# What should stay with people when AI does the routine work?

Canonical: https://fortitudemedia.ai/insights/what-stays-with-people-when-ai-does-the-routine-work
Author: Ross Williams
Published: 2026-11-18
Last reviewed: 2026-11-18
Reviewed by: Ross Williams
Pillar: Running the firm on systems
Summary: What should stay with people when AI takes on routine work: judgement, relationships, exceptions and accountability, how approvals work, and the evidence.
Keywords: human in the loop ai, what work should stay with people ai, ai approval steps, will ai replace jobs uk, ai and staff roles, meaningful human involvement
Short answer: Four things: judgement, relationships, exceptions and accountability. The system prepares the work, and a named person approves anything that reaches a customer or moves money, with the authority to change it. Significant decisions about people stay with a person. The evidence so far shows more change inside jobs than loss of jobs, though junior and clerical roles are the most exposed. Tell staff early what the system will take on, what stays with them and what training they will get.
Publisher: Fortitude Media Limited

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

---
You have started handing routine work to AI, or you are about to. The invoices raise themselves, the follow-up emails are drafted before anyone asks, and the weekly report writes itself. Then two questions arrive. A client asks who checked this. One of your best people asks what happens to their job.

Both deserve a clear answer, and it is the same answer. Four things stay with people: judgement, relationships, exceptions and accountability. Steve Hasker, chief executive of Thomson Reuters, put it this way when the firm published its [2026 survey of more than 1,800 professionals](https://www.thomsonreuters.com/en/institute/articles/future-of-professionals-2026): "the judgment, relationships and accountability remain human, and that won't change."

This guide explains what those four things mean in a working firm, how approval steps work in practice, what changes for roles, what the published evidence says about jobs, and what to tell staff who are worried.

## Key takeaways

- Four things stay with people when AI does the routine work: judgement, relationships, exceptions and accountability.
- A sound approval step gives a named person the work before it goes out, the authority to change it and the training to understand what the system did.
- Approval fails when it turns into a tick-box: the Financial Conduct Authority has seen reviewers accept AI output without scrutiny.
- So far, the evidence shows more change inside jobs than loss of jobs: the ONS found around half of UK businesses reported no effect on headcount from AI in June 2026.
- Junior and clerical roles are the most exposed, so plan how new starters will learn the job.
- Tell staff early what the system will take on, what stays with them and what training they will get.

## What should stay with people?

Judgement, relationships, exceptions and accountability. The system prepares the work, and a person decides, talks to the client, handles what the rules did not foresee and answers for the result.

| What stays | What it means | An example |
|---|---|---|
| Judgement | Deciding what is right when the answer is not in the rules | Whether to take on a client, what to say in a difficult email, what to stop doing |
| Relationships | Being the person a client, candidate or supplier trusts | Taking the meeting, giving bad news, asking for the renewal |
| Exceptions | Handling the case the process was not written for | A complaint, an unusual clause, a customer in difficulty |
| Accountability | Putting a name to the result | Signing off the accounts, approving a payment, sending the advice |

This is how we draw the line in the systems we build. The table below uses the split set out on our [Business Operating System](/business-operating-system) page.

| Area | What the system does | What stays with your people |
|---|---|---|
| Marketing | Drafts posts, emails and pages from customer questions and sales calls | Approving the message and the tone |
| Sales | Builds prospect lists, researches each firm, drafts first emails and follow-ups | Approving what is sent, and taking the meeting |
| Customer care | Answers common questions from your own help pages | Complaints and judgement calls |
| Finance | Raises invoices, chases late payers, sends a weekly cash report | Your accountant signs off the books |
| Contracts | Fills standard contracts and flags unusual clauses | Your solicitor gives the advice |
| Hiring | Writes job ads, books interviews, runs onboarding checklists | Your managers choose who to hire |

## Why can those four things not be handed to a system?

Because AI makes mistakes that look convincing, because clients still want a person at the moments that matter, and because responsibility in law sits with people and firms.

Take relationships first. A [Gartner survey of 645 business buyers](https://www.gartner.com/en/newsroom/press-releases/2026-05-20-gartner-survey-finds-sixty-nine-percent-of-b-two-b-buyers-turn-to-sales-reps-to-validate-ai-generated-insights), published in May 2026, found that 69% prefer to check what AI tells them with a sales rep. Reps were also the most important source of information when buyers were choosing a supplier and finalising a purchase.

