By KYN AI Advisory Team — AI implementation specialists, Singapore
Start with one repetitive task that has a right answer a person can check, and that runs on data you already hold. In most SMEs that means bank and bill reconciliation, a morning operations report, enquiry sorting, or first drafts of marketing content. Name the problem, map where the data lives, score the options, and pilot one. Buy or build only after those steps. For the buy-or-advise question, see why talk to an adviser before buying a tool.
Step 1: What problem are you solving?
Write one sentence per candidate: who spends how many hours a week on what, across which systems, and what goes wrong.
Good: "Our finance lead spends about six hours a week matching card charges to receipts and chasing late invoices." Weak: "We should use AI in finance."
Ask each team for three tasks they repeat weekly and dislike. Collect them without filtering. You want a list of ten, not one idea.
Step 2: Where does the data live?
For each candidate, list the systems the task touches and how data moves between them.
- Accounting platform, bank feeds, ERP
- CRM, email, WhatsApp
- Spreadsheets, shared drives, paper
- Government portals for filing
Mark each handover as automatic, manual or retyped. Retyped handovers cause errors and are where automation pays first. Check the data itself: duplicate customers, inconsistent field formats and missing references all produce wrong output at speed.
Step 3: How do you rank the options?
Score each candidate from 1 to 5 on five questions, then add them.
| Question | 5 means | |---|---| | Frequency | Daily or weekly | | Checkability | A person can verify the output in minutes | | Data readiness | Data is structured and in one or two systems | | Cost of error | Low, or a person approves before it matters | | Time saved | More than five hours a week |
Pick the highest total with a cost-of-error score of 3 or more. A task where a wrong output is hard to detect or expensive is a poor first project, however large the saving.
Which starting points work for most SMEs?
| Area | First project | What stays with a person | |---|---|---| | Finance | Match bank lines to invoices and receipts; capture bills from email | Posting adjustments, payments, filings | | Operations | A morning report joining stock, production and cash | Decisions on the report | | Sales | One inbox for email and WhatsApp; sort replies; log to CRM | Sending to new contacts, quotes | | Marketing | Draft articles from checked facts | Publishing | | Internal | Route approvals and generate documents from templates | Sign-off |
KYN's own books follow this order. Bills are captured from Gmail, receivables are chased at 3, 14 and 30 days, and payroll runs on 2026 CPF rules. Each started as one narrow task.
Why pilot before you commit?
A pilot answers three questions that a plan cannot.
- Is the output right often enough? Take 100 real cases, have a person label them, and compare.
- How long does a person take to review? A draft that takes longer to check than to write is not a saving.
- What does one run cost? For reference, our live systems cost about S$0.02 per outreach email, S$0.01 per reply sorted, S$0.03 per benefits comparison and S$0.05 per morning report. Your figures depend on the model and the length of the text.
Run the pilot on real data in a copy of your environment. Read-only first, then writes.
A worked scoring example
This example is illustrative. The company, the candidates and the scores are invented.
A 25-person trading company lists five candidates and scores each from 1 to 5 on the five questions. Cost of error is scored so that 5 means a low cost of error.
| Candidate | Frequency | Checkability | Data readiness | Cost of error | Time saved | Total | |---|---|---|---|---|---|---| | Match bank lines to invoices | 5 | 5 | 4 | 4 | 4 | 22 | | Morning report for the owner | 5 | 4 | 3 | 5 | 3 | 20 | | Sort enquiries and draft replies | 5 | 4 | 3 | 3 | 4 | 19 | | Quote generation from templates | 3 | 3 | 2 | 2 | 3 | 13 | | Forecast demand with a model | 2 | 1 | 2 | 2 | 3 | 10 |
Read the table in two passes. First, drop anything with a cost-of-error score below 3: quotes and the demand forecast are out for now. Second, pick the top total. Bank matching wins at 22. The enquiry sorter comes third, and the team can start it once the first project has run for a quarter.
Notice what the table does. It stops the loudest idea, the demand forecast, from winning by volume. The scores are opinions, so write them down, have two people score separately, and discuss the gaps.
How do you check data readiness in a day?
