80% of Companies See No ROI on AI (Here's Why)

Everyone talks about AI. But in reality, actual adoption lags far behind the hype.

I often use this image about the world's population as an example: 84% of humans have never used AI, and only 0.3% pay for a $20/month subscription. That's little given the media noise around the topic.

And on the business side, it's not much better. McKinsey shared a fairly striking figure: more than 80% of companies still see no impact on their bottom line, despite the millions invested in the topic.

With Tandem, we've supported 40+ companies (60+ AI projects) across very varied contexts (PayFit, InfraVia, Doctolib, Nomination, KparK, …). I wanted to take the time to share what we observe in the field: the logic that works, and the steps we follow to unlock AI's full potential.

Here are the 5 main steps of this section:

1. The diagnosis
2. Deploying an AI copilot
3. Task automation
4. The success conditions to put in place
5. A new role to define: the AI builder

Spoiler: the most common mistake is skipping steps. People want “an AI agent” right away when the first two steps aren't in place.

1️⃣ Understand before deploying

The classic reflex is to want to plug in the latest trendy AI tool without deeply understanding how the company works, and you miss out on most of the value…

Here are the elements we systematically look at before proposing anything:

• The internal organization: who does what, the size of the teams, the time teams spend on the different tasks…
• The existing processes: those that are documented, those that aren't (and often, many aren't)
• The technical architecture: which tools are used daily, which CRM, which stack, which data is accessible
• The OKRs and business priorities: what the company is trying to improve this year

Prioritize by impact (not just financial)

Once we have the mapping of tasks, we prioritize. And here too, a classic mistake: reasoning only in terms of direct financial ROI.

Impact can take several forms:

• Time saved on repetitive tasks
• Reduced error rate (quality of work)
• Employee experience (making thankless tasks disappear)
• Customer experience (response time, personalization)

In practice, the departments that come up most often in our engagements are support and sales for operational processes. And that's consistent with what we observe on a larger scale. a16z (a VC fund) shared a recent report: code, support and information retrieval are the 3 dominant uses in the companies that really deploy AI in the US.

Why these departments? Because they're often the functions with the most employees, structured processes, repetitive and time-consuming tasks, and above all clear KPIs to measure impact (tickets resolved, qualified leads, conversion rate).

I'll also take the chance to share this edition where I went over the top 5 AI use cases in business:

Top AI use cases in business: concrete feedback and impact

Quantify the ROI upfront

There are so many possible AI use cases today, across every department, that you can quickly spread yourself thin and launch 10 initiatives in parallel without knowing which ones will really pay off.

To avoid this, we structure the impact estimate in 3 stages:

1. Analysis of the team structure: number of FTEs, fully loaded cost per employee, cost per ticket or per operation
2. Understanding of flows and friction: volumes, average durations, % of missed calls, friction points in the journey
3. Estimate of annual gains vs setup cost + usage cost (AI models have a per-call cost that must be factored in from the start)

A recap with a concrete example:

The figures have been modified, but in the example, it was a voice agent project for customer support. The payback was in under 2 months. But the raw figure isn't everything. On this project, the voice agent also brought a better customer experience (fewer missed calls, 24/7 availability), much finer monitoring (automatic categorization of requests, detection of dissatisfaction) and actions that could be triggered downstream (alerting, smart routing). These elements matter as much as the direct ROI.

Factor in technical feasibility

A final, often underestimated angle: what can AI really do today on your specific case?

Because automating doesn't mean 100%. AI covers 80% of standard cases very well, but not every edge case. And that completely changes how you design the solution (with human in the loop notably).

A few concrete examples where we currently recommend waiting:

• Voice agents for cold sales outreach: not yet mature, you quickly fall into the uncanny valley and it can damage the brand
• Very long workflows with many interdependencies: autonomous agents still hit a ceiling on this type of task

Saying “it's not mature today” doesn't mean AI isn't relevant for the company. It means that specific use case is poorly chosen. There are often 5 or 10 others right next to it that are perfectly mature and high-impact.

Quick wins vs long-term projects

Once we've cross-referenced impact and feasibility, we produce a simple matrix per department, like this:

The goal: a global, documented view to make simple decisions. We start with 2-3 quick wins to build the team's skills and generate fast ROI, and we take on 1-2 long-term projects in parallel on the big strategic stakes.

Two examples of classic traps to avoid:

• Too much ambition: tackling the most complex project head-on when you have neither the internal maturity nor the simple use cases to learn from. If you're looking for 100%, AI isn't what you need.
• The “build” instead of “buy” reflex: some companies set off to develop a custom solution in-house (an internal ChatGPT, an image-generation solution..) while, for example, Midjourney, Krea, Leonardo release new features every week. The devs spend their time catching up with the state of the art instead of creating value. For standard building blocks (generation, transcription, web search), very often favor off-the-shelf solutions.

