Hand over the repetitive half —
taking the repetitive halfNine kinds of role, split line by line
Roles covered
5
Tests that draw the line
4
Left with people
4 things
Headcount reduction promised
None
The tests first
The line isn't drawn by seniority — it's drawn by these four
Does it have one correct answer?
Specifications, stock, warranty periods, order status — those answers are looked up, and a machine is steadier at it than a person and never forgets. Anything requiring a judgement call, reading a room, or knowing how far to push has no single correct answer, and a machine will simply produce one for you.
Is a mistake reversible?
An image can be re-run. A quote that has been sent cannot be recalled. Where the cost is irreversible, the final action has to be a person's — not because the machine judges badly, but because there is no second attempt.
Does someone have to sign for it?
External commitments, legal exposure, concessions to a customer — the signature has to belong to someone who can answer for it. That isn't a technical question, and a better model doesn't change the answer.
Is the volume large enough?
Something that happens three times a month costs more to specify than it saves. Handing a task over is only worth it if it recurs every week — repetition is the only economics on this line.
The four together say one thing: what a machine takes is the part that's lookup-able, repeatable, unsigned, and happening every week. The rest isn't beyond it — it's that you wouldn't trust the output.
Nine roles
Line by line: what goes, what stays
Customer support
Taken by the machine
- ✓Specifications, stock, shipping and warranty — anything with a lookup answer
- ✓Order status and standard after-sales flows
- ✓Replies in every language, with terminology held consistent
- ✓First response around the clock, so nothing sits overnight
Left with people
- ✕Complaints and anyone who is upset
- ✕Negotiations involving a concession or compensation
- ✕Taking over when the machine can't answer — and it has to know that it can't
- ✕Reviewing transcripts periodically and adding what was missing to the knowledge base
The key design decision: when it can't find an answer it escalates and never improvises. A support bot that invents answers is worse than none.
Content and SEO
Taken by the machine
- ✓Keyword research and topic scheduling
- ✓First drafts in your brand's voice rather than generic filler
- ✓Every language in sync, against one glossary
- ✓Structured data, internal links and AI-crawler discovery files generated with it
Left with people
- ✕Deciding the direction and the content strategy
- ✕Final review: fact-checking, professional judgement, legal risk
- ✕Adding what a machine can't produce — first-hand experience, real cases, what's actually happening in your industry
The realistic ratio: a long-form article goes from 8 hours to 16 minutes, of which 15 minutes is human final review. What disappears is the first draft, not the checking — and those 15 minutes shouldn't be cut.
Visual assets
Taken by the machine
- ✓Batch background removal, scene replacement and consistent lighting
- ✓Hero, detail and social sizes exported in one pass
- ✓Style held to your visual standard so a batch doesn't drift
- ✓Short-form video source material produced at volume
Left with people
- ✕Setting the visual standard — a machine executes one, it can't decide one
- ✕Physical photography: materials, real usage settings, models
- ✕Selecting from each batch, and sending back what doesn't pass
Control is the hard part of this category, which is why the work is organised as a choice between tiers rather than one method.
Sales development
Taken by the machine
- ✓Finding prospects at volume and assembling background on them
- ✓Scoring and ranking leads so time goes to the right ones
- ✓Personalised first contact rather than a broadcast template
- ✓Follow-up cadence, so nothing is forgotten
Left with people
- ✕The actual sales conversation and the negotiation
- ✕Deciding the price and the concession
- ✕The relationship — that can't be handed to a machine
The goal isn't “AI sells for you”. It's that salespeople only handle qualified leads, with the filtering in front of that given away.
Finance
Taken by the machine
- ✓Invoices, receipts and statements extracted into tables instead of retyped
- ✓Two sets of records reconciled automatically, with only the lines that don't match pushed to a person
- ✓Expense pre-checks: over-policy, missing receipt, duplicate claim, flagged on the spot
- ✓Ageing and collection reminders sent by customer tier
- ✓Month-end aggregation and a first draft of the reports
Left with people
- ✕Every payment and credit approval
- ✕Accounting policy and tax treatment
- ✕Deciding what a discrepancy means — a typo, or something real
- ✕Confirmation letters and anything financial sent outside the company
The value here isn't speed of arithmetic. Finance teams don't spend their days calculating, they spend them hunting for differences — so what actually gets handed over is “surface the lines that don't match”, and none of the judgement moves.
People
Taken by the machine
- ✓CV screening and structuring, ranked against the role's requirements
- ✓Interview scheduling and candidate status kept in sync, so nobody is dropped
- ✓Joiner and leaver checklists dispatched automatically, each department receiving only its own line
- ✓Attendance and leave anomalies surfaced
- ✓Policy questions: how leave accrues, how to file expenses, how contributions work
Left with people
- ✕Hiring decisions and pay bands
- ✕Performance assessment and promotion calls
- ✕Conversations: exits, conflicts, feedback
- ✕Exceptions involving a person — and this function is almost entirely exceptions
This category has the highest data-sensitivity of all nine — salaries, identity documents, medicals, background checks — which is one reason we only do self-hosted work. That material shouldn't be sitting on somebody else's server.
Engineering
Taken by the machine
- ✓Requests and field feedback condensed into tickets: classified, de-duplicated, linked to the module they touch
- ✓Interface and configuration docs updated alongside the code rather than six months behind it
- ✓First drafts of code and config, test cases, migration scripts
- ✓Incident logs clustered, with similar past cases and their fixes surfaced
- ✓Release notes and changelogs drafted
Left with people
- ✕Architecture decisions and technology choices
- ✕The “why is it written this way” part of review
- ✕Trade-offs: what ships first, what gets cut
- ✕Calling a production incident and deciding to roll back
The easiest category to overestimate. First drafts arrive quickly, but review time doesn't shrink at all. The real gain is not queueing behind the one person who knows how to write that thing — not writing less code. Budget it as the former and you'll be disappointed.