Take the candidate with the top score and run five checks on a sample of 100 real records.
- Completeness. How many records have every field the task needs?
- Consistency. How many distinct spellings or formats exist for a field such as customer name, date or currency?
- Duplicates. How many records describe the same thing twice?
- Links. Can you follow a record from source to result, for example a payment to an invoice?
- Access. Can a system read the data without a person exporting it by hand? Most accounting packages, CRMs and email services have an API. A shared drive of spreadsheets may need a conversion step.
Score each check as pass, partial or fail. Two or more fails means fix the data before the pilot, and the fix itself is the first project.
What should the pilot's success test say?
Write the test before the pilot starts, as a table with a number in every row. The values below are illustrative.
| Measure | Today | Target after 8 weeks | How measured | |---|---|---|---| | Hours per week on the task | 6 | 2 | Timesheet of the person doing it | | Errors found at month-end | Count from last quarter | Fewer | Review log | | Items needing a person's decision | Not tracked | Tracked and falling | Review queue | | Cost per run | Not applicable | Within a set limit | System log |
If a target is missed, the pilot has still produced an answer. Decide in advance whether you will extend, change or stop, and who decides.
How should approval work during the pilot?
Start with full review, then relax only where the evidence supports it.
- Weeks 1 to 2. The system proposes, a person approves every item.
- Weeks 3 to 4. The system proposes, a person approves items the system marked uncertain, and spot-checks a sample of the rest.
- Weeks 5 to 8. Low-risk items proceed with logging. Anything involving money, outbound messages, filings or deletions still waits for a named person.
The last rule does not relax. A reconciliation system that posts a wrong entry, or a sales system that sends a wrong message, damages trust faster than a slow system does.
How do you decide between buying a tool and building one?
Decide per task, not per company.
| Situation | Lean toward | |---|---| | The task is standard, such as payroll or e-filing | Buy a package | | Few people use it and the process rarely changes | Buy a package | | The task spans two or three of your systems | Build a layer on top | | Your rules do not exist in any package | Build | | You need the system and data in your own cloud account | Build | | The package needs more than two workarounds | Build |
Many SMEs do both. They keep accounting and payroll on packages and build the layer that connects enquiries, orders and reports. The test is the same as before: list the steps of the task, and count the workarounds a package would need.
What does the second project reuse?
The first project leaves four assets.
- Clean data fields that other projects can read.
- An approval route that works, with named people.
- A log format that records who did what and when.
- A measured baseline that makes the next saving easy to compare.
A second project in the same area, such as bills after bank matching, is usually faster and cheaper than the first. A project in a new area, such as marketing after finance, starts with fresh data work, so plan for that.
What if the team resists?
Resistance has usual causes. Address the cause, not the person.
- "It will replace me." State what the person will do instead, in writing. Show the hours that move from keying data to reviewing exceptions.
- "It makes mistakes." It does. Show the error rate from the pilot and the review step that catches errors.
- "It is another tool to learn." Put the output where the person already works, such as the existing approval queue, the email inbox or WhatsApp.
- "Nobody asked me." Ask the people who do the task to write the problem sentence in step 1.
Involve the person who does the task today. They know the exceptions, and the system will inherit their knowledge.
What do you need before a vendor conversation?
Bring four things.
- The problem sentence for your top candidate.
- A map of the systems and handovers it touches.
- The data readiness result from the sample.
- The success test with numbers.
A vendor who sees these can scope the work in one conversation. A vendor who is shown none can only sell a product. Ask each one who owns the code and the cloud account at the end, what the monthly cost is after launch and what happens if you stop paying.
What does the first 30 days look like?
A concrete calendar makes the plan harder to postpone. This one assumes a small team and no outside help, and you can compress or stretch it.
| Week | Work | Output | |---|---|---| | 1 | Ask each team for three weekly tasks they dislike. Write one problem sentence per task. | A list of ten to fifteen sentences | | 2 | Map the systems and handovers for the top five. Score them on the five questions. | A ranked table | | 3 | Take 100 real records for the top candidate and run the five data checks. | A readiness result | | 4 | Write the success test with numbers. Name the owner and the approver. | A one-page pilot plan |
At the end of week 4, you either have a pilot plan or a clear reason to fix data first. Both are progress.
