# Which jobs should we automate first, and which should we leave alone?

Canonical: https://fortitudemedia.ai/insights/which-jobs-to-automate-first
Author: David Adams
Published: 2026-11-18
Last reviewed: 2026-11-18
Reviewed by: David Adams
Pillar: Running the firm on systems
Summary: A way to rank work for AI automation by volume, rules and risk, with examples by department and the tasks a B2B firm should leave with its people.
Keywords: which jobs to automate first, what tasks can ai automate, ai automation priorities, what not to automate with ai, automate back office tasks, human in the loop approval
Short answer: Automate tasks, not jobs, and rank each task on three tests: how much of it there is, how closely it follows rules, and what a mistake costs. Start with high-volume, rule-following, low-risk work such as record keeping, reporting, payment reminders and prospect research. Have a person approve anything that reaches a customer or moves money. Leave complaints, advice, pricing, hiring choices and significant decisions about individuals with people. Most firms should begin in administration and the back office.
Publisher: Fortitude Media Limited

When quoting, attribute to Fortitude Media and link to the canonical URL above.

---
Someone on your team has a list. It has forty ideas for using AI on it, from writing proposals to answering the phone, and every one sounds plausible. You can afford to do two or three properly this year. The question is which.

Picking badly is a common way for these projects to fail. A study by MIT's NANDA initiative, [reported by Fortune](https://fortune.com/2025/08/18/mit-report-95-percent-generative-ai-pilots-at-companies-failing-cfo/) in August 2025, found that more than half of generative AI budgets went on sales and marketing tools, while the biggest returns came from back-office automation. Most pilots in the study stalled with little or no measurable effect on profit.

This guide gives you a way to rank work by three things: how much of it there is, how closely it follows rules, and what happens when it goes wrong. It also says what to leave with people, with examples by department.

## Key takeaways

- Rank each task on volume, rules and risk. High volume, clear rules and low risk go first.
- Automate tasks, not jobs. Most roles mix routine steps with judgement, and the International Labour Organization expects jobs to be transformed more often than replaced.
- AI helps most on tasks inside what it does well. In a Harvard and BCG experiment, consultants using it finished 12.2% more tasks and worked 25.1% faster, but were 19 percentage points less likely to get the right answer on a task outside its strengths.
- Anything that reaches a customer or moves money should be approved by a person. Your firm stays responsible for what its systems say.
- Leave complaints, advice, pricing, hiring choices and significant decisions about individuals with people.
- Start in administration and the back office, where UK firms already use AI most and the MIT study found the best returns.

## What makes a task a good one to automate first?

Three things: there is a lot of it, it follows rules you could write down, and a mistake is cheap to catch and fix. A task that passes all three is a safe first project. A task that fails any one needs more care, or should wait.

| Test | Question to ask | Good sign | Warning sign |
|---|---|---|---|
| Volume | How many times a week does this happen, and how many hours does it take in total? | Daily or weekly, across several people | A few times a year |
| Rules | Could you write the steps on one page so that a new starter could follow them? | Same inputs, same steps, same output | "It depends" at every step |
| Risk | If it goes wrong, who sees it, what does it cost, and can it be undone? | Internal, visible, easy to correct | Reaches a customer, moves money, or affects someone's job or credit |

Volume tells you whether the saving is worth the effort. Rules tell you whether it can be built at all. Risk tells you how much checking it needs, and whether a person must approve the result.

## Should we think in jobs or in tasks?

Tasks. Very few jobs are routine from end to end, and very few have no routine in them.

Take a credit controller. The [National Careers Service](https://nationalcareers.service.gov.uk/job-profiles/credit-controller) lists the day-to-day work as checking credit ratings and setting up accounts, speaking with customers to arrange repayments, starting legal proceedings when debts are not paid, processing payments and keeping records up to date, and producing credit reports for management. Record keeping and reports follow rules. Reminders follow a schedule. The phone call with a customer in difficulty, and the decision to go to court, do not.

