Skip to content

Keep consequential decisions with a named person

Human-in-the-loop does not mean a person reads every message. The permission differs by route. This guide sets out who approves what, what can run inside approved rules, what stays a draft and which cases always reach a person. It promises no outcome and does not replace your own legal or professional advice.

Deterministic rules versus AI

Decide which steps a rule can settle before you decide where AI helps. AI earns its place when a rule cannot do the information work.

  1. 01

    Let rules check the case first

    Identity, required fields, duplicates, exclusions and suppression can all be tested by rules. They run before AI summarises context or drafts an action, so the AI works from a case that has already passed those checks.

  2. 02

    Give AI a bounded information task

    AI can read, match, summarise, draft, classify and prioritise from the material it is given, using criteria you define. Name the task, for example "draft a reply from this approved answer", and not "handle the customer".

  3. 03

    Keep some decisions out of its reach

    An AI step should not invent buying intent, contact permission or urgency, and should not reach a compliance conclusion. Clinical, legal, regulatory, safety, credit, pricing and other professional judgements stay with a named person.

  4. 04

    Make uncertainty stop the case

    Conflicting records, low confidence, a suppressed contact, a dispute or a decision that needs commercial or professional judgement should route to a named person. The system should not guess.

  5. 05

    Check whether a native feature already does it

    Your booking, CRM or quoting system may already offer the rule you need. Replacing a reliable native workflow only to add an AI component adds risk without adding control.

Campaign-level approval

On the routes that contact people outside your business, approval attaches to the rules of a campaign. It is given before anything is sent.

  1. 01

    Approve the audience, route and message together

    For B2B lead research, the approval covers the audience, the contact source, the channel, the message, the frequency and the hand-off route, with exclusions and suppressions applied first. For quote and account recovery, it covers the record group, the contact route, the message, the frequency and the escalation rules. Both are agreed before activation.

  2. 02

    Own the basis for contact

    The workflow is not there to decide whom you may lawfully contact. The basis for each contact list is recorded with you before a message is sent. Take your own advice on that basis.

  3. 03

    Start with a small group

    The first test runs on a small, agreed group of accounts or records, with no database-wide send. For a quote recovery sprint, the published description is draft actions with human approval.

  4. 04

    Widen only on evidence

    Move from the first group to a recurring campaign only if replies, hand-off speed and the quality of the matches justify it. Keep pause and exclusion controls in place when it does.

  5. 05

    Name an owner for each campaign

    Someone on your side holds the approval, can pause the campaign and takes the replies.

Actions permitted without per-message review

This is the core of the matrix. Each route allows something different, and the permissions are not interchangeable.

  1. 01

    Enquiry handling

    Routine questions are answered from a knowledge base the business approves. A missed call can trigger an approved text, in reply to a call the customer made. Routine, low-risk sends such as a scheduled reminder go out as a template would if a person sent it by hand. Website and messaging answers use your existing booking link and have no write access to the booking platform. Other booking writes need a verified scope and the platform's own API permissions.

  2. 02

    B2B lead research

    Inside an approved campaign, the workflow can send the tailored opening, run the agreed follow-ups, classify replies and route the qualified ones. It is not draft-only by default. Low-confidence entity matches stop for human review.

  3. 03

    Quote and account recovery

    After activation, the workflow can send the agreed message, follow up within the set frequency, apply stop rules and classify replies. Approved recovery is not draft-only by default either, and it stops when asked, when uncertain or at a hand-off point.

  4. 04

    Evidence preparation

    Nothing in this route is sent to anyone on its own. It reads authorised records and prepares an index, exceptions and drafts for review. The draft-only evidence section below covers it.

  5. 05

    Why not review every routine message

    Reading every routine message can turn into clicking approve without reading. Spend reviewer attention on the first small group, the exceptions, a sample of routine cases, and unanswered, escalated and corrected cases.

Sensitive exceptions

Write the stop list before launch. It should read as plain cases, not as a confidence score.

  1. 01

    Clinical, urgent and uncertain enquiries

    A clinical, urgent or uncertain question goes to a named person. Out of hours, that means an immediate scripted handover and not an automated answer overnight, and no triage. This boundary is not something to switch off on request.

  2. 02

    Complaints, disputes and negative feedback

    These are not handled automatically. A negative or ambiguous review signal goes to a named manager, and the system is not given permission to publish or respond to it.

  3. 03

    Positive and complex replies in B2B work

    A positive, sensitive, disputed or complex reply goes to the named salesperson, who also owns pricing, negotiation and the purchase decision.

