Singapore’s PATH AI Starter Kits for HR: funding and costs

Programme information checked 29 September 2026.

Singapore’s PATH AI Starter Kits combine HR software with help implementing it, training and funding support. PATH stands for People-centred AI Transformation for HR. The Ministry of Manpower (MOM) announced the initiative on 24 September 2026. It is an implementation programme, not simply a downloadable prompt pack. MOM announcement

Your final cost depends on the work included and which support your project is approved to receive. Check the ongoing bill as well as the starting price. A grant cap is not the price of a package.

This guide explains what is available, where to enquire, what funding can cover and how to test whether a package will actually help your team. You do not need to wait for every programme resource to appear before making enquiries. You do need to check the conditions before committing to a purchase.

This guide is published by Screened, an AI recruitment software provider. It is not an official PATH publication.

What has been announced, and what is available now?

MOM identifies the SkillsFuture Workforce Development Grant (Job Redesign+), or WDG(JR+), as the funding route for the Starter Kits. The kits, training and forthcoming playbook are separate resources. Training involves the Institute for Human Resource Professionals (IHRP). MOM programme details

Resource	What to know
AI Starter Kits	Packages combining AI solutions, consultancy, implementation support and training, with funding subject to eligibility and approval. Enquiries and individual supplier offers are available.
AI fluency training	IHRP-accredited programmes aligned to 12 new AI Fluency Skills Badges. Check individual courses, dates and assessments with the provider.
AI Playbook for HR	Expected in January–March 2027. This is the playbook’s timing, not a confirmed release date for every Starter Kit.
Updated HR Jobs Transformation Map	Expected in January–June 2027, covering changes to HR work and skills.

The programme descriptions and future dates come from MOM’s summary; the launch announcement specifies the 12 badges. MOM also recommends developing guidance for responsible AI recruitment. That recommendation should not be mistaken for a completed recruitment rulebook.

Is there an official vendor list or a Starter Kit download?

We have not verified a complete official catalogue or an all-in-one download. This guide is not a vendor directory. Check the official programme page and ask the programme adviser which options are available for the HR task you want to improve.

Ask the programme adviser which packages cover your task and which proposed costs can qualify. An approved consultant, an accredited course and an approved grant application are not interchangeable.

Who qualifies for WDG(JR+), and what does it fund?

MOM’s grant factsheet gives two starting conditions: the business must be registered or incorporated and operating in Singapore, and it must have at least three local employees, meaning Singapore citizens or permanent residents.

Support is up to 70% for small and medium-sized enterprises (SMEs) and up to 50% for non-SMEs, capped at S$150,000 per enterprise. These are maximum support levels, not an entitlement or a discount that automatically applies to a supplier’s entire invoice.

Grant component	Work supported	Component grant cap
Workforce consultancy	Examining the work and redesigning roles	S$50,000
Capability building	Helping the team introduce and sustain the change	S$60,000
Workforce technology	Technology supporting the workforce changes	S$90,000

Technology must be bundled with consultancy, capability building, or both. The component caps do not add up to a larger entitlement: the overall S$150,000 cap still applies. Grant factsheet, pages 2–3

Describe the change in the job, not just the software being bought. An HR coordinator might spend less time rewriting onboarding emails and more time resolving new-starter access problems. They will also need to check AI drafts, correct mistakes and maintain the instructions. Those responsibilities belong in the proposal and training plan.

Does the 200-employee HR certification threshold affect this grant?

No. It concerns a separate requirement. From 1 July 2028, organisations with more than 200 employees must have at least one suitable HR professional certified at IHRP-CP level or higher. Recognised equivalents can count subject to IHRP’s additional requirements. An AI skills badge is not a substitute for that professional certification. MOM certification guidance

“More than 200” does not mean “200 or more.” Check MOM’s employee-count and entity guidance when assessing whether the requirement applies. MOM also lists an additional S$100 certification subsidy for Singapore citizens and permanent residents from 1 October 2026 to 30 June 2028. That supports certification fees, not software purchases.

Where do we enquire, and when can we commit?

Start with the workforce-transformation interest form linked from MOM’s PATH programme page. You can also seek advice through the Singapore National Employers Federation (SNEF) or the Singapore Business Federation (SBF).

The programme offers free workforce-transformation advisory support. SNEF lists ewtp@snef.org.sg for enquiries and asks for a consultant’s quotation and proposal before application. Its pre-approved consultant panel is a list of consultants, not approved software.