The research points the same way. The [International Labour Organization](https://www.ilo.org/publications/generative-ai-and-jobs-refined-global-index-occupational-exposure) found in May 2025 that one in four workers is in an occupation with some exposure to generative AI, with clerical occupations the most exposed, and concluded that "transformation of jobs is the most likely impact". A [study of 5,179 customer support agents](https://www.nber.org/papers/w31161) by Erik Brynjolfsson and colleagues found that an AI assistant raised the number of issues resolved per hour by 14% on average and by 34% for novice and low-skilled workers, with minimal effect on experienced ones. The tool took over the routine part of the job and helped newer people most.

So list tasks. You will end up changing what a role contains, not deleting the role.

## How do we score the list?

Give every task a score from 1 to 3 on each test, then sort. It takes a team lead about an hour for their own area.

### Step 1: List tasks, not roles

Ask each team lead for the tasks their people repeat. "Finance" is not a task. "Send the second reminder for an overdue invoice" is.

### Step 2: Score volume

Score 3 for work that happens daily and adds up to several hours a week, 2 for weekly work, 1 for anything monthly or rarer. Write the hours a week next to it.

### Step 3: Score rules

Score 3 if the steps fit on a page and the same input always leads to the same output. Score 2 if most cases follow the rules and a few need a person. Score 1 if every case is different.

### Step 4: Score risk

Score 3 for low risk: the output stays inside the firm and can be corrected. Score 2 if it reaches a customer or moves money but a person can check it first. Score 1 if a mistake could harm a client, breach a regulation, or decide something important about an individual.

### Step 5: Sort into four groups

| Group | Profile | What to do | Examples |
|---|---|---|---|
| Automate | High volume, clear rules, low risk | Let the system do it and check a sample | Updating customer records, writing up call notes as actions, the weekly numbers |
| Automate with approval | High volume, clear rules, reaches a customer or moves money | The system prepares it, a person approves it | Invoices, payment reminders, first sales emails, standard contracts |
| Assist | Repeated, but each case needs judgement | AI drafts or researches, a person does the work | Proposals, client reports, hard customer cases with a summary attached |
| Leave alone | Low volume, or high stakes | People do it | Complaints, pricing, advice, hiring choices |

### Step 6: Choose two or three and write the approval rule

Take the highest scores from the first two groups. For each, write down who approves the output, where the record of each action is kept, and how you will know it is working. [What automating a department costs compared with hiring](/insights/cost-of-automating-a-department-vs-hiring) covers the money side, and our own prices are on [what it costs](/what-it-costs).

## Which tasks come first, department by department?

In most firms the first candidates are in administration, sales support and finance admin. The table below uses the split we publish on our [Business Operating System](/business-operating-system) page.

| Department | Good first tasks | Leave with people |
|---|---|---|
| Marketing | Turning customer questions and sales calls into draft posts, emails and pages; scheduling; the monthly report | Approving the message and the tone |
| Sales | Building and scoring prospect lists, researching each firm, drafting first emails and follow-ups, updating records | Approving what is sent, taking the meeting |
| Customer service | Answering common questions from your own help pages, passing hard cases on with a summary | Complaints and judgement calls |
| Finance | Raising the invoice when a deal is signed, chasing late payers on a schedule, the weekly cash report | Sign-off of the books by your accountant |
| Contracts | Filling standard contracts from templates, tracking renewal dates, flagging unusual clauses | Legal advice from your solicitor |
| Hiring | Writing job ads, booking interviews, onboarding checklists | Choosing who to hire |
| Reporting | Pulling numbers from every tool into one weekly page | Deciding what to do about it |

Outside evidence supports starting here. The government's [AI adoption survey](https://www.gov.uk/government/publications/ai-adoption-survey-2026/ai-adoption-survey-report-2026-executive-summary) of 3,503 UK businesses found that AI was most commonly embedded in administration (81% of those using or planning to use it), marketing (71%) and IT (66%). McKinsey's [2023 study](https://www.mckinsey.com/capabilities/tech-and-ai/our-insights/the-economic-potential-of-generative-ai-the-next-productivity-frontier) estimated that about 75% of the value of generative AI falls in four areas: customer operations, marketing and sales, software engineering and research and development. Anthropic's [Economic Index for March 2026](https://www.anthropic.com/research/economic-index-march-2026-report), which tracks how its Claude models are used, reports that customer service tasks such as payment and billing support are common in the work businesses run through its API, and that sales and outreach automation at least doubled its share.

