Audit Your Enquiry Response Before Adding AI

Decide What Counts As A Useful Response

An automatic receipt can confirm that a message arrived. For this audit, it does not count as a useful response. Count the first reply that answers the question, asks for a necessary detail, or gives a realistic next step with a named owner. Keep the receipt timestamp separately if you need it.

Choose one channel and a fixed review window. Practise with the fictional records below. If you later review real enquiries, keep the audit in your existing system and copy only the fields you need. Do not send customer messages to this website.

Record elapsed minutes consistently. If an enquiry arrives outside opening hours, flag it. Calendar minutes and staffed-hours minutes answer different questions; do not mix them in one median. The example below uses calendar minutes on one fictional working day, so no opening-hours conversion is needed.

1. Define a useful response in one sentence and agree who checks it.
2. Use a random record ID, received time, useful-response time, owner and next action. Leave out names, addresses and message contents.
3. Mark unanswered records as missing. Never turn a missing timestamp into zero minutes.
4. Review a fixed window and include open records, not just completed ones.

Context: NIST: AI Risk Management Framework 1.0.

Run The Six-Enquiry Rehearsal

Imagine a small maintenance business reviewing six invented enquiries at 16:00 UTC. All arrived during its fictional staffed hours. One is still unanswered. The owner and next-action columns are independent of response speed: a quick reply can still leave work unassigned.

Five records have useful responses. Their delays are 12, 18, 40, 240 and 90 minutes. Sorting gives 12, 18, 40, 90, 240, so the median among answered enquiries is 40 minutes. Their mean is 400 / 5 = 80 minutes. One out of six is unanswered at the review cutoff. Report both figures together; the median alone hides the open enquiry.

Two records lack an owner and three lack a recorded next action. These observations suggest an ownership and handoff check before reply drafting. They do not establish that a sale was lost, that the replies were good, or that a product will improve conversion.

from statistics import mean, median

# None means unanswered, not zero minutes.
delays = [12, 18, 40, 240, None, 90]
owners = ["A", None, "B", "A", None, "B"]
next_actions = [True, False, True, False, False, True]
answered = [d for d in delays if d is not None]
print("Answered median:", median(answered), "minutes")
print("Answered mean:", mean(answered), "minutes")
print("Unanswered:", len(delays) - len(answered), "/", len(delays))
print("No owner:", owners.count(None))
print("No next action:", next_actions.count(False))
# Answered median: 40 minutes
# Answered mean: 80 minutes
# Unanswered: 1 / 6
# No owner: 2
# No next action: 3

Repair One Handoff Before Automating

Write one operational rule: every new enquiry gets a named owner and a next-action due time. Choose a backup for absences. A shared inbox label, an existing CRM task or a simple internal checklist may cover this without AI.

Only after the rule is clear should you test automation. A fixed routing rule can assign a known category. A draft assistant might help with varied wording, but a person must still check the facts, the request and the next step. Keep price, availability and service commitments under human control.

For a small rehearsal, set a stop rule before starting: if any sample loses its owner or a draft invents a promise, stop and repair the process. Decide what your team considers an acceptable response window; this guide does not supply an industry benchmark.

Context: NIST: AI Risk Management Framework 1.0.

Use The Audit As A Baseline, Not A Sales Claim

The Scorecard asks about the average first useful human reply during working hours. Our median is a separate descriptive measure, not its input. In this invented staffed-day sample, the answered mean is 80 minutes; the open enquiry still needs separate review. Do not substitute the 40-minute median for an average or count a receipt as a human reply.

A later comparison needs the same definitions, review window and channel mix. Record changes in staffing, opening hours, enquiry type and workload. Include the open-record count and oldest unanswered record, not just an average for completed work.

Six invented records cannot tell you whether customers would buy or be satisfied. Measure those outcomes separately. This exercise records response delays; it does not measure sales or time saved.

The workflow is written for worldwide English readers, including the US, UK, Australia and Canada. There is no country-specific timing target in this method. Any future use of real messages, external processors or marketing follow-up needs a separate local privacy and permission review.