2️⃣ The AI copilot: where every company should start

Whatever your sector, size or maturity, this step concerns you. An internal ChatGPT, whatever the solution (Claude, Copilot, Dust, Mistral, …), is the starting point that unlocks the most value for the least effort.

The stakes are twofold: enabling smart search across your internal data (SharePoint, Drive, CRM, business tools…) and creating specialized assistants connected to the right sources to boost teams' productivity.

In terms of cost, we're talking about $30 to $45/user/month → 1 hour saved per month already pays for the license.

The real difficulty comes when you want to push further: creating genuinely relevant assistants. That's where we talk about skills (task-specialized instructions), connectors (to plug the assistant into your tools), MCP (to go further in the actions). These are technical topics, and making teams autonomous on them isn't easy.

Two points systematically underestimated at this step:

• Change management: installing a tool isn't enough. You have to configure it, train, support, show concrete use cases... Without that, it doesn't work.
• Measuring adoption: how many people really use the tool, on which tasks, with what impact. Without metrics, it's impossible to steer the next steps.

And you also have to factor in that these platforms evolve fast (every 3-6 months), with new features to test, adopt, train on. It's not a “one-shot” project, it's a capability to develop continuously. Do the teams have the bandwidth to do it?

3️⃣ Automation: the real leverage

Once the copilot is well established, we move to the next step: automation. The human is no longer permanently in the loop, AI executes end to end and we validate occasionally (or not at all).

This is where the leverage becomes massive.

Three main families (mostly 2) here, that need to be clearly distinguished:

• AI workflows: a scenario defined in advance, with one or more AI steps in the middle. Hybrid logic (rules + probabilistic). Suited to 90% of business automation cases today.
• AI agents: an autonomous system that decides for itself which actions to launch, which tools to use, how many iterations. Purely probabilistic logic. Much more powerful, but much less deterministic, and therefore riskier to deploy in production.

At Tandem, we use n8n a lot for this part. It's open-source, self-hostable, and lets you build both simple workflows and complex agents in the same tool.

To dig into the possible levels of automation (from the assistant to the autonomous agent) and know which to choose for your case, I refer you to this edition which goes into detail.

The key takeaway: don't jump straight to the autonomous agent because it's the hype of the moment. In 80% of cases, a well-designed AI workflow will give you a better ROI, faster, with far less risk.

4️⃣ The success conditions

Beyond choosing the right step and the right use case, there are cross-cutting criteria we systematically find in the companies that succeed with their AI deployment.

Nothing revolutionary: test, measure, iterate, repeat.

Here are the 5 criteria we observe in those who really extract value:

• Document the context : the more you structure your processes, your business rules, your knowledge base, the more relevant AI becomes. It's a big part of the work on an AI project, and it's rarely sexy. But it's what makes the difference
• Experiment continuously : quickly launch POCs over short sprints. The cost of an AI POC today is very low, the real cost is inaction.
• Measure what matters : time saved, adoption rate, quality of the result, user satisfaction, financial ROI. Without metrics, it's impossible to know what works and what deserves to be scaled.
• An accessible technical stack : AI needs to interact with your tools to be really useful. A custom ERP, an in-house CRM, business applications with no API … it blocks everything. It's often a whole project in itself..
• Continuous training : the tools evolve every 3 months. You have to accept that skill-building is never finished, and set aside recurring time for teams to test, explore, share their findings.

What these 5 criteria have in common: it's not a project with a start and an end, it's an organizational capability to develop. The companies that get stuck are often the ones that treat AI like a classic IT project, with a fixed spec and a single go-live.

5️⃣ A new role to define: the AI builder

It's a fact: since ChatGPT came out 3+ years ago, AI has already changed the way most companies work.

We write faster. We analyze more data. We automate tasks that used to take hours.

But the real transformation isn't only in the tools. It's also organizational, in the new roles that are emerging.

At Palantir, they invented the role of Forward Deployed Engineer to describe these hybrid profiles capable of connecting technology and field needs.

Today, hundreds of companies recruit this type of profile. And we see the same thing happening with AI.

In this video, I explain why the AI Builder role is becoming central in companies (and why companies are so eager to find this profile)

The role: turning AI into concrete impact. It acts on two axes I presented above (copilot and autopilot)

The point not to underestimate is that AI doesn't work on its own. There's all the business logic:
➡️ reproduce the operational logic
➡️ connect the tool ecosystem
➡️ manage security, governance and internal constraints

This is precisely where AI Builders bring value.

And that's how AI will really transform companies. Not only with new tools. But with new roles.

With every technological shift, roles evolve to reflect the new value created. And AI is bringing about exactly the same thing.

The big transformations in companies don't come first from software.
They come from people who use these tools to solve problems differently.

And these people often end up creating… their own role.

So who will be the AI Builder in your company? If you're still looking, maybe we can talk.