← All service demonstrations

Which step is worth automating first?

A fictional B2B services team receives enquiries by form and email, copies them into a CRM, checks fit, drafts replies and tracks follow-up by hand. Here is the kind of decision I would return before building anything.

Fictional exampleNo client records, measured savings, production integration or legal conclusion.

See the handoffs before choosing a tool.

The buyer would verify these steps in a walkthrough. The diagram is a starting hypothesis, not a claim about a real company.

  1. 01Enquiry arrivesForm or shared mailbox
  2. 02Copy into CRMFields and ownership checked manually
  3. 03Assess fitHuman interprets the request
  4. 04Reply and follow upHuman sends and records outcome

Questions first: How many enquiries arrive, what data may be used, which CRM owns the record, who approves a reply, and how often do handoffs actually fail?

Three opportunities. One sensible start.

Open each option to inspect its authority boundary and the evidence that would make it useful. The ranking is illustrative; real priorities depend on the buyer's volume, cost and data constraints.

01Validate intake and identify duplicatesFirst pilot candidate

A narrow rules-first change that can be checked before AI or a CRM write is introduced.

Authority
Stage a normalized record; route ambiguous merges to a human.
Check
Field completeness, duplicate-review accuracy, handling time and visible errors.
02Suggest a fit category and reply draftLater option

Useful only after source fields are reliable and reviewers can see what the suggestion rests on.

Authority
AI may propose and cite; a person approves every external reply.
Check
Reviewer agreement, unsupported claims, edits needed and stale-source rate.
03Track follow-up ownership and outcomesLater option

A reminder is helpful once record ownership and outcome states are agreed.

Authority
Create an internal reminder; no automatic customer message in the first pilot.
Check
Missing owners, overdue tasks and completeness of recorded outcomes.

Prove one path without sending a message.

With the buyer's permission, start with 10–20 representative, privacy-safe examples: clear fit, poor fit, duplicate, missing data, uncertain fit, stale data and a provider error. Redacted or synthetic records come first.

Acceptance path

  1. Normalize and validate each input.
  2. Identify duplicates and missing fields.
  3. Propose a fit decision with its source and uncertainty.
  4. Place any reply draft in a human review queue.
  5. Record approval, edit or rejection; send nothing in this pilot.

Stop if source rights, consent, field quality or decision ownership is unclear.

A decision pack, not a vague promise.

01Current-workflow map

Verified steps, owners and handoffs.

02Opportunity scorecard

Three options with assumptions and trade-offs.

03Authority and failure notes

Tool/data rights, approvals and stop conditions.

04Pilot decision

One first build, acceptance checks—or a reason not to build yet.

The published £500 entry offer does not include production integrations, model tuning, data migration or ongoing operation.