Where AI Creates Measurable ROI in Your Business

Most companies measuring AI ROI are measuring the wrong thing. Seats provisioned, prompts sent, pilots launched — those are activity numbers, and activity is not return. The return shows up in exactly one place: a workflow that got rebuilt around the model, with a metric that existed before the rebuild.

Where does AI actually create measurable ROI?

In high-volume, measurable workflows that get redesigned around the model — not in tools bolted onto process that stays the same. McKinsey’s State of AI research is blunt about this: only about 6% of organizations qualify as AI high performers, meaning they attribute more than 5% of EBIT to AI. What separates that 6% is not better technology. It is workflow redesign.

That finding should reorder your roadmap. The question is not “which AI tool should we buy?” It is “which workflow, if rebuilt, would move a number the CFO already tracks?” Tickets resolved per agent per day. Days from underwriting file to decision. Cycle time from ticket to production. If the workflow has volume, rules, and a metric, AI can move it. If it has none of those, you are funding a demo.

Here is the position, and plenty of vendors will disagree with it: “adoption first, ROI later” is backwards. Usage without a baseline metric is not early ROI — it is cost. Bolt AI onto how you already work and you buy a quarter. Rebuild the operating layer underneath and you buy the decade.

Why do most AI investments show no return?

Because most AI investments are pilots, and pilots almost never reach the P&L. MIT’s GenAI Divide study found that roughly 95% of enterprise generative AI pilots deliver no measurable P&L impact. The root cause the researchers name is not model quality. It is a learning gap: generic tools never adapt to the specific workflow, so the workflow never changes, so the number never moves.

McKinsey’s survey data frames the same gap from the other side: 88% of organizations report adopting AI, but only 39% report EBIT impact at the enterprise level. That is not a technology gap — nearly nine in ten companies already have the technology. It is an operating-model gap. Adoption got funded. Redesign did not.

The same MIT data carries a second finding buyers skip past: externally built, specialized tools succeeded about 67% of the time, while internal builds succeeded roughly a third as often. The failure is not ambition. It is trying to close the workflow gap with generic tooling and part-time attention. We wrote the longer version of this argument in how to move beyond AI pilots into everyday operating processes. The short version: a pilot with no owner, no baseline, and no workflow change is not an experiment. It is AI theater.

Which business functions show AI ROI first?

Customer support, software development, and content-heavy marketing work — and the measured gaps between functions are wide. Per Stanford HAI’s 2026 AI Index, measured productivity gains land around 14-15% in customer support, 26% in software development, and 73% in marketing output. The unevenness is the lesson: gains track how rule-heavy, high-volume, and checkable the work is, not how enthusiastic the team is.

Anthropic’s own usage data points the same direction. The Anthropic Economic Index Cadences report found 93% of Claude conversations produce a concrete artifact — code, a document, a plan. Artifacts are why the ROI is measurable at all. An artifact can be counted, diffed, evaluated, and QA’d. A brainstorm cannot.

There is a practitioner version of this map too. JJ Englert on our team handed his own job to Claude Cowork on camera, and the honest finding was not “AI does everything.” The automatable parts were exactly the high-volume, well-defined ones; the judgment calls stayed human. That is the ROI map in miniature: automate the countable middle of a workflow, keep humans on the edges. If you are deciding which workflow goes first, we published the full sequencing argument in which workflows to automate first — and which should stay human.

How do you measure AI ROI before spending anything?

Baseline the workflow before the model touches it. One metric per workflow, defined tightly enough that a skeptical CFO signs off: cost per resolved ticket, cycle time from PR open to merge, hours per campaign shipped. Measure the current process. Then — and only then — deploy, and measure the same metric on the same work.

This is where most measurement dies. Teams deploy first and reconstruct the baseline from memory, and a reconstructed baseline is a negotiation, not a measurement. Numbers earn trust by carrying their definitions; a percentage with no denominator is a press release.

Three tests before any AI initiative gets funded:

• Volume. Does the workflow run often enough that a 20% gain compounds into something the P&L can see?
• Verifiability. Can the output be checked — by eval, review queue, or diff — without heroics?
• A named owner. Who loses sleep if the metric does not move? No name, no funding.

Pass all three and the ROI conversation gets short. Pass none and no model will save it. We break down one tactical build like this every Tuesday in the ultrathink newsletter — 40,000+ executives and builders read it weekly.

How would Tenex find the measurable ROI in your business?

We embed, trace the workflows, and let the numbers pick the bet. Our AI transformation work starts inside your team, not in a conference room: we map the workflows carrying real volume, record the baselines, and rank the rebuilds by expected movement on a number your CFO already tracks. Glamorous work and garbage work alike — sometimes the blocker is data plumbing, not intelligence, and if the trash needs taking out before the system can work, we take out the trash.

Then we ship. First production system by week four, median — a system running on real work with its baseline already recorded, not a deck. Then we train the operators who own the workflow, and we track the number most vendors will not show you: 90%+ operator adoption, defined as weekly active at handoff, a figure we tag illustrative until our launch audit because unaudited numbers are theater — including ours.

Then we hand off. Repos, evals, agents, dashboards — transferred, our access revoked. Dependency is a design flaw. If the ROI only exists while the consultants are in the building, it was never ROI.

The people doing this are the 0.16%: 50 builders selected from 32,100 applicants, every one through a paid build trial. No leetcode theater. You build. We watch.

Where does chasing AI ROI go wrong?

It goes wrong when you automate a broken workflow, skip the baseline, or expect the return in the wrong currency. The failure modes we see most:

• Automating the mess. AI applied to a broken process gives you a faster broken process. McKinsey’s redesign finding cuts both ways — if the workflow is wrong, redesign comes first, model second.
• No baseline. If the “before” number was never recorded, the “after” number proves nothing, and finance will treat it accordingly.
• Counting the wrong currency. The first returns usually arrive as cycle time and throughput, not payroll. Model the ROI as capacity — more shipped per team — before you model it as cost-out.
• Tool sprawl as strategy. Ten point tools with ten logins is not a portfolio. It is ten unmeasured bets and one very confused workflow.

And the honest disqualification: do not fund an AI ROI push if your workflow data does not exist yet. If nobody can say today what a ticket costs to resolve or how long a file sits in queue, the first project is not AI. It is measurement. We would be wasting your money on anything else — and we will say so.

Here is the challenge. Pick the workflow in your business with the highest volume and the clearest metric, and write down today’s baseline. If you cannot, that is the finding. If you can, bring it to us and we will tell you straight whether there is a bet worth making — or tell you there is not.