# How do you measure AI adoption in a company? | Business Cyborgs

1. [Home](https://businesscyborgs.com/en/)
2. [Services](https://businesscyborgs.com/en/services.html)
3. How do you measure AI adoption in a company?

# How do you measure AI adoption in a company?

Having licenses isn't adoption, and neither is using them. Active users, prompts and hours saved tell you how much the tool gets used. Adoption is measured per workflow: which concrete piece of work changed in a way you can verify, which old step stopped being done, and who decided it. If nothing stopped being done, you're still paying twice.

[Why do AI projects fail in companies?](https://businesscyborgs.com/en/why-ai-projects-fail.html)

## Usage isn't adoption

It's easy to measure what the tool logs: how many logged in, how many prompts, how many tokens. That tells you whether people open it. It doesn't tell you whether the company works differently. Someone can write faster while the proposal still waits for the same approval and the same review.

What matters is simple actions that solve real problems. You measure that in the work, not in the tool.

## Three levels: access, capability and substitution

- Access: the tool is available. The license buys it. Measured with users and usage.
- Capability: the team produces an acceptable result, again and again. It's built. Measured with real cases that pass the tests and substantive corrections going down.
- Substitution: the new cycle passed a threshold and someone retired an old step. It's decided. Measured by what stopped being done.

Many companies celebrate access, work hard on capability and never reach substitution.

## What to measure in each workflow

Before you start, write down six things:

1. Piece of work. Which concrete step are you changing?
2. Starting point. How long it takes today, what quality it delivers, what controls it has, what exceptions show up.
3. What's missing. Data, permissions, method, controls and skills.
4. Overlap. How long both cycles will run and who carries the double work.
5. Threshold. What the new cycle has to show to retire the old one without losing what it protected.
6. Who decides and when. Who can retire, fix or roll back, and on what date.

The threshold comes from the starting point and the risk of each workflow. There's no universal one.

## An example

An analyst builds a weekly sales report from several sources. The new cycle drafts it from authorized sources. For four weeks both run, and time, rework and failures get logged. Before starting, the team agrees on a threshold, for example: four deliveries in a row with traceable sources, zero exposure of restricted data, at most two substantive corrections, and less time than before, without losing controls. At the end of week four, the owner decides: retire the manual consolidation, fix it with one extension, or roll back. It's an illustrative example, not a client.

## Watch out for measuring what doesn't matter

- Counting automations rewards volume. Sometimes the best outcome is a report that disappears or goes to someone else.
- Hours saved are only part of the math. The question is whether the result the process protects got better: on-time delivery, fewer errors, faster answers.
- If whoever frees up capacity gains nothing, they'll defend the old step.

## When Business Cyborgs helps and when it doesn't

We work one workflow at a time with the people who run it, with its starting point, test set and threshold, until the old step can be switched off. If there's already a workflow that hurts, that's an [Agent Sprint](https://businesscyborgs.com/en/services/agent-sprints.html): one problem, one workflow, 2 to 4 weeks, from USD 5,000. Several workflows with shared data and a leadership mandate is a [Transformation](https://businesscyborgs.com/en/services/transformations.html). If there are licenses and curiosity but no clear case, we start with a [Talk](https://businesscyborgs.com/en/services/talks.html).

It's not for you if what you want is a tool usage dashboard or a prompt course for everyone with no workflow behind it.

[See the Agent Sprint](https://businesscyborgs.com/en/services/agent-sprints.html) · [Talk to a Business Cyborg](https://calendar.app.google/WwiEztrut3UkDZSx7)

## Published cases

In the [consumer goods case](https://businesscyborgs.com/en/cases/consumer-goods-marketing-agency.html), the measure wasn't how many people used the tool: the traditional marketing agency left entirely and was replaced by an agent operation, with 8 modules, 3 roles and an automated weekly report. In the [Technical Knowledge Agent](https://businesscyborgs.com/en/cases/technical-knowledge-agent.html), what changed was that answers became more consistent and source-based, and the organization could see which documentation was missing.

## FAQ

### Which metrics work for measuring AI adoption?

Per workflow: time, quality, substantive corrections and exceptions against the starting point, and which old step was retired. Tool usage is a signal, not the goal.

### We bought licenses and almost nobody uses them. Is that an adoption problem?

Usually nobody said what they were for. More in [We bought AI licenses and nobody uses them](https://businesscyborgs.com/en/we-bought-ai-licenses-nobody-uses-them.html).

### How often should we review?

Every cycle of the workflow, with a fixed date to decide. Without a date, the overlap becomes permanent.

### Who should own adoption?

The owner of each workflow inside the team that runs it, not the technology team or the vendor.

Further reading (in Spanish): [BC 56](https://businesscyborgs.substack.com/p/si-nada-deja-de-hacerse-no-migraste), [BC 54](https://businesscyborgs.substack.com/p/bc-54-pedimos-transformacion-medimos) and [BC 64](https://businesscyborgs.substack.com/p/bc-64-tiempos-modernos-automatizar).
