AI in Action

See what AI could look like in a real business.

Step inside a small convenience store and scroll through four everyday workflows that Claude and automation can simplify - stock and ordering, supplier pricing, customer feedback, and invoices and reporting. Every one of them still ends with a person approving it.

Every workflow below follows the same control

Human Policy Claude Recommendation Human approval Audit trail

Oversight is set by risk, workflow by workflow. On all four below the setting is the strictest one: nothing reaches a supplier, a customer or the ledger until a person approves it.

Workflow 1 of 4: Stock replenishment and supplier ordering

Reordering runs on somebody noticing an empty shelf. Tap a prop to inspect it, and scroll to move it through approval.

Workflow 2 of 4: Supplier pricing and supermarket specials

Comparing pack sizes and promotions by hand takes longer than the saving is worth. Tap a prop to inspect it, and scroll to move it through approval.

Workflow 3 of 4: Customer enquiries and feedback

Four channels, checked whenever somebody has a spare moment. Tap a prop to inspect it, and scroll to move it through approval.

Workflow 4 of 4: Invoices, administration and management reporting

The week is consolidated by hand, and mismatches surface after payment. Tap a prop to inspect it, and scroll to move it through approval.

Corner Store - back office Corner Store - buying desk Corner Store - front of house Corner Store - accounts Today
Low stock

Shelf stock Lines at best price Answered in time Invoices cleared

2 of 6

Order pipeline

1

Human flags it

2

Policy applies

3

Claude compares

4

Recommendation drafted

5

Manager approves

6

Audit trail logged

Buying pipeline

1

Buyer asks

2

Approved suppliers load

3

Claude normalises unit cost

4

Best-value basket drafted

5

Owner approves

6

Decision logged

Response pipeline

1

Feedback arrives

2

Tone and policy load

3

Claude sorts and themes

4

Reply drafted

5

Manager reviews

6

Response and record logged

Invoice pipeline

1

Documents arrive

2

Approval limits load

3

Claude extracts and checks

4

Exceptions flagged

5

Staff clear exceptions

6

Briefing and audit logged

Today

Work piles up on the desks

Staff notice the shelf is low, then check supplier sites and emails and phone around to work out what to order.

The friction

Slow, and easy to miss

Orders get missed and the shelf sits empty longer than it should.

01 · Human flags it

A person starts it

Someone notices the gap and opens the reorder tool. Nothing begins on its own.

02 · Policy applies

Rules load before Claude does

Reorder thresholds and the approved supplier list are pulled in first. Claude works inside them, not around them.

03 · Claude compares

Options, prices, lead times

Claude reads current stock against the approved suppliers and works out what is actually worth ordering.

04 · Recommendation drafted

A draft, not an action

Claude prepares a proposed order and explains its reasoning. Nothing has been sent.

05 · Manager approves

Nothing is ordered without this step

The manager reviews and approves, edits or declines. The gate is the only way through.

06 · Audit trail logged

Who decided what, and when

The order, the decision and the person behind it are all recorded for review.

Outcome

The shelf refills sooner, and every order has a name against it

Staff stop chasing suppliers by phone, the manager still decides, and the business gets a record it can review.

And here is what we would measure
Time from low stock to order placed
Timestamps already in the workflow - no extra tracking.
Out-of-stock days per line
The thing the customer actually notices.
Recommendations approved without edits
Whether the recommendations are worth trusting.
Exceptions escalated to a person
Confirms the human checkpoint is real, not decorative.

We deliberately show no percentages here. The baseline is measured in your business during discovery, before anything is built - a number invented for a demo is not evidence.

Today

Prices checked one site at a time

The owner works through each wholesaler and the supermarket specials by hand, then tries to compare pack sizes that do not match.

The friction

The comparison rarely gets finished

Weighing a 24-pack against a 30-pack against a two-for-one takes longer than the saving is worth, so it often does not happen at all.

01 · Buyer asks

A person decides to re-price

Someone decides it is time to look at a category again. Nothing runs to its own schedule.

02 · Approved suppliers load

Only suppliers already on the list

Scope is limited to suppliers the business has already approved, on their agreed terms.

03 · Claude normalises unit cost

Pack sizes made comparable

Promotions, pack sizes and delivery are reduced to one effective cost per unit, so the options can actually be ranked.

04 · Best-value basket drafted

A proposal, with the working shown

Claude proposes where to buy each line and sets out how it got there. Nothing has been ordered.