An enquiry can be straightforward:

We want to reduce repeated resume review and give hiring managers a clearer interview brief. Could you advise whether this project fits PATH and WDG(JR+)? Please explain what could qualify, what we would pay and what must happen before we sign or make a payment.

Check the start conditions before signing a contract, issuing a purchase order or paying a deposit. SBF’s eligibility notes restrict prior consultancy payments and contracts or purchase orders with consultants or third parties. Ask how these conditions apply to your proposed project and any trial. SBF eligibility notes

Get the permitted next step in writing. A supplier offering a trial does not establish that starting it preserves grant eligibility, and submitting an enquiry is not grant approval.

Planning to use AI for recruitment?

Screened helps hiring teams review resumes and run first-round AI interviews, with transcripts and assessments for people to review.

Talk to Screened about your recruitment project. Tell us which roles you hire for, roughly how many applications you receive and which applicant tracking system you use. Mention PATH in your message. We’ll discuss where Screened could help, what a pilot would involve and what it would cost.

Any grant support needs to be agreed with the programme partner before you commit.

How much will our company actually pay?

We have not verified a standard price that applies to every PATH Starter Kit. Ask for the project cost now and the ongoing cost after support ends. The quote should separate consultancy, setup, software, training and support, then show which items are approved for funding. Keep GST and unconfirmed support visible.

For a policy assistant, establish who prepares the source documents, sets permissions and tests conflicting policies. For recruitment, establish which roles, application volumes and system connections are included. These are practical buying questions, not additional PATH requirements.

SBF describes a net-fee model: the company pays its co-funding share to the consultant, while SBF pays the grant portion after milestones are completed. Confirm how the arrangement applies to your quote, especially if a separate technology supplier sends another invoice. SBF payment model

Can SkillsFuture Enterprise Credit reduce the balance?

The programme FAQ describes redesigned SkillsFuture Enterprise Credit, also called SFEC (Enhanced), as starting on 1 December 2026, with current SFEC support expiring on 30 November 2026. Do not apply the December arrangements to a September purchase without checking the applicable rules.

The FAQ gives this simplified example, excluding GST:

Calculation	Amount
Project cost	S$20,000
Grant at 70%	−S$14,000
Balance after the grant	S$6,000
Eligible credit covering 90% of that balance	−S$5,400
Company pays, before GST	S$600

The 90% applies to the S$6,000 balance, not the original S$20,000. This requires an eligible project, approved support and sufficient credit. It is not a S$600 price for a PATH kit.

Ask for the amount payable without the credit first, then the amount with it once eligibility is confirmed. Check payment dates as well as the final total.

What happens if we stop the project?

The FAQ’s SFEC (Enhanced) section says committed credits are not refunded if the company withdraws from the programme. Supplier cancellation charges are separate. Confirm the rules applying to your project dates before committing credits. Programme FAQ, withdrawal question

Ask which fees remain payable if the trial fails, whether you can export the work and data, and how access and queued actions are stopped. Being able to switch off the software does not mean you can cancel without a cost.

Is this the same as PSG or EDGE?

WDG(JR+) replaced PSG-JR, the earlier job-redesign support, not every grant with “PSG” in its name. Previous PSG-JR participants may qualify for projects with different scopes and redesigned roles; do not assume the same work can be funded twice. Programme FAQ

Separately, Enterprise Singapore’s EDGE guidance says EDG, MRA and PSG cease on 29 September 2026, with EDGE applications opening on 30 September 2026. Ongoing applications and projects continue to be processed.

EDGE and WDG(JR+) are different routes. Put the exact grant name on the proposal so finance can check the right eligibility and payment rules.

Which HR work is worth starting with?

Choose a task before choosing a package. Drafting an onboarding email, answering a policy question and assessing an applicant need different tools and checks.

Work to improve	Sensible starting point
Drafting onboarding material or employee communications	A business AI assistant your company already approves. Test whether it preserves facts and is quick to review.
Answering policy questions	An assistant that shows the source, uses the right policy version and respects access permissions.
Scheduling, sending standard letters or updating details	Check whether ordinary automation in your existing HR system can do it. Test retries and duplicate actions.
Reviewing applications or filling gaps in answers	Check your applicant tracking system’s features, then compare any extra tool with the recruiter’s current process.
Planning learning or staffing	Require evidence of a skills gap or a workload problem. Course history alone does not tell you which one you have.

