When AI Saves Time And When It Creates More Work

AI saves time when checking its output is much faster than doing the work yourself and a missed mistake is cheap. Break either condition and it usually creates more work than it removes. That one rule explains most of the research. Support agents with an AI assistant resolved 14% more issues per hour. Experienced developers working in code they knew well took 19% longer with AI tools, and still believed they had been faster. The tools were similar. The work they were pointed at was not.

We build AI features and automation for companies that want hours back, so we spend a lot of time on the question before the build: should this task go to AI at all? Below is the test we'd run on any candidate task, what the studies found in software teams and in operations, and the work AI adds after launch that rarely makes it into the business case.

What The Research Actually Shows

Four studies get quoted most often in this debate. They look contradictory until you compare the tasks.

Compare the tasks and the pattern shows up. AI helped most where the worker lacked context the tool could supply: a new support agent suddenly had the best agents' answers on screen. It helped least where the worker already had all the context, like a maintainer working in a codebase they had shaped for years. And output that arrives faster than the next step can absorb simply piles up at that step. In the Faros data, that step was code review.

One caveat we'd add ourselves. The METR developers used early-2025 tools, coding assistants have improved since, and METR calls its result a snapshot. The mechanism behind the slowdown hasn't changed, though: time spent prompting and then checking the answer still counts as time.

The Five-Question Test Before You Automate Anything

Run each candidate task through these questions before it goes anywhere near an estimate. An hour with the person who does the task today is usually enough.

1. How long does checking take compared with doing? If a person needs four minutes to do a task and three to verify the AI's version, the ceiling on savings is one minute, and that's before you count prompting and fixes. Strong candidates take seconds to check: a field matches the source document, a ticket sits in the right queue.
2. What does one missed error cost? A misrouted ticket costs a few minutes. A wrong figure on an invoice costs money and a client's trust. The more an error costs, the more review you keep, and review eats the savings.
3. Is the knowledge written down? AI works from what it can read. If the rules live in one manager's head, the system will guess, and someone will spend their week correcting the guesses. This is the METR effect outside software.
4. Does the task repeat in the same shape? Two hundred similar invoices a week is a pattern. A dozen unique negotiations a month is not. Volume is what pays back the setup cost.
5. Where does the saved time go? If the next step (an approval, a review, someone who has to sign) is already the bottleneck, faster input only makes the queue longer. Fix that step first or the savings never reach delivery.

A task that clears all five is a strong candidate. Two or more "no" answers usually mean the first win is writing the process down, which often saves time with no model involved.

Tasks Where AI Usually Saves Time, And Where It Doesn't

These are typical outcomes, not guarantees. Your volume and error costs can move any row.

Look at the rows where AI saves time. In almost all of them AI is a narrow job inside a workflow, not a chat window someone opens when they remember to. That difference accounts for much of the gap between the 2.8% from Denmark and the 14% from customer support.

In Software Teams: Faster Code, Same Delivery Date

For a CTO the trap is measuring the wrong step. Coding assistants clearly speed up writing code. Delivery is writing plus everything after it: review, testing, integration, release. The Faros numbers show where the extra output went. Pull requests got bigger and the review queue nearly doubled, while delivery metrics stayed where they were.

Where Coding Assistants Earn Their Keep

Tests for code that already works, migrations, one-off scripts, API client boilerplate, first passes in a framework the developer hasn't used before, prototypes nobody will ship. In each case the developer can tell in seconds whether the output works.

Where They Cost Time

Subtle changes in a mature codebase, and code whose correctness depends on business rules nobody wrote down. Security-sensitive code belongs here too, since it needs a careful human read whoever wrote it. Developers in the METR study accepted less than 44% of AI-generated code and spent about 9% of their time cleaning up AI output.

What Actually Speeds Up Our Delivery

Where we see our own delivery speed up most is reuse of production-proven code from earlier projects: RAG pipelines and vector database setups, prompt frameworks, API integration layers, data preprocessing logic. Adapting something that already runs in production beats generating something new and debugging it. QA works in parallel with development and checks each feature as soon as it's ready, so finished work doesn't wait for a testing phase at the end. More on how we run sprints in our software development methodology guide.

In Business Operations: Built-In Beats Bolted-On

The Danish study measured people using general chatbots on their own initiative: about an hour a week saved, part of it spent on new tasks the AI created, such as checking its answers. The support study measured a tool built into the agents' workflow and fed with the company's own best answers. Most of the gap between those numbers comes down to that: a chatbot on the side versus a tool built into the work.

That's why we build system-first. The model is one component. What decides whether a task gets faster is the plumbing around it: data pipelines and retrieval feeding the model, business logic checking what comes out. For most operations work that means RAG over your own documents and data, wired into the systems where the work already happens. Our business process automation page covers the integration side, and the guide on adding AI to an existing product goes deeper into architecture.

Documentation Is The Cheapest Accelerator

The Work AI Adds After Launch

Business cases count the hours AI removes and skip the hours it adds. Put these in the plan before you commit.

• Review time. Someone checks outputs, at least on a sample, for as long as the system runs. Budget it as part of a role, not as a favor someone does on Fridays.
• Exceptions. Whatever the AI can't handle goes to a person, and those are by definition the hardest cases. That queue needs an owner.
• Upkeep. Prompts and knowledge-base documents need updating, and model providers ship versions that behave differently. Our post-launch support for AI systems runs $700 to $1,500+ a month and covers model updates and infrastructure monitoring.
• Running cost. Every model request costs money. At volume, cost per request decides whether the saved time is worth buying. We route across OpenAI, Claude, Gemini and OpenRouter partly to keep that number in check, and for narrow, high-volume tasks a small language model on your own servers can bring it down further.

A good use case still pays back after you add these up. If yours doesn't, better to find out on a spreadsheet.

How To Measure Whether AI Saved Time

The METR developers believed they were 20% faster while measuring 19% slower. So don't ask the team how it feels. Time the task.

Time The Task Before You Touch It

Take 20 to 50 real cases and record how long each one takes today, start to finish, hand-offs included. That's your baseline. Without it, every later number is a feeling.

Pilot On Real Work, End To End

Run the AI version on the next batch of real cases and time the whole loop: triggering it, waiting, checking, fixing, and the person downstream who receives the result. Count errors that slipped through separately from errors someone caught.

Pick The Number Up Front

Agree before the pilot what saving justifies the running and upkeep cost. If you need 30% off handling time and the pilot shows 8%, you have your answer, and it cost you a pilot instead of a full build.

How Our Team Decides Where AI Belongs

Discovery comes first: up to five business days and $1,000–$1,500, covering product and competitor analysis plus the choice of AI approach for the use case. It's the right moment to run the five questions with the people who do the work today. Sometimes the outcome is a plain integration with no model in it, which costs less to run and is easier to check.

After that, about half of the SRS is written before development starts. It describes every page and every button and is handed over in Notion. MVP development typically costs $2,000–$10,000, and RAG solutions usually take around two months; our AI development cost breakdown walks through every stage. Sprints last one or two weeks and end with a demo and a written progress document. The budget sits in our tracker next to the scope, so you can see what the hours are buying without asking for a report.