Pick one workflow, count volume, time each step, find the leak, name the metric, set exit criteria. The baseline method behind honest AI ROI.
Every AI ROI claim you have ever read had a baseline behind it, or it was made up. Before we build anything for a client, we spend a short, unglamorous stretch measuring how the workflow runs today. This is the method, step by step, with a hypothetical example so you can run it yourself before you talk to any vendor.
WRITER's 2026 research found 79% of organizations face challenges adopting AI, and Anaconda and Forrester research puts the share of agent pilots that never reach production at 88%. A large part of that gap is that nobody defined what "working" meant before the pilot started, so nobody could prove it worked.
Our rule is that hours are the honest currency. Not "efficiency," not "productivity," not a percentage from a slide. Hours of human work the workflow consumes today, and hours the system absorbs after.
Not a department. One workflow with a clear trigger and a clear end state. "Handle inbound quote requests from the website form until a quote is sent" is a workflow. "Improve sales" is not.
Good candidates happen often, follow a recognizable pattern, and someone on your team is visibly tired of them. Pick the one whose owner can describe the steps without hesitating.
How many times did this workflow run last month? Pull it from the system that triggers it (form submissions, tickets, invoices received), not from memory. If you cannot count it, you have found your first problem.
Distribution matters too: twenty runs a day on weekdays with spikes on Monday is a different automation problem from six hundred runs on the first of the month.
Break the workflow into its steps and time each one across several real runs. You are looking for two numbers per step: the active handling time and the wait time before the next step starts.
Wait time is where most of the money hides. A quote that takes 20 minutes of active work but sits in an inbox for two days is a two-day workflow, and the customer experiences it that way.
Multiply volume by per-step time. Which step consumes the most hours per month? Which has the longest wait? Those are usually different steps, pointing to different solutions.
Pick one primary metric that the build must move. Hours absorbed per month is the default. For customer-facing workflows, time-to-first-response is often more valuable, because revenue follows response time: leads answered in minutes tend to convert better than leads answered in days. Pick one secondary guardrail metric that must not get worse, such as error rate or customer complaints.
Write the metric down with its current value and the date. This becomes the "before" line on every future report.
Before building, decide what ends the pilot. Two explicit outcomes:
A pilot without a stop condition is a subscription.
Suppose a 12-person marketing agency receives about 90 inbound project inquiries a month through its website form. This is an illustrative scenario with invented numbers, not a client result.
Timing five real runs, the team finds:
Active handling is about 45 minutes per inquiry, or roughly 67 hours a month. Time to first response averages about a day and a half. The leak is the reply draft: high hours and the longest wait.
The metric: hours absorbed per month, with time-to-first-response as the secondary. The guardrail: no increase in replies the account lead has to rewrite. Exit criteria: absorb at least 40 hours a month within six weeks with first response under two hours, or stop.
The build that fits: an agent that reads the inquiry, retrieves the three closest past projects, drafts a reply, and queues it for a human to approve, with automatic CRM logging. The human stays in the loop exactly where the judgment is.
Once the system runs, measure what each automated task costs: model usage, retrieval calls, infrastructure, and any human review time. Cost per task lets you compare the system against the baseline hours honestly, and tells you when a cheaper model tier is good enough.
Anthropic's Claude Opus 5, released July 24, 2026, is described by Anthropic as coming close to the frontier intelligence of Claude Fable 5 at half the price. If your evaluation shows the cheaper tier passes on your real inputs, the cost-per-task math changes immediately.
The report you show leadership has a before column and an after column, from the same counting method, with dates. The ROI figure is a division of two numbers you can show, not an estimate from a vendor deck. And the pilot can fail visibly, which is the only way successes are believable.
Baselining is the first deliverable in most of our engagements, whether the eventual build is AI automation, an AI agent, or a broader AI consulting roadmap. We prove value on one workflow, then scale with evidence, and every system we ship includes cost-per-task instrumentation and evaluation before launch. Human-in-the-loop is a design choice, not an afterthought.
For teams that want that measurement built in without standing up their own infrastructure, My Cloud Company runs managed agents with approval gates, hard budget caps, an audit trail, and a weekly report. If you already have a workflow in mind, send it to us with your best guess at the monthly volume, and we will help you run the baseline.
Answers
For a single workflow, usually one to two weeks of light effort alongside normal work. Most of that is timing five or more real runs and pulling volume from the source system. The write-up itself is a page: steps, hours, the leak, the metric, and the exit criteria.
Treat that as the first thing to fix. If runs are not captured anywhere, add a lightweight log or a form so the trigger is recorded. Without a volume count, any ROI figure you produce later will be an estimate, and estimates are what vendor decks are made of.
A license fee tells you nothing about whether the work got done. Cost per task combines model usage, infrastructure, and human review time for each run, which lets you compare directly against the hours in your baseline. It also shows when a cheaper model tier is sufficient.
Put it into practice
The XISLABS services closest to what this article covers.
Keep reading
OpenAI released GPT-6 Astra on September 3, 2026, the fourth major release this year. A practical evaluation harness so you can decide in days, not quarters.
Stop picking one winner. A three-tier model strategy (frontier, workhorse, self-hosted) with routing by task, data residency, and eval-driven switching.
Anthropic shipped Fable 5.1 and Mythos 5.1 on September 1 and made Fable 5 legacy after 12 weeks. How to tier models and plan for lifecycle churn.
We build the AI agents, automation, and software behind ideas like these — scoped to a metric, shipped in weeks, operated after launch.