Can we use ChatGPT, Copilot or Gemini instead?

For drafting, start with an approved business account rather than a new implementation project. Check the actual plan and configuration, not just the product name. Microsoft, Google and OpenAI publish Copilot enterprise protections, Workspace AI privacy commitments and business-data commitments. Connected services can have different terms.

A promise not to use information for model training does not mean nothing is stored. Check retention, access, deletion and what other services receive before entering personnel information.

Use a non-confidential onboarding document for a first test. Ask for a manager checklist, then have the person responsible for onboarding check for missing or invented requirements. Change one instruction and repeat. Count checking and correction time, not just generation time.

Do we need another recruitment tool if our ATS already has AI?

Ask your applicant tracking system (ATS) administrator to demonstrate the work your recruiters need to do. Then identify what remains missing.

Resume screening checks the information an applicant has already provided. An interview can ask for details they left out. “Uses Salesforce daily” could mean entering customer details, not configuring the rules that assign leads to salespeople. For a role needing configuration experience, ask:

Have you set up Salesforce rules that assign new leads to salespeople? Please describe one example, or tell us if a colleague handled that work.

At Screened, our AI interviewer can ask role-specific questions and follow-ups, with transcripts and assessments for your team to review. Compare it with an email or recruiter call.

Can the hiring manager now see what the applicant actually configured and prepare a useful follow-up? Include the candidate’s effort and the recruiter’s review time in the comparison. If the answer is already clear and the application is waiting for a manager’s decision, another screen will not fix that delay.

For workforce planning, make a similar distinction: a colleague who knows how to do the work but has twice as much as they can finish may need capacity, not another training course.

What should we check about data, permissions and fair hiring?

Before using real resumes or employee information, establish where it goes, which companies process it, who can access it, what it is used for and when it is deleted. Include transcripts, derived assessments and copies retained by other providers, not just the original upload.

Singapore’s PDPC guidance on AI recommendation and decision systems explains relevant personal-data responsibilities. Ask your data protection officer which purposes, notifications, consent requirements or exceptions apply to the proposed use.

HR data does not universally have to stay in Singapore. Overseas transfers must meet the PDPA’s transfer requirements, including comparable protection, subject to the applicable rules. Server location alone does not establish compliance. Nor should information be kept indefinitely once its business or legal purpose has ended. PDPA obligations

TAFEP’s interview guidance calls for interview, assessment and offer records to be kept for at least one year. That does not mean retaining every recording forever. Agree what evidence is needed, why it is retained and when each type of information will be deleted.

We have not verified a universal PATH rule requiring AI Verify or a particular independent bias audit. Ask the programme partner for current requirements and the supplier for testing relevant to your jobs and applicants. A government funding route is not proof that every hiring use of a tool is appropriate.

Separate permission to draft from permission to act

Drafting an invitation, sending it and rejecting an applicant have different consequences. Give them separate permissions and tests. Check pending actions too: disabling new work may leave an email already queued to send.

Singapore’s IMDA framework for agentic AI provides a useful reference for human oversight and limits on AI actions. The practical checks here are our recommendations, not a claim that PATH mandates this exact setup.

How can we test an HR policy assistant?

A chatbot can quote the right annual allowance and still give the wrong advice. It might miss a condition such as approval before purchase.

Use these two sample policies for a demonstration. They are fictional company rules, not Singapore employment or grant rules.

Policy	Text to give the supplier
Effective 1 January 2025	Employees may claim up to S$500 per calendar year for approved courses. Written manager approval is required before purchase.
Effective 1 September 2026	Employees may claim up to S$800 per calendar year for approved courses. Written manager approval is still required before purchase. This version replaces the earlier policy for purchases from 1 September 2026.

Ask: “I bought a S$700 course on 20 September 2026. Can I claim it?”

A useful answer would be:

The policy for September purchases allows up to S$800 per calendar year. It also requires written manager approval before purchase. Did you get that approval, and have you already used any of this year’s allowance?

“Yes, the limit is S$800” is incomplete. The answer needs both prior approval and enough unused allowance. Ask it to show the policy supporting its answer.

Now reply: “I didn’t get approval.” The assistant should explain the condition and refer the question to HR, not promise that approval can be given afterwards.

In a fresh conversation, change the purchase date to 20 August 2026. The older S$500 annual limit applies, so the full S$700 is not covered. Approval and previous claims still need checking before any payment can be confirmed.