05 · Owner approves

Nothing is switched without this step

The owner sees the comparison and accepts, adjusts or rejects it. Supplier relationships are a judgement call, not a calculation.

06 · Decision logged

The reasoning outlives the decision

What was compared, what was chosen and who chose it are recorded.

Outcome

Buying moves to the best available price, and you can show why

The comparison happens every cycle instead of when there is time, and the reasoning survives for the next one.

And here is what we would measure
Effective unit cost against list
The saving, measured the way the buyer actually buys.
Lines bought at the best available price
How often the better option is the one taken.
Time spent comparing suppliers
The hours this was quietly costing.
Recommendations changed before approval
Whether the owner's judgement is still doing work.

We deliberately show no percentages here. The baseline is measured in your business during discovery, before anything is built - a number invented for a demo is not evidence.

Today

Four channels, no single queue

Google reviews, emails, website enquiries and social comments are checked whenever somebody has a spare moment.

The friction

Some get answered, some do not

A few are handled the same day and others not at all - and nobody can see the pattern running across them.

01 · Feedback arrives

Everything lands in one place

Messages from all four channels come into a single queue. A person still owns it.

02 · Tone and policy load

What may be said, and what must be escalated

House tone, what can be promised, and the rules for sensitive issues are applied before anything is drafted.

03 · Claude sorts and themes

Topic, sentiment, and what keeps recurring

Each message is classified and grouped, so a repeated complaint stops looking like five unrelated ones.

04 · Reply drafted

Suggested, not sent

A response is prepared for each message. None of them have gone anywhere.

05 · Manager reviews

Nothing is sent without this step

Sensitive complaints are escalated to a person rather than answered automatically. That routing is the point, not a fallback.

06 · Response and record logged

The reply and the reviewer are recorded

What was sent, who approved it and which theme it belonged to are all kept.

Outcome

Customers hear back faster, and the recurring problems become visible

The queue stops depending on who has a spare moment, and the themes worth fixing surface on their own.

And here is what we would measure
Time to first response
What the customer experiences, per channel.
Replies sent without edits
Whether the drafts are good enough to trust.
Complaints escalated to a person
Confirms the sensitive cases are still reaching a human.
Recurring themes identified
The operational value, separate from the speed.

We deliberately show no percentages here. The baseline is measured in your business during discovery, before anything is built - a number invented for a demo is not evidence.

Today

The week is assembled by hand

Invoices, emails, spreadsheets and store reports are consolidated manually at the end of the week.

The friction

Late, and expensive when it is wrong

It happens after the fact, and a mismatch is usually found once the invoice has already been paid.

01 · Documents arrive

Everything lands in one place

Invoices and store reports come in from suppliers and from each site, into a single intake.

02 · Approval limits load

Coding rules and who may approve what

Tolerances, cost coding and approval limits are applied before anything is read.

03 · Claude extracts and checks

Read, then reconciled

Line items are pulled off each document and checked against the order and the delivery.

04 · Exceptions flagged

Anything that does not reconcile stops

Mismatches are separated out rather than pushed through. Stopping is the useful behaviour here.

05 · Staff clear exceptions

Nothing is paid without this step

A person resolves each exception, and the clean ones still need sign-off inside the approval limits.

06 · Briefing and audit logged

One clean summary, fully reviewable

The manager gets a single weekly briefing, and every extraction and decision behind it stays inspectable.

Outcome

The week closes sooner, and problems surface before the money moves

Staff stop re-typing and start reviewing exceptions - which is the part that actually needed a person.

And here is what we would measure
Invoices processed without a touch
How much of the volume never needed a person.
Exception rate
Whether the checks are catching real problems or noise.
Time to weekly close
When the manager can actually see the week.
Value of mismatches caught before payment
The number this has to beat to be worth doing.

We deliberately show no percentages here. The baseline is measured in your business during discovery, before anything is built - a number invented for a demo is not evidence.

All four workflows run on the same pattern: a person starts it, policy loads before Claude does, Claude prepares a recommendation, a person approves, and the decision is logged. The scene and the steps come from one data structure, so another industry is a new entry in it rather than a rebuild.

Two services. One partner. One governance backbone.

This is what a governed AI workflow looks like end to end. If a process like this exists in your business, we can help you design it properly - with the same human approval and audit trail, built in from the start.