Will an AI agent actually save my business time?

In a careful 2025 study, experienced developers expected AI to make them 24% faster. Afterwards they still believed it had made them 20% faster. The clock said they had taken 19% longer (METR, July 2025).

That is the most useful fact I know about AI and time. The lesson is simple: a feeling is not a measurement.

So, will an AI agent save your business time? Yes, often, on short, repeated jobs that are easy to check. Not on everything, and not as much as the demo suggests. The only way to know for your business is to time the job before and after, and count the checking.

The time ledger

An agent’s speed is only one line on the ledger. The full ledger has five:

Time it gives back

• The work it does while you do something else

Time it takes

• Setting it up and teaching it your rules
• Checking what it did
• Fixing what it got wrong
• Redoing the jobs it failed

Most claims count only the left column. Anthropic, which makes Claude, says its own estimate of task speed-ups “can’t account for additional time humans spend on tasks outside of their conversations” (Anthropic, November 2025). The right column is where your saving is won or lost.

What the real results look like

1. Company story3+ hours a week → under 15 minutes a dayA two-person people team at Figma, checking background checks
2. Company story3x less timeQonto, making client invoices, with a person approving
3. Trial with a comparison group19 minutes a dayUK Department for Work and Pensions, Copilot users
4. Survey2.2% of work hoursUS workers, what they say AI saves
5. Survey paid for by a vendor74% save an hour a week or moreAustralian business owners who use AI
6. Controlled study19% slowerExperienced developers, who felt 20% faster

From the biggest saving to the smallest:

• A narrow team job, measured by the team: a two-person people team at Figma cut background-check monitoring “from more than three hours a week to a daily scan of under 15 minutes”, with a person still deciding (Zapier, 2026).
• A finance agent with approval: Qonto, which serves small businesses, says creating client invoices takes “3X less time”, and “any action with a financial or customer relationship impact requires user confirmation” (Anthropic customer story).
• A careful government trial: the UK Department for Work and Pensions compared Copilot users with non-users. Users saved about 19 minutes a day across eight routine tasks (DWP, January 2026).
• Across a whole economy: US workers say AI now saves 2.2% of all work hours, up from 1.6% (St Louis Fed, August 2026).
• Australian owners, self-reported: in a survey paid for by OpenAI, 74% of owners who use AI said they save at least an hour a week (Dynamic Business, September 2026).
• The slowdown: the developers above, 19% slower while feeling faster (METR).

The first two are company stories, picked because they went well. The middle ones are big and careful. Read them together and the shape is clear: large savings on narrow jobs, modest savings on average.

Why the speed-up shrinks

Agents do not always finish. Anthropic looked at how businesses use Claude through its API. About 49% of tasks succeeded, and for jobs that take a person five hours or more, about 45%. Counting the failures “roughly halves the implied gains” (Anthropic, January 2026).

Checking eats time. In the UK trial, people “consistently reviewed and edited” the output, and “in some cases, this additional step negated the time savings” (DWP). The Australian Government’s own Copilot trial found the same: reading a whole document to check it “significantly reduced any efficiency gains” (Digital Transformation Agency).

Some jobs just do not get faster. A UK department found that scheduling, making images and producing slide decks with Copilot “did not save time” (Department for Business and Trade, September 2025).

None of this is a reason to wait. It is a reason to pick the first job carefully.

A good first job has four signs

1. 1. It repeats.Daily or weekly, many times.
2. 2. The finished result is clear.A reply sent, an invoice made, a booking in the calendar.
3. 3. A person can check it in a minute.
4. 4. A wrong answer is cheap to catch.Nothing goes to a customer or a bank before someone looks.

Australian owners already use AI for exactly these jobs: turning site notes and photos into quotes, drafting customer emails, tracking jobs (Dynamic Business). One Queensland manufacturer cut its quoting time “from a week to 10 minutes” (NAB, July 2026). If your first job needs your calendar, inbox and job system to work together, that is the kind of small custom AI app I build. For the difference between a chatbot, an automation and an agent, see chatbot, automation or AI agent: which one does your business need?

The four-week test

Weeks 1–2: without the agentWeeks 3–4: with the agent

Date	Job	Start	Finish	Checking (min)	Fixing (min)	Redo?

vaisoftlab.com

Weeks one and two: no agent. Every time the job is done, write the start time, the finish time, and whether it needed fixing later. “If you want to save time, know where the time is going now,” as Lee Hickin of Australia’s National AI Centre puts it (National AI Centre, September 2026).

Weeks three and four: with the agent. Same job, same sheet, plus three columns: minutes checking, minutes fixing, and jobs it failed. Keep a person approving anything that goes to a customer or touches money.

End of week four: compare totals. Keep the agent if the total time per job is clearly lower and the quality is the same or better. Change the job, or stop, if not.

Two honest limits. Four weeks is short: Microsoft found most people only felt the full benefit after about 11 weeks (Microsoft WorkLab). And people get better with practice; Anthropic saw experienced users succeed more often (Anthropic, March 2026). So a close result after four weeks is worth another month. This is the same before-and-after check I build into every project.

Decide what to do with the hour

The National AI Centre’s post on this is called “You’ve got an hour back. Now what?” Saved time does not turn into money by itself: Danish researchers found “precise null effects” on hours and pay two years after ChatGPT arrived (Humlum and Vestergaard, NBER), and in NAB’s July 2026 survey only 9% of small businesses linked AI to higher profit, while 22% reported “transformational improvements” (NAB). Write down, before the test, what the saved hours are for: more site visits, faster replies to leads, an earlier finish on Friday.

The numbers, by year

Share of US work hours saved by AIWhat US workers report, 2025 to 2027.

1. 20251.6%
2. 2026 (second quarter)2.2%
3. 2027Dotted and empty: no good number yet.

Sources: St Louis Fed, On the Economy, 13 Nov 2025; FRED Blog, 27 Aug 2026.

Australian businesses using AITwo counts that ask different questions.

1. ABS: small and micro businesses with AI built into their work, 2024–2511%
2. NAB survey: small and medium businesses using AI, July 202640%

Sources: Australian Bureau of Statistics, 25 Jun 2026; NAB News, 1 Jul 2026.

• 2025. US workers report AI saves 1.6% of work hours (St Louis Fed). Australian SMEs using AI: 45% in the National AI Centre tracker, September to November 2025 (National AI Centre).
• 2026. US: 2.2% (FRED Blog). Agents in real business use are still “in the single digits across nearly all business functions” (Stanford AI Index). Australia’s official count, which asks a stricter question, puts AI use at about 11% of small and micro businesses (ABS); NAB’s survey says 40% of SMEs (NAB).
• 2027. No one has a good number for time saved yet. In Deloitte’s survey, 74% of leaders expect their companies to use AI agents at least “moderately” by 2027. My guess: more of the routine office work moves to agents, and the businesses that measured first will know exactly which jobs to hand over.