Skip to content

Examples

What a process looks like once it is automated.

These are illustrative scenarios rather than delivered projects — worked examples of how a workflow of each shape could be designed. For work we have actually built, see the case studies. Each one shows how the work runs today, what changes, what gets built, and which parts stay with a person.

Illustrative example See work we have delivered

Start here

What does your team still do manually?

Pick the part of the business it sits in. Each example shows the work as it runs today, the same work automated, what gets built and which parts stay with a person.

Showing all 12 examples.

Supplier invoices into the finance system

Finance

Today

Invoices arrive by email in a dozen layouts. Someone opens each one, reads the totals, keys them into the finance system and files the PDF.

Automated

The workflow reads each invoice, extracts supplier, line items, VAT and totals, checks them against the purchase order, and creates the record.

What gets built
Document extraction with per-field confidence, matched against existing purchase orders through the finance system API.
What stays human
Anything that does not match its purchase order, or where a field came back with low confidence, goes to a person.

Inbound enquiries sorted and answered

Customer service

Today

A shared inbox fills up. Someone reads every message, works out what it is, answers the easy ones and forwards the rest.

Automated

Each message is classified on arrival, routine replies are drafted and sent, and anything unusual is routed to the right person with a summary.

What gets built
Intent classification, entity extraction, a reply generator working from your own approved wording, and routing rules.
What stays human
Complaints, anything with a legal or safety flavour, and any message the classifier was not confident about.

Quotes generated from an enquiry

Sales

Today

Staff read the enquiry, chase the missing details, look up pricing, build the quote in a spreadsheet and email it over.

Automated

Details are extracted, missing ones requested automatically, pricing rules applied, and the quote produced and sent.

What gets built
A pricing rules engine, an automated follow-up sequence for missing information, and document generation.
What stays human
Quotes outside the normal range, or for non-standard work, are held for approval before they go out.

New starter onboarding

People

Today

HR works through a checklist by hand: accounts, equipment, policy documents, training bookings, reminders to managers.

Automated

Accepting an offer triggers the sequence. Accounts are requested, documents sent for signature, training booked, and progress tracked.

What gets built
Workflow orchestration across the HR system, IT service desk and document signing, with a status view for the manager.
What stays human
Right-to-work checks and anything involving a document that looks wrong go to a person.

Assessment and feedback preparation

Education

Today

Educators read each submission, apply the rubric, draft feedback, record the mark and return it — the same shape of work every time.

Automated

Draft rubric-aligned feedback is prepared before the educator opens the submission. They review, adjust and approve.

What gets built
Rubric-aware analysis of submissions with the reasoning shown next to each criterion.
What stays human
The educator makes every academic judgement and nothing reaches a student without their approval.

Weekly operational reporting

Operations

Today

Someone exports from three systems every Monday, pastes into a spreadsheet, fixes the formatting and writes a summary.

Automated

Data is pulled automatically, reconciled, and the report is produced with a written summary of what changed and what needs attention.

What gets built
Scheduled integrations with each source system, reconciliation logic, and generated commentary on the movements.
What stays human
The summary is a draft. Whoever owns the report reads it before it circulates.

Chasing missing information

Admin

Today

A case stalls because a form is incomplete. Someone eventually notices, emails the customer, and waits. Sometimes nobody notices.

Automated

The system detects exactly what is missing, asks for those specific items, follows up on a schedule, and restarts the case when they arrive.

What gets built
Validation rules per case type, a follow-up sequence with sensible limits, and automatic resumption.
What stays human
After a set number of attempts the case is escalated rather than chased indefinitely.

Policy and procedure questions

People

Today

Staff ask HR or a manager. The answer is in a document somewhere, but finding it takes longer than asking.

Automated

Staff ask in plain English and get an answer with a citation and a link to the section it came from.

What gets built
Retrieval over your own policy library, answering only from those documents and saying so when the answer is not there.
What stays human
Questions touching discipline, grievance or anything personal go to a human, not an assistant.