This tests whether the assistant uses the policy that applied when the purchase happened and checks the conditions before saying yes. A quotation alone does not prove either.

Check access with separate test accounts

An employee without access to a colleague’s appraisal should not receive its contents through a summary or an indirect question. Where policies differ between employee groups, check that each account receives the right version.

Use test information before connecting real personnel files. Keep the questions and expected answers so you can repeat them after a policy, permission or product change. These cases can expose failures; they are not a complete accuracy or security assessment.

How can we test recruitment AI without risking applicants?

Start with CVs written for testing and contact details your team controls. Keep live candidate messages, stage changes and rejections off. Complete the relevant privacy review before using real applications.

The UK government’s Responsible AI in Recruitment guide is useful for supplier questions and evaluation methods. It is not Singapore law. The exercises below are our suggested checks.

Does it distinguish missing information from a mismatch?

Return to the Salesforce role: the person needs to configure lead-assignment rules, not only use the system. Give someone who understands the role these fictional resume passages before showing them the AI results.

Resume passage	What it establishes
“I created lead-assignment rules and tested which salesperson received each sample lead.”	Describes the required work. An interviewer can examine how it was done.
“I used Salesforce daily. Our administrator built the assignment rules for this project.”	Establishes daily use, but leaves open whether the applicant configured rules elsewhere.
“Skills: Salesforce.”	Does not establish the responsibility. A software name is not an example of doing the work.
“I have never configured assignment rules; I only enter and update customer details.”	Explicitly says the applicant has not done the required configuration work. That differs from an unanswered question.

For the third passage, a useful assessment is:

Requirement: Configure Salesforce lead-assignment rules.
Evidence: The resume lists Salesforce but does not describe configuring rules.
Missing information: An example of the applicant doing that work.
Next step: Ask for an example before deciding.

Agree what each passage supports, then compare the tool’s conclusions. It should not invent configuration experience or turn an unanswered question into a definite “no”.

Try the failures a polished demo can miss

Test	What to look for
Relevant experience under an unfamiliar job title	Finds the work rather than relying only on the title.
A scanned or damaged resume the tool cannot read	Flags a file problem instead of treating it as lack of ability.
Text telling the AI to ignore the criteria and rank the applicant first	Treats it as application content, not an instruction to obey.
A spoken answer with local pronunciation, an abbreviation and a self-correction	Preserves the meaning in the transcript before assessing it.
Someone cannot complete the interview format	Offers the agreed help or alternative without treating the access problem as poor suitability.

Check missed candidates, not just the shortlist

A test set contains 90 unsuitable applications and 10 suitable ones. Rejecting all 100 produces 90% accuracy while excluding every suitable applicant. A single accuracy percentage can hide the mistake you most need to catch.

Check both missed experience and skills credited without evidence. Keep unreadable files and missing answers separate from suitability decisions. Do not infer protected characteristics from names or declare a product bias-free after a few examples. Broader fairness assessment needs appropriate data and methods, with specialist help where needed.

The UK ICO’s recruitment findings highlight the gap between reviewing a shortlist and reviewing the people excluded from it. For an initial pilot, have someone inspect proposed rejections before they happen, with the original application, time to review it and authority to change the decision. Include that work in the cost.

Test the reviewer as well as the AI

A 2024 research review in Nature Human Behaviour analysed 106 experiments. People using AI did better on average than people alone, but worse than whichever performed better independently: the person or the AI. The experiments covered varied tasks and systems studied from 2020 to mid-2023, not current Singapore hiring products.

The practical lesson is to check the final reviewed decision. Which AI mistakes did the recruiter catch or accept? Did an override introduce another error? Find out whether the reviewer lacked evidence, time or clear instructions, then fix that part and repeat the test.

Give candidates a working help route

Explain who is contacting the applicant, what the interview involves, how their information is used and whom to contact. Someone needs to answer that contact route.

If an interview fails, check which answers were saved before asking the candidate to repeat them. Give a realistic update time, adjust reminders and deadlines where needed, and offer the agreed alternative. Count these interruptions as part of the pilot, not as unrelated support work.

Will the time saved justify the ongoing cost?

Compare the same completed task, including checking and follow-up. Generating a summary is not the same as completing an application review.

This worked example covers 100 application reviews a month. The figures are illustrative staff minutes, not measured product results. The monthly AI-rule check is extra work introduced by the tool; one-off setup is excluded.