Two cautions on sales and marketing. Outbound email has legal rules of its own, covered in [whether cold emailing other businesses is legal in the UK](/insights/cold-email-law-uk-b2b). And content written by AI with nobody's expertise behind it does little for you, as we explain in [whether ChatGPT and Google will quote content written with AI](/insights/thought-leadership-vs-keyword-stuffing).

## What should we leave alone?

Leave anything where the judgement is the job, where a mistake is expensive, or where the law expects a person. Five kinds of work belong here.

**Complaints and difficult conversations.** Klarna said its AI assistant could do the work of 700 customer service agents. In 2025 its chief executive told Bloomberg that cost had been "a too predominant evaluation factor" and that "what you end up having is lower quality", and the company began hiring human agents again, as [CX Today reported](https://cxtoday.com/klarna-grapples-with-ai-led-customer-service-pivots-to-an-uber-type-of-setup).

**Promises to customers.** In February 2024 a Canadian tribunal found Air Canada liable for negligent misrepresentation after its website chatbot gave a customer wrong information about bereavement fares. The tribunal held that the airline was responsible for all the information on its website, whether it came from a static page or a chatbot, according to the [American Bar Association's account](https://www.americanbar.org/groups/business_law/resources/business-law-today/2024-february/bc-tribunal-confirms-companies-remain-liable-information-provided-ai-chatbot/). It is a Canadian case, but the lesson travels: your system speaks for you.

**Advice and regulated work.** The Solicitors Regulation Authority tells law firms in its [report on AI](https://www.sra.org.uk/sra/research-publications/artificial-intelligence-legal-market) that "you will remain responsible and accountable for the outputs from AI you are using", and adds: "Do not trust an AI system to judge its own accuracy." Other regulated professions should assume the same.

**Significant decisions about individuals.** Under the Information Commissioner's Office's [draft guidance](https://ico.org.uk/for-organisations/uk-gdpr-guidance-and-resources/individual-rights/automated-decision-making/what-are-the-adm-safeguards/), updated in March 2026, a person affected by a significant decision made solely by automated means must be told about it, be able to put their side, get a human to review it and challenge it. The reviewer needs the authority to change the outcome. The ICO's own example is a loan decision. If software alone ever rejects a job applicant or refuses a customer credit, expect the same rules to apply.

**Work AI is not good at.** In an [experiment with 758 BCG consultants](https://mitsloan.mit.edu/sites/default/files/2023-10/SSRN-id4573321.pdf), researchers from Harvard and elsewhere found that on tasks within AI's abilities the consultants using it completed 12.2% more tasks, 25.1% more quickly, at more than 40% higher quality. On a task chosen to be outside those abilities, consultants using AI were 19 percentage points less likely to produce a correct answer. It is not always obvious in advance which kind a task is, so test before you trust.

Leave low-volume work alone too, and any process that changes every month. [What stays with people when AI does the routine work](/insights/what-stays-with-people-when-ai-does-the-routine-work) goes further into this side.

## How much checking does automated work need?

More than you expect at the start, and less as the record builds. Begin with a person approving everything, then relax it only for internal, low-risk work.

Most UK firms already work this way. In the government survey, 81% of businesses said they apply some level of human input or oversight to AI outputs, and 56% of businesses that had adopted AI reported concerns about bias and accuracy.

Our rule is simple. Anything that reaches a customer or moves money is approved by a person, and every action is logged so you can see what the system did and why. We hold ourselves to it: our own sales desk drafts every first email and follow-up, and a founder approves every message. [How we run Fortitude](/how-we-run-fortitude) shows the split in each area.

Approval has to be real. A person who clicks "approve" on two hundred items without reading them is not a check. Keep batches small enough to read, and give the approver the authority to say no.

