Why this matters
AI tools are sold on demos, and fluent copy says nothing about whether a tool reads the right guest data or respects consent. The guide to choosing and evaluating tools puts integration, identity and data handling at the heaviest weights because they cannot be fixed after purchase. The data flow and integrations guide explains why: a tool that keys on the wrong identifier or arrives by file upload becomes a manual process that eventually stops.
The operator and its shortlist
Pinecrest Holiday Parks is an invented operator of lodge and caravan parks. Its booking system is the system of record for guests and consent; there is no customer data platform (CDPUnified first-party customer profiles.). The email platform already receives bookings through a native connector. A marketing officer writes pre-arrival, post-stay and re-booking emails by hand.
Pinecrest: evidence gathered on the two candidates
| Criterion | Weight | Candidate A: AI drafting feature inside the current email platform | Candidate B: standalone AI personalisation service |
|---|---|---|---|
| Use-case fit | 3 | Drafts and varies copy within templates; no per-guest product choice | Writes each email per guest, picks park activities by party |
| Integration | 3 | Uses the existing booking connector | Nightly file export from the booking system, to be built |
| Identity and consent | 3 | Reads the platform's contact record and consent flags | Keys on email only; the export carries no consent field |
| Data handling | 3 | Inputs not used for training under current terms; processing region covered | Inputs used to improve the service unless a higher tier is bought |
| Usability for the owner | 2 | Marketing officer used it unaided in the trial | Needed vendor help to change a template |
| Reporting | 2 | Results in the platform's existing reports | Own dashboard; no export of results to analytics |
| Governance | 2 | Approval step before send; change history | Approval flows and versioned prompts |
| Vendor | 1 | Established; feature labelled as new | Young company; responsive support |
| Total cost | 3 | See cost table | See cost table |
| Exit | 1 | Templates stay in the platform | Configuration not exportable |
| Two-year cost item | Candidate A | Candidate B |
|---|---|---|
| Licence | £75 × 24 = £1,800 | £400 × 24 = £9,600 |
| Set-up and integration | £400 | 40 hours × £60 = £2,400 |
| Maintaining the nightly export | £0 | 2 hours × 24 × £60 = £2,880 |
| Training | £300 | £300 |
| Total | £2,500 | £15,180 |
Pinecrest's marketing officer spends about 2 days a month on these emails, at about £30 an hour.
Scoring the tools
Weight the criteria before scoring, from what cannot be fixed later. Score each candidate from one to five with the evidence beside each score, and say where evidence came from a demo rather than a trial. Then apply the guide's rule: a candidate that fails an essential requirement does not win on the sum of the rest. For Pinecrest, consent read from the system of record and guest data kept out of a vendor's training are essential.
Task
- Mark which requirements are essential, important or useful, and justify the essentials.
- Score both candidates against the weighted criteria and total them.
- Compare two-year cost, adding any workarounds the integration checks reveal.
- Choose a candidate and write the decision record's key lines, including known gaps.
- Design a time-boxed pilot: use case, data, success measure, holdout, end date and decision rule.
- Say which onboarding steps must happen before real guest data reaches the tool.
Worked answer
Essentials. Consent read from the booking system; guest data not used for vendor training; results returned to the platform's reports so outcomes can be measured. Candidate B fails the first two outright.
Pinecrest: weighted scores out of 115
| Criterion | Weight | A score | A weighted | B score | B weighted |
|---|---|---|---|---|---|
| Use-case fit | 3 | 3 | 9 | 5 | 15 |
| Integration | 3 | 5 | 15 | 2 | 6 |
| Identity and consent | 3 | 5 | 15 | 2 | 6 |
| Data handling | 3 | 4 | 12 | 2 | 6 |
| Usability | 2 | 4 | 8 | 3 | 6 |
| Reporting | 2 | 4 | 8 | 2 | 4 |
| Governance | 2 | 3 | 6 | 4 | 8 |
| Vendor | 1 | 4 | 4 | 3 | 3 |
| Total cost | 3 | 5 | 15 | 2 | 6 |
| Exit | 1 | 3 | 3 | 1 | 1 |
| Total | 95 | 61 |
Time currently spent: 2 days × 7.5 hours × £30 = £450 a month. If A halves it, the saving is £225 a month, or £5,400 over two years against £2,500 of cost: a return on investment (ROIProfit relative to cost.) of (£5,400 − £2,500) ÷ £2,500 ≈ 116%, before any change in bookings.
Choice and record. Candidate A, on the essentials and on cost. B's better fit is real but rests on the demo and on a nightly export that would key on email alone and lose consent, exactly the identifier mismatch and consent leakage the data flow guide warns against. The record notes the gap: no per-guest activity choice. B could be revisited with a higher tier and a governed feed.
Pilot. Post-stay re-booking emails at two parks for eight weeks, with a fixed end date. Departing guests are split at random: half receive AI-drafted, human-approved emails, half the current hand-written template. The success measure is time per email and defects found in review, with re-booking rate as a guardrail KPIA metric tied directly to an objective. that must not fall below the control. The case rests on time saved, because a pilot this small cannot reliably detect a booking lift. Decision rule, written before the start: adopt if time per email falls by at least a third with no rise in defects and no fall in re-bookings; extend once if results are mixed; stop otherwise.
Before real data. The marketing tool onboarding SOP applies even to a pilot that touches real guest data: approval with a named owner, a data and privacy review with the data protection lead, access by named account, a test of consent handling and training before editing rights. Send only the fields each email needs. The privacy review gives general direction, not legal advice.
Common mistakes
- Scoring from the demo. Fit looked strongest for B and was the least proven.
- Comparing licence prices. B's export and its upkeep cost more than its licence difference suggests.
- A pilot with no end date or rule. It runs until someone likes it.
- Measuring what the pilot cannot detect. A small booking lift needs far more guests than eight weeks provide.
Next, take on the AI adoption plan capstone or return to the intermediate AI in marketing level.