Audit evidence gathering

Compliance

Today

Preparing for an audit means weeks of collecting records from multiple systems and cross-referencing them against requirements.

Automated

Evidence is collected continuously against each requirement, with gaps flagged as they appear rather than discovered at audit time.

What gets built
Scheduled collection from source systems, mapping to the control framework, and a running gap report.
What stays human
Every judgement about whether evidence is sufficient belongs to the compliance team.

Shipment status updates to customers

Operations

Today

Customers email to ask where their order is. Staff check two systems and write a reply, several times per order.

Automated

Status changes trigger proactive updates, so customers are told before they ask.

What gets built
Integration with carrier and internal tracking, event-driven notifications, and a customer-facing status view.
What stays human
Delays and exceptions are written by a person, because those messages need judgement.

Meeting notes into CRM actions

Sales

Today

Notes are written up after the call, if there is time. Actions get remembered, or they do not, and the CRM drifts out of date.

Automated

The call is summarised, actions extracted with owners and dates, and the CRM record updated.

What gets built
Transcript summarisation, action extraction, and a write-back into the CRM through its API.
What stays human
The owner reviews the summary before it is saved, so nothing incorrect enters the record.

Contract review support

Compliance

Today

Every incoming contract is read end to end to find the clauses that differ from the standard position.

Automated

Incoming contracts are compared against your standard terms, with differences and missing clauses listed up front.

What gets built
Clause-level comparison against your template library, with each finding linked to the passage it came from.
What stays human
This shortens the read, it does not replace it. Every contract is still reviewed and signed by a person.

Your process is not on this list

It rarely is. These are patterns rather than products — most projects start with a process nobody has written down. If yours involves reading something, checking it, moving it between systems or chasing someone, it is worth a conversation.

Describe your process

In depth

Two use cases, end to end.

Two processes set out in full, including the objection a careful buyer would raise. The second decides who gets an interview, so it is deliberately the one with the most said about fairness and human judgement.

Internal IT support

Support requests over WhatsApp

An employee messages the company WhatsApp number instead of filling in a ticket form. The assistant asks what is wrong, gets the details an engineer will need, raises the ticket, and assigns it to whoever is on for that kind of problem.

  • WhatsApp
  • Microsoft Teams
  • Email
  • Service desk portal

How it works today

People message a colleague directly, or catch someone in the corridor, because the ticket form asks for a category they cannot pick and a priority they cannot judge. Requests arrive in a dozen places, nobody knows the real backlog, and the person who fixes things fastest quietly becomes everyone’s first port of call.

Step by step

  1. The employee just describes it A person

    They send a message in their own words — "laptop won’t connect to the VPN since this morning". No form, no category, no priority to choose.

  2. The assistant asks what is missing

    It asks the two or three things an engineer will need and the employee has not said — which device, when it started, whether anyone else is affected, whether there is a deadline it blocks. A photo of the error is welcome.

  3. Classified and prioritised

    The request is categorised — access, hardware, software, network, or something else — and given a priority from the impact described, not from what the employee guessed.

  4. Ticket raised

    A ticket is created in your service desk with the full conversation attached, so the engineer opens it already knowing what was asked and answered.

  5. Assigned to the right person

    Routing uses the category, the current rota and who is available, then notifies them where they already work — Teams, Slack or email — rather than waiting for someone to check a queue.

  6. The engineer takes it from here A person

    A person diagnoses and fixes the problem. The workflow has done the intake, not the support.

  7. The employee is kept informed

    They get the ticket reference immediately, then updates on the same WhatsApp thread as it progresses, without having to chase or log in anywhere.