Accountability is the firmest line. The Financial Conduct Authority's [review of AI in financial services](https://www.fca.org.uk/publications/good-poor-practice/ai-financial-services), published in October 2026, says that firms using AI tools "remain responsible for consumer outcomes". The same review describes good practice as choosing which tasks AI can support or automate, which should stay with staff, and where the tool should stop and pass the matter to a person.

Data protection law draws a similar line for decisions about people. The Information Commissioner's Office's [draft guidance on automated decisions](https://ico.org.uk/for-organisations/uk-gdpr-guidance-and-resources/individual-rights/automated-decision-making/what-does-the-uk-gdpr-say-about-adm/), updated in March 2026, covers decisions with a legal or similarly significant effect on someone, such as recruitment or pay. If no person is meaningfully involved, extra safeguards apply, including the person's right to have a human look at the decision.

## How do approval steps work in practice?

The system prepares the work, a named person approves it before it reaches a customer or moves money, and every action is logged. The skill is in deciding which items need approval, and in making the approval real.

### Step 1: Sort every output into three kinds

Not everything needs a person. Sort what the system produces by what happens if it is wrong.

| Kind | Examples | Who acts |
|---|---|---|
| Runs on its own | Updating a record, filing a document, internal reminders, the weekly numbers | The system, with a log |
| Approve before it goes | Emails to clients and prospects, invoices, payments, published pages, proposals | The system drafts, a named person approves |
| A person decides | Hiring and pay, complaints, advice, pricing an unusual deal, ending a client relationship | A person, with the system preparing the facts |

### Step 2: Name the approver and give them authority

The ICO's draft guidance sets out what makes human involvement meaningful, and it works as a checklist for any approval step. The person should review the work at a point where they can still change the outcome, have the authority to alter it, be trained to understand the system's logic, outputs and limits, and take the relevant facts into account. The ICO says involvement must be "active and not just a token gesture".

### Step 3: Show the approver what they need

Put the draft, the source it was built from and the reason for it on one screen. An approver who has to open four systems to check an invoice will stop checking.

### Step 4: Make sending it back easy, and count it

Record how often approvers change or reject what the system produced. A rate near zero for months may mean the system is very good, or that nobody is reading. A rising rate means the system needs fixing.

### Step 5: Guard against the tick-box

The FCA review describes one firm where the reviewer's role "became a tick-box exercise over time", with staff accepting AI recommendations even where there were signs of error. It describes another firm that had reviewers assess each case without seeing the AI's proposed answer first, then compared the two. Keep the number of approvals per person manageable, and have a second person sample the approved items each month.

### Step 6: Review the lines every quarter

As evidence builds, move items between the three kinds. Work that has run cleanly for six months may no longer need approval. Work that produced a complaint goes back to a person.

We hold ourselves to the strict version. Nothing reaches a client, a prospect or a bank account at Fortitude without a person approving it, and you can see the full list on [how we run Fortitude](/how-we-run-fortitude).

## What does this mean for roles?

Most roles keep their title and lose their routine tasks. A smaller number of roles made up mostly of routine work shrink or change shape. New work appears in checking, handling exceptions and spending time with clients.

The best-known study of AI inside a real job supports this. [Brynjolfsson, Li and Raymond](https://www.nber.org/papers/w31161) followed 5,179 customer support agents who were given an AI assistant. Productivity rose by 14% on average and by 34% for new and lower-skilled workers, with little effect on the most experienced. The authors found signs that the tool passed on the practices of the best agents to everyone else.

UK job adverts point the same way. [PwC's 2026 AI Jobs Barometer](https://pwc.co.uk/press-room/press-releases/research-commentary/2026/ai-jobs-barometer-2026.html) found that roles most enhanced by AI had grown 39% since 2018, against 17% for roles where AI mainly simplifies tasks. Postings for roles that use AI rose 65.8% in 2025 while overall vacancies fell 6.6%.

