs to run now, and the original output when done by humans, it’s also a yes or no situation – it’s either better or faster with AI or it’s not.

Create a benchmark for what outputs looked like before, when a human owned 100% of the workflow – how long it took them, how much their time and any other resources cost, and what the output was.

Then put together a pilot for the AI-enabled workflow and measure against that benchmark after 30, 60, and 90 days. If it saves time, costs less, and generates the same or better outputs, you’ve got the start of ROI. If it doesn’t pick another workflow to pilot and try again.

Your AI ROI honeymoon won’t last forever

It's obvious what's happening in the enterprise right now: Other SaaS contracts are being downsized or canceled to make way for AI spend in organizations.

This makes sense – we all bought TOO much productivity and workflow automation software over the last 10 years, so we can afford to lose a few systems and no one will notice. At Section, we spend about $700K annually on all our SaaS platforms, and we are determined to reduce that by at least 10% in 2025 – so we will have room to add in more AI spend.

If you’re looking for that 10% productivity gain from an off-the-shelf LLM, deploy it to a significant enough number of employees to have a valid test – probably 10-20% of the org (or 100% of a small team) – and see how it affects productivity. If the results are good, and the risks are manageable, scale your deployment.

If you’re considering AI agents with a human in the loop, form an AI tiger team to audit team workflows. Build a long list of possible AI-assisted use cases, narrow that list to 10 use cases (based on expected ROI - e.g. how many people in the org would benefit from this new AI workflow), then pick 3 to focus on, and build 1 that has high value and is fast to measure.

None of this is “easy”, and we should not expect it to be – notwithstanding the sales pitches from OpenAI and other AI vendors. They need us to deploy (and pay) for their own investments to pay off, but this is NOT typical workflow automation software. Software typically automates a workflow that we already know and handle – e.g. updating customer records, so deploying a new CRM is the easy choice.

Off-the-shelf AI, like ChatGPT or Copilot, is not software. It’s a general purpose technology that can assist with hundreds of potential tasks in the knowledge workforce. It will change how many knowledge workers do parts of their job everyday – with potential productivity gains much greater than 10% per year per employee. But it takes conviction, a plan, and discipline to measure the ROI – only then can you keep your CFO from looking over your shoulder.