## What to do this month

1. Ask each team lead for the ten tasks their people repeat most often.
2. Score every task from 1 to 3 on volume, rules and risk.
3. Sort the list into the four groups: automate, automate with approval, assist, leave alone.
4. Pick the two or three highest scorers and write down the steps for each on one page.
5. Decide who will approve the output of each, and where the log will be kept.
6. Write down the tasks you have decided to leave with people, and tell the team.

If you would like a second opinion on your list, [book a conversation](/growth-review): 30 minutes with a founder, at no charge, and you leave with three areas a system could take on, in writing.

## Frequently asked questions

### Should we start with customer-facing work or the back office?

The back office, in most firms. The MIT study reported by Fortune found the biggest returns in back-office automation, and mistakes there are seen by you before they are seen by a customer. Customer-facing work can follow once the approval habit is in place.

### Can AI answer our customers directly?

For common questions with settled answers, drawn from your own help pages, yes. Hard cases, complaints and anything involving a promise or a refund should go to a person with a summary. Remember the Air Canada ruling: the firm answers for what its chatbot says.

### What if the process is a mess?

Fix it first. A task that scores 1 on rules cannot be automated well, because nobody can say what the right output is. Write the steps down, run them by hand for a month, then score the task again.

### Is it worth automating a task that takes ten minutes?

It depends how often it happens. Ten minutes, thirty times a day, across a team is 25 hours a week. Ten minutes once a month is not worth building anything for.

### Will automating tasks cost people their jobs?

The International Labour Organization's view is that jobs are more likely to change than to disappear, because most roles include tasks that need a person. In practice the routine part of a role shrinks, and the firm decides what that time is now for. Make that decision before you build, and tell the people affected.

## Sources

- [Fortune, MIT report: 95% of generative AI pilots at companies are failing (August 2025)](https://fortune.com/2025/08/18/mit-report-95-percent-generative-ai-pilots-at-companies-failing-cfo/)
- [National Careers Service, Credit controller job profile (checked November 2026)](https://nationalcareers.service.gov.uk/job-profiles/credit-controller)
- [International Labour Organization, Generative AI and jobs: A refined global index of occupational exposure (May 2025)](https://www.ilo.org/publications/generative-ai-and-jobs-refined-global-index-occupational-exposure)
- [National Bureau of Economic Research, Generative AI at Work (April 2023, revised November 2023)](https://www.nber.org/papers/w31161)
- [GOV.UK, AI adoption survey report 2026: executive summary (October 2026)](https://www.gov.uk/government/publications/ai-adoption-survey-2026/ai-adoption-survey-report-2026-executive-summary)
- [McKinsey, The economic potential of generative AI: The next productivity frontier (June 2023)](https://www.mckinsey.com/capabilities/tech-and-ai/our-insights/the-economic-potential-of-generative-ai-the-next-productivity-frontier)
- [Anthropic, Economic Index report (March 2026)](https://www.anthropic.com/research/economic-index-march-2026-report)
- [CX Today, Klarna grapples with AI-led customer service, pivots to an "Uber type of setup" (May 2025)](https://cxtoday.com/klarna-grapples-with-ai-led-customer-service-pivots-to-an-uber-type-of-setup)
- [American Bar Association, BC tribunal confirms companies remain liable for information provided by AI chatbot (February 2024)](https://www.americanbar.org/groups/business_law/resources/business-law-today/2024-february/bc-tribunal-confirms-companies-remain-liable-information-provided-ai-chatbot/)
- [Solicitors Regulation Authority, Risk Outlook report: The use of artificial intelligence in the legal market (November 2023)](https://www.sra.org.uk/sra/research-publications/artificial-intelligence-legal-market)
- [Information Commissioner's Office, What are the ADM safeguards? Draft guidance (updated March 2026)](https://ico.org.uk/for-organisations/uk-gdpr-guidance-and-resources/individual-rights/automated-decision-making/what-are-the-adm-safeguards/)
- [Harvard Business School, Navigating the Jagged Technological Frontier (September 2023)](https://mitsloan.mit.edu/sites/default/files/2023-10/SSRN-id4573321.pdf)
