AI Opportunity Map · October 2026

Harlowmere Fixings Ltd

The 26 recurring tasks we mapped take about 5,200 hours a year, costing £128k. On cautious starting assumptions, AI could release £23k–£45k of that time each year. Three pilots come first, worth £7k–£15k a year between them, and each pays back within 5 months of released time in the low case. The pilots will measure the real figures.

Worked example. Harlowmere Fixings is a fictional business. Its tasks, times, salaries and costs are illustrative and show how the method works. They are not benchmarks.
1 · The matrix

Where AI is worth it, and where it isn't yet

Pilot nowPlanSimplify firstControls firstBundleNot now
Value against ease for 26 tasks. The full list is in the table below the chart.QUICK WINSBIG BETSFILL-INSNOT NOWT14T01T15T02T11T22T03T25T24T23T21T2600252550507575100100Ease score →Value score →
Next stepTasksSaving a year
Pilot now3£7k–£15k
Plan2£4,000–£7,600
Simplify first4£2,800–£5,800
Controls first3£800–£1,400
Bundle7£4,400–£8,300
Not now7£3,300–£7,300

Ranges run from the low to the high case, are rounded, and rest on assumed AI shares. Pilot now means valuable, easy, and paying back within 12 months in the low case. 9 tasks sit within five points of a threshold (T03, T10, T15, T17, T20, T21, T23, T25, T26), so treat their placement as provisional.

By tool

ToolTasksPayback
AI assistant for the office team (8 users)93 months
Document extraction into the ERP218 months
Customer service assistant connected to the ERP1No payback
Reorder analysis add-on for the ERP150 months
All 26 tasks as a table
IDTaskTeamHours a yearSaving a yearValueEaseNext step
T01Turn emailed enquiries into quotationsSales960£3,500–£7,0008065Pilot now
T02Key customer PDF orders into the ERPSales384£2,300–£4,4007965Plan
T03Re-key web shop orders into the ERPSales180£1,000–£2,2004840Simplify first
T04Answer order status and stock questionsCustomer service256£900–£2,2003765Not now
T05Chase overdue invoicesAccounts80£500–£9002285Bundle
T06Process supplier invoices and match to ordersAccounts168£1,100–£2,1003875Bundle
T07Month-end management report packManagement96£400–£8002040Not now
T08Weekly sales report for the MDManagement72£600–£1,3002780Bundle
T09Write product descriptions for the web shopMarketing200£900–£1,8002380Bundle
T10Load supplier price list changesPurchasing288£1,100–£2,1003845Not now
T11Weekly reorder analysisPurchasing384£1,600–£3,2006240Plan
T12Research alternative suppliersPurchasing72£200–£500665Bundle
T13Meeting notes and actionsManagement64£900–£1,5002195Bundle
T14MD inbox triage and repliesManagement240£2,700–£5,3009075Pilot now
T15Customer supplier questionnaires and tendersSales144£1,300–£2,6005460Pilot now
T16Job adverts and interview packsPeople48£200–£300460Bundle
T17Update risk assessments and method statementsOperations72£200–£5001645Not now
T18Paper picking from printed pick listsWarehouse640£0–£300210Not now
T19Delivery scheduling and driver routesWarehouse264£200–£7001830Not now
T20Returns and credit notesCustomer service120£400–£8002145Not now
T21Credit checks on new accountsAccounts48£200–£4001545Controls first
T22Staff policy and handbook questions to the MDPeople45£400–£1,0003375Simplify first
T23Complaint investigation reportsCustomer service72£500–£8002145Controls first
T24Stock count variance investigationsWarehouse96£400–£8002060Simplify first
T25Print, file and retrieve paper delivery notesWarehouse198£1,000–£1,8003550Simplify first
T26Sickness absence and return-to-work recordsPeople36£100–£3001345Controls first
2 · Your three pilots

Prove it before you rely on it

Pilot 1

MD inbox triage and replies

£2,700–£5,300 a year

Pays back in 1 month of released time (low case) · Management · 4-week pilot

Baseline
Log inbox time for two weeks: emails received and minutes spent
Success
Inbox time down by a third, with no missed or wrong replies
Owner
Managing Director
Human review
AI sorts and drafts; the MD reads every reply before it is sent
Go-ahead
Go
Pilot 2

Turn emailed enquiries into quotations

£3,500–£7,000 a year

Pays back in 3 months of released time (low case) · Sales · 4-week pilot

Baseline
Time 20 quotations from enquiry to sent, and count revisions
Success
Quotation time halved with no rise in revisions or errors
Owner
Internal Sales lead
Human review
Prices come from the ERP, never from the AI; every quotation checked by the sender
Go-ahead
Go
Pilot 3

Customer supplier questionnaires and tenders

£1,300–£2,600 a year

Pays back in 4 months of released time (low case) · Sales · 8-week pilot

Baseline
Time the last three questionnaires from receipt to submission
Success
Next three completed in half the time, with no answer the MD would change
Owner
Managing Director
Human review
Answers come only from an approved answer library; the MD signs off every submission
Go-ahead
Go with conditions
3 · Simplify first

Don't automate the waste

These 4 tasks exist because something upstream isn't working. Fix the cause and most of the work disappears, with no AI needed.

