Skillnaija Start free

A support-desk pack for one real business

The capstone project for AI Agents for Customer Service: Run an Honest AI-Assisted Support Desk on WhatsApp, Instagram, Email and Web Chat — a free AI & Emerging Tech course. Pass it and you earn a certificate anyone can verify.

Start the course free Download the project workbook

The project unlocks once you complete the lessons. The workbook is free to download now so you can see exactly what is expected.

What you will submit

One shared document link with the audit and intent list, the front door and desk log, the knowledge base with sources and a never-say list, the persona card, the tier and escalation sheets, five handled conversations, the complaints, privacy and retention lines, a quality scoresheet for ten replies with the weekly numbers and a monthly report, and a retainer offer or a portfolio page; plus a write-up (5,000 characters at most) that pastes the words proving each part.

What to do, step by step

  1. Choose one real business: your own, or one whose owner has agreed in writing to let you run or rehearse its customer desk. If you have neither, run the desk for your own service. Never use a business's chats without the owner's written yes. Use free tools only: no card, paid plan, laptop, CAC registration or website is needed to pass, and a phone is enough for every graded task; a channel your phone cannot set up may be left out if you say what stopped you. Remove every customer's name, number, address and order before anything goes into an AI tool or into your submission: write Customer C, 080X XXX XXXX and an order number with X's. Never write a password or login code anywhere. A business with few chats may role-play conversations with a friend reading a script you wrote from your audit; say which ones were role-played.
  2. Part 1. The audit: the places customers message the business, who answers each and in what hours; one sentence saying what the desk is for and what it will never do by itself; and the transcript audit of four weeks (or the last 50 conversations; for a new business, the first question each of ten likely customers says they would ask, and say that is what you did; never make up a count), de-identified, with the intent list, counts, one example of each intent in the customer's own words, and the three kinds: answered from written facts, needs a record checked, needs a person's decision.
  3. Part 2. The front door and the queue: your greeting and away message (one message each, saying AI is used and a person checks, the real hours and the way to a person); up to five quick replies with no product prices; any Instagram FAQs; the copies list; what you saw when you looked for Meta Business Agent in Tools (if it is not there, write that down), with the phone and the date. If you could not set these on the business's own phone, the written texts count in full; never move your personal WhatsApp number to the Business app to practise; if no phone you may use has the WhatsApp Business app, write the dated line 'No WhatsApp Business app on a phone I may use; Tools not checked'. Then the channel map (each channel's tool, who answers, the plan's limit and the date you read it); the desk log's columns; and a three-line meter note for a client who asks about a WhatsApp bot: the WhatsApp count from your audit (only WhatsApp messages go through the WhatsApp rate), the working in US dollars, and the date of the rate.
  4. Part 3. The knowledge base, persona and hard lines: every entry with its source, who confirmed it and the date; the instructions you gave the AI assistant, and what you saw in its settings about chats training its models, with the date; your launch test (one control question per entry, asked three ways, with a Nigerian reader checking any Pidgin) and what you fixed; the persona card with your disclosure line and your answer to 'are you a robot?'; and the never-say list, at least eight lines, each with its reason.
  5. Part 4. Triage and escalation: the tier sheet (every intent in a tier, who does what, the first-reply target and when the real answer comes), the five checks before send, and the escalation sheet (every issue that needs a person, with an owner, a backup, when the customer hears back and how), your keyword exit and your rule for automatic handover.
  6. Part 5. Five handled conversations: one Tier 1, one Tier 2, one Tier 3, one complaint or refund, and one about AI disclosure or a customer's data. For each: the customer's message, the AI's draft, what you changed and why, the first reply you sent, and the human finish.
  7. Part 6. Complaints and data: who decides what (with no blanket 'no refunds'), the complaint log's columns, your payment line, the privacy notice with every item NDPA 2023 s.27(1) asks for (a starting draft for the owner, its retention line matching the data map), the data map with retention periods, the recorded basis for using an AI tool outside Nigeria, your processor-agreement lines, and the NDPC registration note with where the owner will confirm it and the date you wrote it.
  8. Part 7. Quality and numbers: a quality scoresheet for ten replies on the five checks, with the cause of every fail fixed; your weekly numbers with their definitions, reporting resolved conversations divided by all conversations and never counting silence as resolved; and a monthly report to the owner for the weeks your desk ran, saying how many (for a role-played desk, from your role-play set, saying so).
  9. Part 8. The offer or the portfolio: a one-page retainer offer priced from your own logged hours and your own hourly rate (with its working), saying whether being reachable through your desk hours is priced or not promised, what you do and what you do not promise; or a job-application portfolio page showing the de-identified pack with the owner's written permission. No income, job or sales promise, and no claim that AI runs the desk unsupervised.
  10. Put everything in one shared document (a Google Doc works; on a phone, use the Google Docs app), set sharing to 'anyone with the link can view', and submit the link. A table may be written as one line per row with the fields separated by ' / ', and a clear photo of a handwritten sheet may be pasted in, provided no customer is identifiable. In the write-up box (5,000 characters at most) paste the words that prove each part, because your write-up is what gets marked: your three biggest intents with counts; your greeting; two knowledge-base entries with their sources and dates and three lines of your never-say list; your answer to 'are you a robot?'; one escalation-sheet row; two of your five conversations in full (draft, change, sent reply, finish); your privacy notice; your quality total and the cause you fixed; your resolved figure with its definition; and the price line of your offer with its working, or the permission line of your portfolio page.