There are two cautions. The first is training. The ONS found in June 2026 that only 11% of businesses with ten or more staff said over half their workforce had received AI-related training. The second is the career ladder. If routine work was how your juniors learned the job, you need another way to teach them, such as having them check the system's output with a senior person and explain what they changed.

Our guide to [which jobs to automate first](/insights/which-jobs-to-automate-first) covers how to choose the work. For a marketing example of the same shift, see [whether to keep paying an agency now that AI can do the work](/insights/ai-disrupting-marketing-agency-model).

## What does the evidence say about how jobs change?

So far, more change within jobs than loss of jobs, with the earliest pressure on junior and clerical roles. The studies below are the most useful we have found. Several are forecasts or use data from the United States, so treat them as direction, not prediction.

| Source | What it found | What it means for a firm your size |
|---|---|---|
| [International Labour Organization](https://www.ilo.org/publications/generative-ai-and-jobs-2025-update), May 2025 | One in four workers worldwide is in an occupation with some exposure to generative AI, and most jobs will be transformed, not made redundant | Expect tasks to move before jobs do |
| [Office for National Statistics](https://www.ons.gov.uk/businessindustryandtrade/business/businessservices/articles/artificialintelligenceinukbusinesses/2023to2026), July 2026 | Around half of UK businesses reported no effect on headcount from AI; among those using it to improve operations, 6% reported a decrease | Most firms are not cutting, yet |
| [CIPD](https://www.cipd.org/en/about/press-releases/one-in-six-employers-say-ai-shrink-headcount-autumn-labour-market-outlook/), autumn 2025 | 17% of employers expected AI to shrink their workforce within a year; 62% of those said clerical, junior managerial, professional or administrative roles were most likely to go | The pressure is on routine office roles |
| [Stanford](https://siepr.stanford.edu/publications/working-paper/canaries-coal-mine-six-facts-about-recent-employment-effects-artificial), August 2025 | Workers aged 22 to 25 in the most AI-exposed occupations in the US saw a 13% relative fall in employment; experienced workers in the same jobs did not | Entry-level roles feel it first |
| [World Economic Forum](https://www.weforum.org/publications/the-future-of-jobs-report-2025/digest/), January 2025 | Employers expect 170 million jobs created and 92 million displaced by 2030, and 39% of skills to change | Skills change faster than job counts |

The honest reading is mixed. Most firms report no change in headcount so far, and people who work with AI are in demand. But the first rung of the ladder is getting harder to reach, and employers who expect to cut are looking at exactly the administrative roles that most automation touches.

## What do we say to staff who are worried?

Tell them the truth, early and in detail: what the system will take on, what stays with them, what happens to roles and what training they will get. Vague reassurance makes it worse.

The worry is common. [Acas](https://www.acas.org.uk/1-in-4-workers-worry-that-ai-will-lead-to-job-losses) found in a 2025 poll of 1,023 employees that 26% were worried AI would lead to job losses and 17% were worried about AI making errors. Its advice to employers is to have clear policies, consult employees on the introduction of AI, and show how it can improve people's roles.

In practice that means four things. Show each team the list of tasks moving to the system and the list staying with them. Involve the people who do the work in writing down how it is done, because they know where the exceptions are. Say what the freed hours are for. And do not promise that no role will ever change. If you expect to need fewer people in an area, say so, and take employment advice before you act.

Most firms use a system like this to grow without adding headcount, not to cut. If that is your aim, say it in those words. Our guide to [writing an AI plan for the board](/insights/ai-plan-for-the-board) covers how to put the same commitment in front of directors. The related worry about tone, that work drafted by a system will not sound like you, is covered in [whether you can outsource content without losing your voice](/insights/outsourcing-content-brand-voice).

## What to do this month

1. List what your AI tools and automations produce today, and sort each item into the three kinds: runs on its own, approve before it goes, a person decides.
2. Name an approver for every item in the second kind, and check they have the authority to change it.
3. Start counting how often approvers change or reject what the system produced.
4. Write down, for one team, the tasks moving to the system and the tasks staying with people, and share it with that team.
5. Decide how new starters will learn the judgement that routine work used to teach them.
6. Put a date in the diary to review the lines in three months.

If you would like help drawing the lines for your own firm, [book a conversation](/growth-review): 30 minutes with a founder, at no charge.