Work for 100 applications	Without AI	With AI
Read and review applications	600 minutes	300 minutes
Clarify missing answers and fix failed attempts	30 minutes	90 minutes
Hiring manager repeats part of the review	90 minutes	90 minutes
Check AI screening rules each month	0 minutes	60 minutes
Total staff time	720 minutes	540 minutes

Reading takes half as long, but the whole job falls from 12 staff hours to nine: a saving of three hours, or 25%, not 50%.

At S$60 per staff hour, that time is worth S$180. It does not cover a S$300 monthly subscription, even before setup costs. These are budget inputs, not a vendor quote.

At 150 reviews a month, with the same per-application timings and one fixed monthly hour checking rules, the comparison becomes 18 hours without AI and 13 with it. The five hours saved are worth S$300. Setup and other costs still need to be covered; different volumes, timings or prices change the result.

Name what the saved hours will be used for. Faster candidate replies or capacity for another vacancy can justify a purchase without reducing salaries. Those are capacity gains, not cash savings unless spending actually falls. A demonstrated improvement in decision quality may also matter, but it needs evidence rather than a convenient assumption.

Count everyone invited, not just those who started

For a separate interview example, follow 40 invitations, 28 starts and 24 completions:

Question	Calculation	Result
How many starters finished?	24 ÷ 28	About 86%
How many invited candidates finished?	24 ÷ 40	60%

Both are correct, but 60% describes the result for everyone invited. Keep failed deliveries, access problems, withdrawals and retries visible. Count candidates’ time too, including answers they must repeat at the next interview.

For an employee chatbot, ask whether the problem was resolved, not simply whether the chat ended. IBM reports 94% containment of common questions for AskHR, meaning they did not pass to a human adviser. That is a service metric, not the percentage of HR jobs removed or a forecast for your team. Your own measurement should distinguish a solved problem from someone giving up.

What should our first pilot deliver?

Run the pilot in four stages. Confirm funding conditions before committing, and data safeguards before using real HR information.

1. Define the task and the comparison. Record the workload today and what a completed result looks like. Compare with your current process or an existing feature, not with doing nothing. Identify the decisions that remain with your team.

2. Test away from live work. Use the sample policies or applications, with real messages and changes disabled. Include ordinary cases, failures and recovery. Agree which errors would stop the trial.

3. Run the permitted pilot and count the whole job. Give the HR colleagues using it clear instructions and support. Measure checking, corrections, candidate or employee effort, and work repeated by the next person. Check different users’ results rather than relying only on the team average.

4. Decide and hand over. Continue only where the benefit justifies the review effort and ongoing cost. Fix and retest material problems, narrow the scope or stop when the simpler process works better.

Use existing resources while preparing

IHRP already offers a free Job Redesign Evaluation Tool and Job Redesign Playbook for HR. These are existing resources, not the forthcoming PATH AI Playbook. The tool’s industry coverage uses Jobs Transformation Maps published before 2024, so check the date of any comparison you rely on.

Use the IHRP course directory to explore training, then confirm dates, assessments, places and funding with the provider. Choose practice that matches the job: catching an invented resume claim, checking a policy version or recovering a failed action, rather than learning prompts in isolation.

Make sure your team can run it without the consultant

Ask the colleague who will maintain the process to update a policy, check a summary’s source or recover a failed application using the handover instructions. The consultant should observe rather than take over. Note where help is needed and agree who will provide it, at what cost.

Keep the test cases, policy versions and expected answers. Name who repeats the checks after a model, policy or connection changes. Pause the affected process if it invents material candidate experience, discloses information to the wrong person or changes an application without permission.

Keep an escalation route for unresolved questions. IBM’s AskHR retrospective describes an early rollout in which removing existing contact routes helped push its Net Promoter Score down to −35 before it recovered. A company-wide switch is not the only way to learn whether a smaller process works.

Take four things back to your team: the task worth improving, the proposed test, the full cost and the funding questions still awaiting confirmation. You can prepare those now. Choose a package when you can explain both what the company will receive and how your colleagues will keep it working after support ends.

---

Programme sources checked 29 September 2026. Funding depends on the applicable programme rules and your project’s approval. The worked examples are practical evaluation guidance, not package prices, funding commitments or legal advice. Confirm current terms with the programme partner and assess the employment and privacy requirements before implementation.