Which tasks in your business could AI or automation help with?

01

What is an AI opportunity audit, and why start with tasks rather than tools?

It is a structured review of the work your business does each week to find tasks where AI or automation could help safely. Starting with tasks avoids buying a tool and then hunting for a use.

The audit produces a ranked list. Each task has a score, any risk flags, a decision (automate, automate with review, or leave manual) and a next step. The Department for Science, Innovation and Technology (DSIT) AI Adoption Research listed "no identified need for AI" (71%) and limited AI skills, expertise or knowledge (60%) as the most cited barriers to adoption. A task list answers the first barrier directly, because it shows where a need exists.

Automation and AI are not the same. Automation follows fixed rules, such as "when a form is submitted, send a confirmation". Generative AI produces or interprets text, images or data, and its output can vary. Many tasks only need simple automation, and the audit should say so.

02

How many UK businesses use AI, and which figure should I trust?

It depends on who is counted. DSIT's 2025 survey of businesses with five or more employees found 16% using at least one AI technology. The 2026 UK Business Data Survey found 41% of businesses that handle digitised data use AI for something. Compare the bases, not the headlines.

Among adopters, DSIT found text generation and natural language tools were the most common technology (85%), and administrative and support tasks were a common use (70%). Adoption figures count use, not results, so a business can be in either percentage and still gain little.

03

How do I list the tasks in my business?

Record every recurring task for one normal week: what it is, how long it takes, how often it happens, who does it, which systems it uses and what information it handles. Write each as a verb and an object, such as "answer pricing questions".

1. 01
   Capture for one week

   Ask everyone, including the owner, to jot down tasks and minutes as they work. Owner tasks are often the most expensive.

2. 02
   Split broad jobs

   "Handle enquiries" becomes receive, categorise, reply with standard information, book a call and log the contact.

3. 03
   Add the factsFull step in the download ↓
4. 04
   Note who checksFull step in the download ↓

The list is the raw material for scoring. A vague entry such as "marketing" cannot be scored, so keep splitting until each line describes one kind of work.

04

How should I score each task for automation?

Rate each task against five tests: how repetitive it is, how rules-based, how high-volume, how low-risk if it goes wrong and how low in data sensitivity. Add the ratings together. A higher total means a better candidate for automation.

The first three tests say whether automation can work and pays back. The last two say whether it is safe. The template in this library scores volume, repetition and error risk, so the rules-based and data sensitivity checks need your own judgement.

05

Which tasks should I not give to AI?

Leave tasks where an error could harm someone or breach the law, where the answer needs professional judgement or where special category or confidential data is involved. AI output can be wrong, so a person must be able to check and correct it.

Accuracy is a real concern. DSIT's research found smaller and micro businesses mostly raised the accuracy of AI as their main concern. Treat generated text as a draft that someone checks against the source.

On data protection, ICO guidance says using AI will involve processing likely to result in a high risk to people's rights and freedoms, so a data protection impact assessment (DPIA) is usually needed when personal data is involved. A DPIA is a written assessment of the risks to people and how you will reduce them. If you decide you do not need one, document why.

06

How do I test an automation safely before I rely on it?

Test it on one high-scoring task at small scale, with a person checking every output and a clear rule for when to stop. Reduce the checking only once the error rate is low enough by a standard you set in advance.

1. 01
   Baseline

   Time the task and count errors for two weeks before changing anything.

2. 02
   Set the boundary

   Decide which tool, which inputs and which data it may see. Keep sensitive data out.

3. 03
   Run in shadow modeFull step in the download ↓
4. 04
   Log errors by typeFull step in the download ↓
5. 05
   Agree a stop ruleFull step in the download ↓
6. 06
   Review at four weeksFull step in the download ↓

Name one person as owner, and make sure customers can always reach a human.

07

How do I estimate time and money saved without fooling myself?

Use your own timed tasks. Hours removed per week, minus review hours, times your hourly value, times 52, minus annual tool and set-up costs. Quoted averages are not a substitute for your own measurements.

Hours saved only turn into money if you use the time for paid work or remove a cost. Count the review time and the time spent learning the tool as well.

Google's blog post of 25 April 2025 on its AI Works pilot says workers across sectors estimate generative AI saves them roughly over 122 hours a year, compared with a modelled estimate of 100 hours. The pilot involved 1,700 people. The post describes the figure as workers' estimates and does not describe the method behind it. Treat it as what some pilot participants reported, not as a promise.

08

What do I do with the scores, and when is outside help worth it?

Rank tasks by score, start with the highest-scoring safe one and review the result after four weeks before adding the next. Seek outside help when data protection, system connections or output quality are beyond your own time or skills.

Repeat the audit about every six months. Tasks change, tools improve, and rules and guidance move. The ICO says its guidance on AI and data protection is under review following the Data (Use and Access) Act.

Reeve Consult offers AI consultancy for UK businesses, starting from the task rather than the tool. We cannot promise any particular saving, because results depend on the task, the data, the tools and how the process is used.