Store operations
Taken by the machine
- ✓Daily trading summaries with anomalies flagged: footfall, conversion, returns
- ✓Walk-in records structured, with follow-up tasks assigned to a named person
- ✓Sample and stock count lists, and transfer requests routed automatically
- ✓Inspection photos and remediation items filed, with overdue items chased
- ✓Product knowledge and talk tracks staff can ask for at any moment, answers with sources
Left with people
- ✕Serving the customer in front of them, and negotiating
- ✕A manager's discretion over people and rotas
- ✕Merchandising, fit-out and promotion decisions
- ✕Handling a complaint on the floor
The hard part on the shop floor has never been technical, it's the entry point. Nobody installs a back-office app to file three records a day. The entry point has to sit inside the tool they already use, or the work simply won't happen.
Back office (admin and legal)
Taken by the machine
- ✓Approvals routed automatically by rule, and chased
- ✓Key contract clauses extracted and compared line by line against your standard
- ✓Meeting notes drafted and actions assigned
- ✓Documents and policy questions answered, so new starters stop asking colleagues
- ✓Daily and weekly reports generated
Left with people
- ✕Every approval that requires a signature and accountability
- ✕Discretion on exceptions
- ✕Setting the rules themselves
- ✕The final call on legal risk
Expenses and CV screening used to live in this category. On 2026-09-10 they moved into Finance and People, where the approach is settled enough to state a boundary — buried in “back office” they were invisible.
The half that stays
What the person actually ends up doing
Set the standard
Decide what “done right” means
A machine executes a standard, it doesn't set one. Set it wrong and it will do the wrong thing very efficiently, at volume.
Handle exceptions
Everything outside the rules
A process covers 85% of cases. The remaining 15% is all difficult. That's where people's time should go.
Carry accountability
The person who signs
External commitments, legal exposure, concessions to a customer — never a machine's decision to make.
Design the process
How this should work at all
The most overlooked one. Whoever understands the business best should decide the process itself — that can't be outsourced, and shouldn't be.
Which is why the return on this shouldn't be counted as headcount saved. It should be counted as whether those four things have ever been done properly. In most companies the answer is no — not from unwillingness, but because the other half filled the day.
Where to start
Pick the most repetitive one, not the hardest one
Look for the backlog
No survey needed. Look at where tasks pile up, who is catching up after hours, and what customers chase most. Backlog is where repetition is, and it surfaces on its own.
Run one for two months
Two months is enough for the run log to speak: how often it ran, how often it was wrong, how often a person stepped in, and where the saved time went. With that, the second choice doesn't need a meeting.
Don't start everything at once
Five or six lines at once means five or six problems at once with nobody able to tell which layer failed — and nothing usable in the first month, which exhausts the budget's patience before the work does.
The inverse of these three is on common mistakes; to start from one specific task, see a single workflow module.
Questions
The four we get asked most about roles
01If we follow this map, what do our people do next?+
This map moves actions, not people. A role has never consisted of one thing — take it apart and what can be handed over is the few actions that repeat, while complaints, exceptions, judgement calls and external commitments are all still there, and those are precisely the parts that consume people. So the step from this map to “how many people can we cut” is a broken inference. What genuinely changes is the composition of everyone's time: most of the day used to go on moving things and waiting, and afterwards most of it goes on judgement and exceptions. Whether that should change your headcount depends on where your volume is going, which isn't a question this system answers — in most cases we've seen, the same people absorbed more work. We also advise against writing “remove N people” into a project's objectives: that objective can't be stated openly, and an objective that can't be stated openly is quietly resisted during rollout. Nobody cooperates with something built to replace them. The reverse works far better — put the tool in people's hands and let them use it on the part of their own job they hate most.
02What if our staff resist it?+
Resistance usually isn't about effort. It's about two fears: will it make decisions for me, and will it be used to measure me. Neither is answered by reassurance — it's answered by how the thing is delivered, and our approach is to put it in people's hands rather than install it above their heads. Delivery includes training aimed at business staff, and what it teaches isn't how to operate a system: it's how to hand over the repetitive parts of their own work — reworking copy in bulk, tidying spreadsheets, drafting replies, aligning terminology and translations. Things they already do daily, where the time saved shows up the same day. Templates, wording and parameters are made changeable by them, tuned to their own habits, without waiting for a development queue; and the first person who gets a result out of it goes and shows everyone else, which beats any amount of top-down push. There's also a fixed rule about sequencing: what goes live first is always the part that saves them time, with monitoring and reporting coming later, and the entry point sits inside a tool they already use so nobody has to log into anything new. Finally, the right-hand columns above should be explained before launch — people who know their half is untouched behave completely differently. Once someone realises this is their tool and it makes their own work lighter, they stop needing to be persuaded and start asking you for more of it.
03Our roles aren't in these nine. What then?+
These nine are simply where the approach is most settled and the delivery boundary can be described precisely — not the limit of what can be built. The test was never the job title, it's the shape of the task: a clear trigger, a process you can describe, a clear output. Manufacturing, wholesale, export trade, engineering, clinics, law firms, training providers — we apply the same test. Where nothing matches a full line, starting with a single module and one specific task works.
04Should we hire someone who understands AI first?+
No, and hiring first usually gets the order backwards. What you actually need is someone who understands the business, can describe the process, and has the authority to change it — someone who can answer “why is this step done this way” and “who approves an exception”. The technical side is delivered by us, with training for your existing technical people at handover; we write the platform layer ourselves but hand over the source, and the foundation is standard tooling — so any engineer familiar with that class of stack can continue.
Next