4 · Decisions for you

What to decide this month

  1. Approve the three pilots, their owners and their human review rules. Measure the baselines before anything starts.
  2. Confirm one office AI assistant for the pilots and the 6 smaller "Bundle" tasks that use it, so the team learns one tool rather than several. Get a current, dated price before committing.
  3. Drop the customer service assistant for now: in the low case it costs more each month than the time it saves.
  4. Revisit the Plan tasks once the Simplify first fixes are done. The document extraction tool is costed against only two tasks today; its payback improves as more tasks share it.
  5. Agree the Simplify first fixes, starting with connecting the web shop to the ERP.
  6. Take data protection advice before using AI on complaint investigation reports, credit checks on new accounts and sickness absence and return-to-work records.
  7. Decide what the released time is for. Without that decision it simply disappears into the day.
The months after

What the agent did, and what people decided

Fictional and illustrative. This continues the worked example from the interviews in September 2026 to January 2027. Every name, figure and event is invented.

Harlowmere set up the agent before the interviews, in its own accounts: its AI assistant, a shared drafts document, and a restricted folder in its own storage for the interview notes. The Managing Director approved and sent every invitation and reminder; the consultant held the interviews and checked every task row the agent drafted. The people who do the work gave every time, and the MD and the consultant chose every pattern, answer and score in the working session. No interview was recorded. The map was handed over on Tue 20 October 2026. After handover the agent reminded the pilot owners to measure their baselines before the pilots started on 2 November 2026, and prepared the re-run of the map from 11 January 2027, 14 days after the longest pilot ended. At the 30-day check-in on Thu 19 November 2026, the MD and the consultant read the interview log and the re-run plan together; the consultant's upkeep of the set-up ended there, and Harlowmere has looked after the agent itself since. Nothing left the business, and no decision changed, without one of Harlowmere's people approving it.

People interviewed
12 of 13
across all 9 teams; one person was observed instead
Task rows drafted
27
26 on the register after the interviewer's check
Times from a time log
3
3 tasks rest on the owner's estimate
Re-run prepared
11 Jan
for the MD to decide on the 3 pilots' measured shares
DateWhat the agent prepared, collated, coded or flaggedKindWho decidedWhat happened
18 Sep 2026Interview invitations drafted for 13 people across the 9 teams, by role: 20 to 30 minutes each, what it is for (the recurring work, not anyone's performance) and a request to bring a rough list of last week's work
Roles sheet; the interview guide
DraftManaging Director
Approved with changes
The MD added one line on why (time back for the team, said only because it is true) and sent all 13 on Mon 21 Sep.
24 Sep 2026One reminder drafted for the 2 people with no reply after 3 working days
Replies logged against the invitations; Inputs (3 working days)
DraftManaging Director
Approved
Sent on 24 Sep. One booked; the warehouse operative preferred not to be interviewed, so picking was observed instead.
9 Oct 202612 interviews held 29 Sep to 9 Oct, none recorded: the consultant's notes filed in the restricted folder within the day; read-back of each person's task list drafted for the consultant to send
Interview log; Inputs (recorded: No)
LoggedThe consultant
Decided
Every read-back sent within a working day; two people corrected a time.
9 Oct 202627 task rows drafted from the notes: task, team, role, people, times a month, minutes and where they came from, with each person's own answers to the five task questions quoted beside the row. No AI pattern, Lean view, returns, tool or score
The consultant's notes; observation notes for 5 tasks
CodedThe consultant
Approved with changes
The consultant merged two rows into one (chasing overdue invoices was raised by two people), leaving 26, and chose every pattern and answer in the working session with the MD.
9 Oct 2026Three tasks with rough times: time-log requests drafted for the people to time the next three occurrences (order status questions, supplier invoices, returns and credit notes)
Interview notes (times given as a range)
DraftThe consultant
Approved
Timings back by 16 Oct; the three rows now show Time log as their source.
12 Oct 2026Three tasks raised only by the MD (product descriptions, supplier research, job adverts) rest on the owner's estimate: 3 of 26, under the 30% limit
Task register (where the data came from)
FlagThe consultant
Decided
Kept as Owner estimate and named in the map's confidence note.
12 Oct 2026Sickness absence records hold health data (special category): Data involved left for the consultant to set; notes describe the data by type and role only
Interview notes
FlagManaging Director
Decided
Set to Special category: Controls first, with credit checks and complaint reports. Advice before any AI use.
26 Oct 2026Reminders drafted for the two pilot owners to measure their three baselines before the pilots start on 2 Nov
Pilots sheet (baseline, owner); Inputs (start date)
DraftManaging Director
Approved
Sent on 26 Oct. The MD's inbox log and the questionnaire timings were ready on 2 Nov.
2 Nov 2026Quotation baseline not complete at the start: 12 of the 20 quotations timed
Internal Sales lead's baseline sheet
FlagManaging Director
Decided
The quotation pilot started a week later, on 9 Nov, once all 20 were timed. The other two started on 2 Nov.
11 Jan 2027Re-run prepared: each pilot's measured time beside its baseline and the assumed pattern share, in a working copy, for the MD to decide whether to replace the shares on Patterns
The pilot owners' logs; Pilots sheet; Inputs (start date, 14 days)
DraftManaging Director
Pending

The agent never filled in a time or chose an answer: three rough times were timed by the people who do the work, and the three tasks only the MD raised stayed as owner estimates, named in the confidence note. The sickness absence records stayed at Controls first, with the notes describing the data by type and role only. The re-run is prepared but not yet decided, so the savings on this page are still the map's assumed ranges, not results. The consultant spent about 20 minutes a working day on the drafts while the interviews ran, and the MD about 10 minutes a week since (estimate).

5 · How confident are we?

The weakest inputs