By KYN AI Advisory Team — AI implementation specialists, Singapore
A small team should run an AI SDR with one narrow job, a named approver, a source for every claim, and a weekly check against real replies. Start with research and drafting only. Add reply sorting second. Leave sending, pricing and closing with people. A team of two to ten gets the most from the agent when it removes typing and logging, not when it replaces judgement.
What is an AI SDR, and what should a small team ask of it?
An AI SDR is an agent that does the work of a sales development rep: finding accounts, writing the first touch, sorting replies and booking meetings. For a small team, the useful question is not "how much can it do?" but "which task do we hate and can check?"
Good first tasks have a right answer you can verify:
- write one factual line about an account from a page you can open;
- label a reply as interested, not now, wrong person, unsubscribe or question;
- write the call or email result to the CRM.
Poor first tasks depend on taste or carry risk: setting a discount, deciding who to chase, replying to a complaint.
For a plain definition and the full list of limits, see what an AI sales agent is, and what it cannot do.
Practice 1: Define one job and one metric
Pick one job for the first 30 days. Pick one number that tells you if it works.
| Job | Metric | Test | |---|---|---| | First-touch drafting | Share of drafts approved with no edit | Review 50 drafts | | Reply sorting | Agreement with a human label | Label 100 replies | | CRM logging | Records with a complete log | Audit 30 records |
Running three jobs at once makes failures hard to trace. If reply rates fall, you will not know whether the drafts, the list or the sorter is at fault.
Practice 2: Put a source behind every claim
The agent should write only what it can point to. A first line such as "Your site lists a new Surabaya warehouse" carries a URL in the draft record. The reviewer clicks it, sees the claim holds, and approves.
Three rules keep this honest:
- No source, no draft. The agent returns "no checkable fact found" rather than inventing one.
- The source is stored with the message, so a complaint can be traced.
- Generic praise ("impressive growth") is rejected by the reviewer. It tells the reader nobody looked.
This also protects you. If a recipient asks how you got their details or why you wrote, the record answers.
Practice 3: Keep a person on every outbound send
Anything outbound, anything about money and anything irreversible waits for a named person. For an SDR that means each first-touch message, each message quoting a price or date, and each deletion or merge of a contact.
Approval is quick when the screen is right. The reviewer sees the draft, the cited source and a one-click approve or edit. A batch of 30 clears in minutes. If review feels slow, fix the screen before you remove the gate.
Move the gate only when the numbers justify it. Our guide to the sales agent autonomy ladder sets out when a task can move from approve to send.
Practice 4: Respect the sending rules from day one
Rules are easier to build in than to retrofit. For email to Gmail users, Google's sender guidelines ask every sender to authenticate with SPF or DKIM and to keep the spam rate in Postmaster Tools below 0.10%. In Singapore, the Spam Control Act 2007 sets the unsubscribe requirement for unsolicited commercial email, and the PDPC's Do Not Call Registry guidance covers marketing messages to Singapore numbers. Other markets have their own rules, so check the recipient's country.
Build these as code:
- a suppression list checked before every send;
- an unsubscribe that stops the contact on every channel;
- a daily cap per inbox;
- a registry check step before any message to a number in a jurisdiction that has one.
Practice 5: Check the agent against reality every week
A weekly 30-minute review is enough. Bring four numbers:
- drafts approved without edit, out of drafts reviewed;
- replies the agent mislabelled, from a sample of 20;
- unsubscribes processed against unsubscribes received;
- cost per unit, such as per draft or per reply sorted.
Unit cost is worth tracking because it is small and stable. KYN's live outreach runs at about S$0.02 per email and reply sorting at about S$0.01 per reply. Those depend on the model and the text length, so measure your own.
When a number drifts, change one thing and measure again. Changing the prompt, the list and the schedule in one week makes the result unreadable.
What does a 90-day rollout look like?
| Weeks | Scope | Exit test | |---|---|---| | 1 to 2 | Review the CRM and list quality; fix duplicates | Duplicate rate you can state | | 3 to 6 | Research and drafting into a review queue | 50 drafts reviewed, sources clickable | | 7 to 10 | Reply sorting with a human override | 100 replies tested | | 11 to 13 | Handoff alerts and CRM logging | Warm replies reach an owner within your set limit |
Sending stays human-approved throughout. Autonomy for any single step is a decision you make once the numbers back it.
What does a worked example look like?
Take an illustrative five-person consultancy that contacts 40 accounts a month. The founder spends about six hours on research and logging. The team picks research and drafting as the first job and sets one number: share of drafts approved with no edit.
| Week | Drafts reviewed | Approved with no edit | Action | |---|---|---|---| | 1 | 20 | 8 | Tighten the source rule; reject generic praise | | 2 | 20 | 12 | Add the firm's own wording for the offer line | | 3 | 20 | 15 | Hold the prompt steady and read replies | | 4 | 20 | 16 | Keep drafting at approve-to-send |
These numbers are illustrative, not a benchmark. The pattern is what counts: one change per week, one number to read, and a decision at the end of month one. If the figure stalls below your limit, the fault is usually the source material, not the model.
What are the common failure modes?
- The list is the problem. An agent writes a fine first line to the wrong person. Check 30 accounts by hand before blaming the draft.
- Reviewers stop reading. Approval rates of 100% for three weeks mean nobody is checking. Seed one deliberately wrong draft a week and see if it is caught.
- Sources go stale. A page cited in March may have changed by June. Store the retrieval date next to the URL.
- Scope creep. Someone asks the agent to chase invoices or answer support mail. Add jobs one at a time, each with its own number.
Where KYN fits
KYN builds SDR work as part of a system in your own cloud account. For an insurance brokerage, we built outreach with one checkable first line and a reply desk that sorts each reply. Every action is logged with the actor, and outbound messages, money and deletions wait for a named person. Agents run on their own restricted accounts.
The first 14 days are a review of your stack and interviews with your team. A full build ships in under 90 days. Pricing is a one-off build fee, then a monthly retainer of at least twelve months, quoted after the review. See sales systems and pricing.
Frequently asked questions
Can a team of three use an AI SDR?
Yes, if the volume of accounts justifies it. The agent removes research and logging time, which matters most when each person also does delivery. If you contact fewer than a few dozen accounts a month, a template and a good list may be enough.
What should I automate first?
Reply sorting or CRM logging. Both have right answers you can test, and neither carries your name outbound. Add first-touch drafting once those two pass your checks.
Do I need to tell recipients they are talking to an AI?
Rules vary by country and channel, and the law is moving. Ask a lawyer for the markets you contact. A safe default is that a named person approves and sends every message, so the sender is a person.
Checked 30 October 2026.
Curious whether Marketing agents fits your business? Talk to KYN on WhatsApp — no forms, just a conversation.