What gets built

  • A WhatsApp Business number connected to the workflow, with Teams and email as alternative front doors
  • Conversational intake that asks only for what is missing, in the employee’s own language
  • Classification and priority derived from the described impact, with your own rules applied
  • Ticket creation in your existing service desk, with the whole conversation attached
  • Assignment from the on-call rota and skill routing, with notification where the team already works
  • Status updates back to the original thread, and escalation when a ticket goes unanswered

What stays human

  • The engineer diagnoses and resolves. The assistant handles intake and routing, and nothing else.
  • It never resets a password, grants access or changes a permission — those go to a person every time.
  • Anything with a security flavour, or an employee who says it is urgent, is escalated rather than triaged.
  • If the assistant cannot make sense of a request after a couple of exchanges, it hands over instead of guessing.

On the obvious risk: social engineering

A support channel that can be messaged from any phone is a target — "hi, it’s the new finance manager, I need my MFA reset". So the assistant has no ability to act on an account. It cannot reset credentials, grant access or change permissions; those requests are routed to a person with the identity check your policy requires. It also only recognises numbers belonging to known employees, and an unrecognised number is flagged rather than served.

Recruitment

From application to shortlist

A role opens and applications arrive — by email, through the careers page, from an agency. Each CV is read against the criteria set for that role, and the hiring manager opens a shortlist with the reasoning already attached to every candidate.

  • Careers page
  • Applications inbox
  • Agency submissions
  • Job boards

How it works today

Someone opens every application, reads the CV, works out whether the experience matches, keeps a spreadsheet of who is worth a call, and writes back to everyone else. A popular role attracts a few hundred applications, so screening happens in the evenings and the good candidates have taken another offer by the time anyone replies.

Step by step

  1. Criteria agreed first A person

    Before the role goes live, the requirements are written down as criteria the workflow can be measured against — must-haves, nice-to-haves, and anything that disqualifies. Fixing this up front is what makes every application judged on the same basis.

  2. Application captured

    The application is recognised wherever it arrives, and the CV, covering note and any form answers are pulled together into one record for that candidate.

  3. CV read and structured

    Contact details, roles, dates, employers, skills and qualifications are extracted from whatever layout the CV happens to use, with a confidence score on each field.

  4. Matched against the criteria

    Each criterion is assessed against the CV, with the passage it came from shown next to it — so a rating is always traceable to something the candidate actually wrote.

  5. Shortlist prepared

    Candidates are ordered by how well they meet the criteria and presented as a ranked list, with the strongest evidence and any gaps summarised for each one.

  6. Hiring manager decides A person

    The manager works down the shortlist, agrees or overrides the assessment, and chooses who to progress. Nobody is rejected by the workflow.

  7. Candidates hear back

    Approved outcomes trigger the emails — invitations to interview with booking links, requests for missing information, and decline messages for everyone else, so no application goes unanswered.

What gets built

  • Capture from the careers page, the applications inbox and agency submissions into one candidate record
  • CV parsing that copes with real-world layouts, scoring each extracted field for confidence
  • Criteria-based assessment with the supporting passage shown against every rating
  • A ranked shortlist view the hiring manager works through and overrides freely
  • Templated candidate emails, interview booking, and write-back to your ATS or HR system
  • A record of the criteria, the assessment and the decision for every applicant

What stays human

  • Every decision to progress or decline a candidate is made by a person, not the workflow.
  • The assessment is a starting point for the manager to disagree with, and overrides are expected rather than exceptional.
  • Low-confidence extractions and unusual CVs are flagged for a proper read instead of being scored anyway.
  • Interviews, offers and any conversation with a candidate stay entirely with your team.

On fairness, and why nothing is auto-rejected

Screening decides who gets a chance, so the design has to be defensible. Criteria are fixed before applications arrive rather than adjusted as they come in; each rating is traceable to a passage in the CV; the workflow ranks and evidences, but never rejects; and every assessment and decision is recorded. Under UK GDPR a candidate can ask how their application was handled — this shape means you can answer with the criteria used, the evidence found and the name of the person who decided.

Bring us your process.

In a free 30-minute discovery call we will look at how it works today, where the manual effort sits, what is worth automating, what is not, and where people should stay in the loop. No technical brief, no budget, no obligation.

Book a free process discovery call See how AI automation works