Which first project fits which kind of business?
The right first project depends on where the hours go. These are patterns, not rules.
| Business type | Common time sink | Typical first project | |---|---|---| | Trading or distribution | Matching payments, chasing invoices, order status | Bank and invoice matching, daily order report | | Professional services | Chasing documents, drafting standard letters, timesheets | Document chasing, first drafts for review | | Manufacturing | Stock counts, purchasing, production status across systems | A morning report joining stock, production and cash | | Retail or F&B with several outlets | Sales roll-ups, staff scheduling, supplier orders | Daily sales and cash report per outlet | | Agency or consultancy | Enquiries, proposals, follow-ups | Enquiry sorting and proposal drafts for approval |
If your business does not fit a row, ask where the person with the most spreadsheets spends their week.
How should you budget the first project?
Budget in three parts, and keep each visible.
- One-off build. The cost of designing, building and testing the first system. Quotes should follow a review of your stack, not precede it.
- Monthly run cost. Hosting and model usage. Ask for a per-run figure so you can multiply by your volume.
- Retainer. Changes, monitoring and fixes after launch. Ask for the minimum term and what is included.
Add your own internal time. Someone on your team will spend hours in reviews, tests and approvals. Count those hours at their salary cost. If the total over three years is less than the value of the hours saved, the project pays. If it is close, choose a different project.
What can go wrong in the first project?
- The scope grows. The team adds "just one more" report or rule. Hold the scope to the one-page pilot plan and record every request for the second phase.
- The owner is too busy. A project with a nominal owner stalls. Give the owner two hours a week, in the calendar.
- The data turns out worse than the sample. Stop and fix the data. Continuing produces wrong output at speed.
- The reviewer is not trained. Show the reviewer ten examples of correct and incorrect output before the pilot begins.
- Results are measured against a feeling. The baseline numbers from step 1 are the only fair comparison.
If any of these happens, say so in the monthly review and record the decision. A project that changes course on evidence is working as designed.
What are the common mistakes?
- Starting with the tool. A platform bought before the problem is named solves nothing in particular.
- Automating messy data. A CRM with three formats for one field produces wrong labels. Clean the records first.
- Skipping approval gates. Outbound messages, money, filings and deletions should wait for a named person.
- No owner. Every system needs a named person who answers for its output.
- No measure. Decide the number before the pilot: hours saved, errors found, time to first reply.
What should the first six months look like?
- Month 1: Review. Inventory systems, interview the team, and choose one project.
- Months 2 to 3: Build and pilot the first project on real data, with a person approving.
- Month 4: Measure against the number you set. Keep, change or stop.
- Months 5 to 6: Start the second project from the ranked list, reusing the data work from the first.
Keep the list ranked and short. Revisit it each quarter.
What support exists in Singapore?
Government schemes may offset part of the cost of qualifying projects. Check current eligibility and rates on the official pages, as terms change. Start at Enterprise Singapore and see our note on grants.
Where KYN fits
KYN builds custom systems for SMEs in Singapore, Hong Kong, Indonesia and Malaysia. Systems run in a cloud account in the client's name. Agents have their own restricted accounts, and every action is logged with the actor.
The first 14 days are the review: an inventory of your stack and interviews with the team. A full build ships in under 90 days. Pricing is a one-off build fee plus a monthly retainer of at least twelve months, quoted after the review. See finance systems and operations systems.
Frequently asked questions
What should be the first AI project in a small business?
A repetitive task with a checkable answer on data you already hold. Bank reconciliation, enquiry sorting and a morning report are common.
Do I need clean data before I start?
You need data clean enough for the one task you pick. Fix the fields that task uses. Do not wait for a company-wide clean-up.
Should I build in-house or use a partner?
Build in-house if you have an engineer who will own the system for years. Use a partner if you do not, and make sure the code and cloud account end up in your name.
Checked 30 October 2026.
Curious whether Marketing agents fits your business? Talk to KYN on WhatsApp — no forms, just a conversation.
