# Measuring AI ROI: a baseline-first method for 2026 budgets

- **Published:** 2025-12-30
- **Updated:** 2026-10-12
- **Category:** Economics & Pricing
- **Author:** Sinthetix
- **URL:** https://sinthetix.com/blog/measuring-ai-roi

> Most organizations use AI but few can show profit impact. A simple way to define a baseline, count real cost per task and decide what to scale.

Budget season is when AI pilots meet finance. The uncomfortable pattern in the data is that use is widespread and measured return is not. In McKinsey's November survey, as reported by IT Brief, 88% of organizations used AI in at least one function, yet only 39% reported any effect on enterprise-wide earnings, and most of those attributed less than 5% [1]. This post sets out a method for making your own number defensible.

**Key takeaways**

- Measure a workflow before you automate it, or you will have no baseline to compare against.
- Count cost per completed task, not cost per token; efficiency claims between models differ widely [2][3].
- Include review time, errors and rework, not only the subscription or API bill.
- Decide in advance what result would make you stop, scale or change course.

## What the survey says separates the few

Per the same IT Brief coverage, among respondents, about 6% qualified as "AI high performers" attributing more than 5% of EBIT to AI. Compared with others, senior executives at these firms were three times more likely to own AI initiatives, and more than a third of them devoted over 20% of digital budgets to AI versus under 10% for others [1]. Over half of all respondents had experienced at least one negative consequence from AI in the past year, with inaccuracy the most common [1]. The survey shows association, not cause, but the pattern is consistent with ownership and focus mattering more than tool choice.

## Step 1: Pick a workflow and record the baseline

Choose one process with a countable output: invoices processed, support tickets resolved, quotes drafted, documents reviewed. Before changing anything, record for a few weeks:

- Volume per week
- Average handling time
- Error or rework rate
- People involved and their loaded hourly cost

If you cannot get these numbers, that is the first finding. A simple template to fill in with your own measurements:

| Metric | Baseline (before) | After (pilot) | Threshold to scale |
| --- | --- | --- | --- |
| Volume per week |  |  |  |
| Average handling time |  |  |  |
| Error or rework rate |  |  |  |
| Review time per task |  |  |  |
| Cost per completed task |  |  |  |

The cells are blank on purpose: they are yours to measure, not figures from a study.

## Step 2: Count the real cost per task

The model bill is only one line. A fuller view:

| Cost | What to include |
| --- | --- |
| Model usage | Input and output tokens per task, times current prices |
| Platform | Hosting, retrieval, monitoring, integrations |
| Human review | Time spent checking or correcting output |
| Rework | Cost of errors that reach customers |
| Build and upkeep | Initial development and ongoing maintenance |

Token efficiency alters the model line materially. Anthropic says Opus 4.5 at medium effort matched Sonnet 4.5's best SWE-bench Verified score with 76% fewer output tokens [2], and Google says Gemini 3 Flash uses about 30% fewer tokens on average than Gemini 2.5 Pro on typical traffic [3]. These are vendor claims on vendor tests, but they make the point: two models with similar list prices can have different cost per task. Techniques such as prompt caching can cut input cost by up to 90% on repeated long prompts, per Anthropic [4]. Measure with your own prompts.

## Step 3: Measure quality, not just speed

A faster process that produces more errors can lose money. Take a sample of 50 to 100 outputs, have a qualified person grade them against the baseline process, and track the pass rate over time. Record failures by type so you know whether the fix is better instructions, better retrieval or a different model.

## Step 4: Set decision thresholds before the pilot

Write down, in advance:

1. The cost per task at which the pilot pays back.
2. The minimum quality pass rate.
3. The review time you expect people to spend.
4. What happens if results land between thresholds: extend, narrow the scope or stop.

Deciding thresholds after seeing results invites rationalization in either direction.

## Step 5: Watch for what is not in the model

Savings often appear as time freed, not cost removed. Freed time counts as return only if it is redeployed to something measurable, such as faster quote turnaround or more jobs scheduled. Otherwise it is capacity, and should be reported as capacity.

## The takeaway

Return on AI is a property of a specific workflow, measured against a recorded baseline. Start 2026 by choosing two or three workflows, capturing today's cost per task and quality, and reviewing results quarterly against thresholds you set in advance.

## Sources

1. [McKinsey report shows AI interest but slow scaling (coverage of McKinsey's State of AI in 2025)](https://itbrief.co.uk/story/mckinsey-report-shows-ai-interest-but-slow-scaling), IT Brief, 2025-11-11
2. [Introducing Claude Opus 4.5](https://www.anthropic.com/news/claude-opus-4-5), Anthropic, 2025-11-24
3. [Gemini 3 Flash: frontier intelligence built for speed](https://blog.google/products/gemini/gemini-3-flash/), Google, 2025-12-17
4. [Prompt caching with Claude](https://claude.com/blog/prompt-caching), Anthropic, 2025-08-14

## How Sinthetix helps

An engineering studio for AI agents, automation, and internal tooling. We build the systems that make operators scale.

[Our services](https://sinthetix.com/services) | [Talk to us](https://sinthetix.com/contact)

Sinthetix is an AI engineering studio. We ship agents, automation, and internal tools for operators who need results, not reports. [About Sinthetix](https://sinthetix.com/about)