  4. 04

    Objections and relationship context in recovery

    Objections, offers, commercial terms and any sensitive situation stay with the account owner. A quote record shows that a quote exists. It does not show why the customer has not replied.

  5. 05

    Hand over the context with the case

    The person receiving a case should get the source, the message history and the suggested next action, so that the customer is not asked to repeat themselves.

Draft-only evidence

Evidence preparation is the route that starts read-only or draft-only. That applies to this route, not to the others.

  1. 01

    The pack prepares, the professional decides

    For a monthly operator review, the workflow brings approved records into one source-linked pack, flags what is missing or conflicting and prepares questions. The Transport Manager keeps the compliance judgement, the investigation, the corrective action and the sign-off.

  2. 02

    Requests stay drafts until approved

    A prepared request for missing evidence remains a draft until a named person approves it. The pack lists an exception. It does not declare a vehicle or an operator compliant or non-compliant.

  3. 03

    Some actions are not available at all

    The route does not take roadworthiness, safety, disciplinary or regulatory decisions, and it does not submit anything to a regulator. It also does not act as the Transport Manager.

  4. 04

    Check every match during a pilot

    In a draft-only pilot, the reviewer checks every match, exception and proposed request against the source documents. That is how you learn how often the pack is wrong.

  5. 05

    Widen after you have the numbers

    Add operators or approved task creation only once you understand the correction rate, the completeness of the packs and the time the reviewer spends. Expansion is a decision taken on the evidence, not a scheduled step.

Audit trail

A record is what lets a person check afterwards what ran, why and who decided. Ask for it in the design, not after something goes wrong.

  1. 01

    Record why a case appeared

    Keep the original source and the material fields behind each case, which checks passed and what remains uncertain. A reviewer cannot judge a case whose reason has been lost.

  2. 02

    Record who decided and what happened

    Log the human correction, the decision, the owner and the outcome against the case, as well as what the system did. A record of messages alone does not show who approved the campaign.

  3. 03

    Keep the action and escalation log where the team can see it

    Put the log where the team works, and look at unanswered, escalated and corrected cases on a regular schedule.

  4. 04

    Review the approvals as well as the cases

    Note when the campaign rules were approved, by whom and what changed afterwards. Rules change, and the record should show which version applied to which message.

  5. 05

    Check our factual position on the security page

    The security page carries the current status of Durward's data-handling and logging controls. Check it there. This guide does not restate it, and it is not a statement that any workflow is compliant.

Failure and recovery

An automation can stop working. Decide before launch what the business does when it does.

  1. 01

    Agree the manual route first

    For each workflow, write down how the work continues by hand if the automation is down, and who does it. Enquiries should be preserved and not lost while the automation is stopped.

  2. 02

    Test the failure paths before go-live

    Use scripted and edge cases, including a complaint, urgent wording, uncertainty and an unavailable system. Each channel should have its own tested route to a person before launch.

  3. 03

    Keep a pause control and use it

    Someone must be able to pause a campaign or an exclusion quickly. Agree who may do so and what they do next.

  4. 04

    Make the AI step recoverable

    When an AI output is unusable, the workflow should remain understandable and have a route back to a person. Make it easy for a person to correct, pause or reroute a case.

  5. 05

    Ask what happens to your setup if you leave

    Ask where the approval rules, message templates and logs are held, and what you keep if you switch the automation off or change supplier.

Eight checks before you approve a route

Answer these in writing for each route. The downloadable matrix has a row for each step, with columns for the default permission, the approval level, who receives the case and an owner.

  1. 1.Is each step written down as runs inside approved rules, draft only or a person decides?
  2. 2.Does the approval name the audience or record group, the contact route, the message, the frequency and the hand-off?
  3. 3.Is the stop list for clinical, urgent, disputed and uncertain cases written before launch?
  4. 4.Is a route described as draft-only only if nothing in it is sent without a person?
  5. 5.Does every handover name a person and carry the source, the history and the next action?
  6. 6.Does the record show why a case appeared, who approved the campaign and who decided each case?
  7. 7.Is the manual fallback written down and tested, with a named person and a pause control?
  8. 8.Do you know what you keep if you switch the automation off?

Download the permission matrix (CSV) A plain spreadsheet file with one row for each step across the four routes. The last two columns are for your own decision and owner. Edit it freely; it opens in any spreadsheet program.

Bring one route and its stop list

Tell us the workflow, who approves it and which cases must reach a person. We will say plainly if a feature you already pay for covers it, or if the work should stay manual.