Files to work with

Project workbook (fill in, then submit)The whole brief, an evidence checklist, the grading rubric as a self-check and the link-sharing steps in one file. Opens in Word, Google Docs or WPS.Word · 9 KB

How it is graded

CriterionWeight
The audit is real (or the allowed substitute, clearly labelled) and de-identified, and the knowledge base comes from it: every entry has a source and a date, a copies list exists, a launch test was run and its failures fixed, and the never-say list covers invented prices and promises, refund decisions, blanket 'no refunds' and payment secrets20%
The front door is honest and fits the business's real channels: one-message greeting and away message that say AI is used, give real hours and a way to a person; no product prices in quick replies (written texts count in full where the business's phone could not be used); the Meta Business Agent check is dated, or the dated reason it could not be done; each channel's tool is given with its limit and the date read; the WhatsApp meter note is worked in US dollars with the date of the rate15%
Triage and escalation: every intent has a tier; nothing that needs a decision is left to the AI; every issue that needs a person has an owner, a backup and a time; there is a keyword exit and a rule for automatic handover20%
The five conversations: each first reply is correct and complete in one message, invents no price, policy or delivery promise, is truthful about AI when asked, and the human finish closes the case; no customer is identifiable anywhere; role-played conversations are labelled and are not penalised15%
Complaints and data: no blanket 'no refunds', the FCCPC line and a complaint log; a privacy notice with the NDPA s.27(1) items, retention periods, a recorded basis for the foreign AI tool, processor lines and a dated NDPC registration note10%
Quality and numbers: ten replies scored with the causes of failures fixed; resolution reported as resolved conversations divided by all conversations, with silence not counted; the monthly report states its definitions10%
The offer or portfolio: priced from the learner's own logged hours and rate, with the availability window priced or not promised and what is not promised, or a portfolio page with the owner's written permission; no income, job or sales promise; no claim that AI runs the desk unsupervised10%

You need 65% overall to pass. A failed submission comes back with feedback and can be revised and resubmitted once free; further attempts come with the All-Access Pass or as a paid retake.

Before you submit: make your link public

If your work lives in Google Drive or Google Docs, open Share → General access and change “Restricted” to “Anyone with the link” (Viewer). Then paste the link into a private browser window: if it opens without sign-in, you are ready. A private link cannot be graded — it is the single most common reason a good project fails.

Frequently asked questions

Do I have to finish AI Agents for Customer Service: Run an Honest AI-Assisted Support Desk on WhatsApp, Instagram, Email and Web Chat before submitting the project?

Yes. The submission screen opens once every lesson in AI Agents for Customer Service: Run an Honest AI-Assisted Support Desk on WhatsApp, Instagram, Email and Web Chat is marked complete. The lessons are where the methods, the local context and the worked examples the project depends on are taught.

How is the project graded?

An examiner scores each rubric criterion from 0 to 100 and weights them as shown on this page; you need 65% overall to pass. Most submissions are graded within minutes and you get written feedback on what was strong and what to improve.

Can I resubmit if I fail?

Yes. A failed project comes back with feedback; revise the weak parts and resubmit. A passed project is final — the certificate is issued and cannot be re-rolled.

What do I actually submit?

A write-up of what you did (under 5,000 characters) plus a public link to your work — a Google Drive folder, Google Doc, spreadsheet, GitHub repository or video. The link must open without sign-in; a private link cannot be graded. The free project workbook on this page walks you through all 10 steps.

Is the certificate real?

Yes. Passing this project issues a certificate with a unique verification code. Anyone — an employer, a client, a school — can open the verification page and see that the certificate is genuine and which project earned it.

More AI & Emerging Tech projects

Ready to earn this certificate?

AI Agents for Customer Service: Run an Honest AI-Assisted Support Desk on WhatsApp, Instagram, Email and Web Chat is self-paced. Finish the lessons, complete this project, and the certificate is yours to share.

Start learning free