Fluent, Confident and Wrong: The Cost of Sending AI Output to Customers Unchecked

The tolerated pain

A customer asks whether a delayed order qualifies for a refund. The support assistant, powered by an AI tool, drafts a friendly and confident reply that cites a refund policy the company does not have. A busy agent clicks send, because the previous forty drafts were fine. Two days later the customer quotes the reply back, and the business must decide whether to honour a promise it never intended to make.

AI tools now draft quotes, proposals, support answers, collections messages and marketing copy. They are fast and usually fluent, which is exactly why unreviewed output is risky: errors arrive in the same polished tone as correct answers. This article looks at the cost of letting AI-generated work reach customers without suitable review, and how to set up review that is proportionate rather than a bottleneck. It complements the discovery note on AI outputs accepted without defined review thresholds.

Why unreviewed output slips through

The output looks finished. Fluent writing signals competence, even when a figure, date or policy is wrong.
Volume wears down attention. After many correct drafts, people stop reading closely. Review becomes a click.
Speed was the point. The tool was adopted to save time, so any step that restores the time feels like failure.
Nobody defined what review means. "Check it before sending" can mean a glance or a proper comparison with the source.
Accountability is blurry. If the tool wrote it, who answers for it? In practice, the business does.

Operational symptoms

customer replies, quotes or summaries sent with no record of who approved them;
staff who cannot say which messages were AI-drafted;
customers receiving prices, dates, policy statements or technical claims that do not match company records;
complaints that a message "sounded right but was wrong";
the same AI tool used for low-risk and high-risk messages with the same level of checking;
automated flows that send AI output directly to customers with no human step;
no log of corrections, so the same mistake recurs.

How the pain becomes money

Correction work: someone has to find the error, contact the customer, resend the right information and update the record.
Concessions: credits, discounts or honoured promises to keep goodwill, especially when a customer relied on a wrong figure.
Rework downstream: an incorrect quote becomes an incorrect order, an incorrect invoice and an argument. See fixing preventable mistakes: rework and customer time cost.
Reputation: customers remember a confident wrong answer more than a slower correct one, and may share it.
Legal and regulatory exposure: a message that misstates terms, or that discloses another customer's personal information, can create contractual and POPIA consequences. The Information Regulator's POPIA page is the starting point for the obligations that apply to personal information.
Loss of trust in the tool: after a visible mistake, staff may abandon a useful tool completely and the investment is lost.

An illustrative cost model

This example is illustrative. All figures are assumptions for a hypothetical Johannesburg service company that sends 300 AI-drafted customer messages a month with no structured review.

Assumption	Value
AI-drafted customer messages a month	300
Share containing a material error	2% (6 messages)
Staff time to correct each error	1.5 hours at R300
Average goodwill credit or concession	R600
Serious incident a year (a wrong quote honoured)	R40,000
Proposed review: share of messages checked	20% of messages, 3 minutes each

Correction time: 6 × 1.5 × R300 × 12 = R32,400
Concessions: 6 × R600 × 12 = R43,200
Serious incident: R40,000
Illustrative total: R115,600 a year
Review effort for comparison: 300 × 20% × 3 minutes = 3 hours a month, or 36 hours a year at R300 = R10,800

If the review catches even half of the errors, the cost is recovered several times over. The numbers are assumptions, and a lower error rate would change the result, but the structure holds: review costs minutes, while errors cost hours and goodwill.

Why the pain is tolerated

Most AI-drafted messages are fine, so the business sees a high success rate and a small number of embarrassing exceptions. The exceptions are handled one at a time and rarely connected. The time saved by the tool is easy to count, while the time spent repairing its mistakes is spread across support, sales and finance, so the net benefit looks better than it is. There is also an understandable reluctance to add steps to a tool whose appeal is that it removes them.

When it becomes commercially urgent

a customer acts on an incorrect AI-written statement and asks the business to honour it;
an AI-written message contains another customer's information;
a regulator, key client or insurer asks how the company governs AI use;
the business plans to connect AI output directly to automated sending;
complaints begin to mention inaccurate or generic replies.

What better looks like

The aim is a review step sized to the risk of each type of message, not blanket approval of everything.

Sort messages by consequence. A thank-you note is low risk. A quote, refund decision, legal statement, technical instruction or message containing personal data is high risk.
Set review levels. Low risk: periodic sampling. Medium: a quick check against the source record. High: named human approval before sending.
Give the reviewer the source. A reviewer who has the price list, policy or order record beside the draft can check facts. One who has only the draft can only check tone.
Keep a human in the loop for sending. Do not connect AI output to automated customer delivery until the error rate for that message type is known and acceptable. See how to reduce human error without removing human judgment.
Record who approved what. A simple log of draft, approver and date allows learning and answers customer disputes.
Track and use corrections. Count errors by type and feed them back into prompts, templates and approved source material.
Use a recognised framework as a checklist. The NIST AI Risk Management Framework organises AI risk work around governing, mapping, measuring and managing risk, and its playbook offers suggested actions. A small business can borrow the questions without adopting the whole programme.

Review works best when the underlying information is sound. If the tool is drawing from outdated or inconsistent records, the draft will inherit those faults; see AI on poor-quality data: the financial cost.

Practical next actions

List every place where AI-generated text can reach a customer today, including email assistants and chat tools.
Mark each as low, medium or high consequence.
Pick one high-consequence type, such as quotes, and require named approval from this week.
Start a simple log of AI-related errors and corrections.
Remove any direct, unreviewed AI-to-customer automation until it has been tested.
Tell staff plainly that they, not the tool, are accountable for what is sent.

Bottom line

AI drafting is useful, and review does not have to cancel the benefit. The risk lies in treating fluent output as verified output. A proportionate approach, where low-risk messages are sampled and high-risk ones are approved by a person who can see the source, protects customers and margins while keeping most of the time saved. The expense is a few minutes of attention. The alternative is paying in credits, correction time and customer trust.

Sources and further reading

AI Risk Management Framework (last verified 10 October 2026)
NIST AI RMF Playbook (last verified 10 October 2026)
https://inforegulator.org.za/popia/ (last verified 10 October 2026)