## Frequently asked questions

### Does a person have to approve everything the AI produces?

No. Internal, low-risk and easily reversed work can run on its own with a log. Anything that reaches a customer or moves money should be approved by a named person. Significant decisions about people should be made by a person.

### Is "human in the loop" a legal requirement in the UK?

For some decisions. Under UK data protection law, a decision with a legal or similarly significant effect on a person, made with no meaningful human involvement, needs safeguards. The ICO's draft guidance lists them: telling the person, letting them make representations, giving them human intervention and letting them contest the decision. Routine emails and invoices are not in that category.

### Who is accountable when the AI gets something wrong?

Your firm, and the person who approved the work. The FCA says regulated firms remain responsible for consumer outcomes when they use AI tools. An approval step makes that responsibility visible by putting a name against each item.

### Will AI replace our junior staff?

It will change what they do. The Stanford and CIPD findings show junior and clerical roles are the most exposed. Firms still need people who will become senior, so move juniors on to checking, exceptions and client contact earlier, with supervision.

### How do we stop approvers rubber-stamping?

Limit how many items each person approves, show them the source alongside the draft, count how often they send work back, and have a second person sample approved items. The FCA review found that a reviewer role with no clear purpose turns into box-ticking.

### What should we train people in?

How the system works and where it goes wrong, how to check its output, how to handle exceptions and how to spend more time with clients. In the World Economic Forum's survey, seven in ten employers called analytical thinking essential, more than any other skill.

## Sources

- [Thomson Reuters Institute, Future of Professionals 2026: As AI adoption grows, so do the challenges (June 2026)](https://www.thomsonreuters.com/en/institute/articles/future-of-professionals-2026)
- [Gartner, 69% of B2B buyers turn to sales reps to validate AI-generated insights (May 2026)](https://www.gartner.com/en/newsroom/press-releases/2026-05-20-gartner-survey-finds-sixty-nine-percent-of-b-two-b-buyers-turn-to-sales-reps-to-validate-ai-generated-insights)
- [Financial Conduct Authority, AI in financial services: good practice and areas for improvement (October 2026)](https://www.fca.org.uk/publications/good-poor-practice/ai-financial-services)
- [Information Commissioner's Office, What does the UK GDPR say about ADM? Draft guidance (updated March 2026)](https://ico.org.uk/for-organisations/uk-gdpr-guidance-and-resources/individual-rights/automated-decision-making/what-does-the-uk-gdpr-say-about-adm/)
- [National Bureau of Economic Research, Brynjolfsson, Li and Raymond, Generative AI at Work (working paper 2023, published 2025)](https://www.nber.org/papers/w31161)
- [PwC UK, UK AI hiring and wages surge as they outpace jobs market: AI Jobs Barometer (June 2026)](https://pwc.co.uk/press-room/press-releases/research-commentary/2026/ai-jobs-barometer-2026.html)
- [International Labour Organization, Generative AI and jobs: a 2025 update (May 2025)](https://www.ilo.org/publications/generative-ai-and-jobs-2025-update)
- [Office for National Statistics, Artificial intelligence in UK businesses: 2023 to 2026 (July 2026)](https://www.ons.gov.uk/businessindustryandtrade/business/businessservices/articles/artificialintelligenceinukbusinesses/2023to2026)
- [CIPD, One in six employers say AI will shrink headcount: autumn Labour Market Outlook (November 2025)](https://www.cipd.org/en/about/press-releases/one-in-six-employers-say-ai-shrink-headcount-autumn-labour-market-outlook/)
- [Stanford Institute for Economic Policy Research, Canaries in the coal mine? Six facts about the recent employment effects of artificial intelligence (August 2025)](https://siepr.stanford.edu/publications/working-paper/canaries-coal-mine-six-facts-about-recent-employment-effects-artificial)
- [World Economic Forum, The Future of Jobs Report 2025 (January 2025)](https://www.weforum.org/publications/the-future-of-jobs-report-2025/digest/)
- [Acas, 1 in 4 workers worry that AI will lead to job losses (April 2025)](https://www.acas.org.uk/1-in-4-workers-worry-that-ai-will-lead-to-job-losses)
