A personal 24/7 AI assistant for SME Centre advisors

This proposes a small, self-hosted AI assistant on one advisor's machine, at a running cost of about US$10 a month, to do three things: remove the meeting administration that consumes each advisor's day, turn the Centre's scattered client knowledge into one searchable record, and give the Centre an evidence base for events, workshops and the group-based upgrading work.

The problem it answers

• Advisors spend ten to fifteen minutes of administration on every client meeting: reading the chat, checking two calendars, offering slots by hand, creating the meeting, pasting the link, then reconstructing the client's history before the call.
• The only shared account of a case is what one advisor typed into PEP, and it may not capture what the SME actually wanted. A client who meets a different advisor starts cold, because PEP is normally opened after the meeting rather than before it, and the case summary and minutes are composed from scratch every time.
• The Centre cannot query its own client base: contacts, interests and pain points sit case by case, so workshop themes, group-based upgrading projects and EDM targeting are decided on judgement, and physical events pay for it in attrition.
• Mid-meeting, an advisor cannot answer what the rest of a client's sector is doing or what comparable companies have applied for.
• There is no AI SOP, so advisors are already using personal chatbots on confidential client material, which shares nothing back to the Centre and sends it outside the Centre's control.

What is proposed

• One open-source assistant (Hermes Agent) installed on an advisor's own machine, reachable from the chat tools we already use, plus one local meeting transcriber (Meetily) that never uploads audio.
• A consolidated client fact base, built from existing case history, held as ordinary readable files on that machine.
• A written AI standard operating procedure (SOP) covering approved tools, data classification, human review, meeting recording, retention, incident response and vendor assessment, adopted before any client data is introduced.

Why a decision is needed now

Client-facing AI is already in use informally. The choice is not whether advisors use AI at work; it is whether the Centre defines the boundaries and the approved tools, or leaves it to each individual's judgement.

In scope

• One assistant, one machine. An open-source assistant running on an advisor's laptop, with the client record held locally as readable notes.
• Meeting capture that stays local. Transcription and summarising on the same machine, so client audio is never uploaded.
• A consolidated client fact base built from existing case and advisory history: who the client is, what they applied for, when, what for, and who the contact is.
• Advisor-facing automation: meeting scheduling against two calendars, pre-meeting client briefs, follow-up reminders, a daily brief and an end-of-day summary.
• A governance layer: a written AI SOP, a data classification rule, an incident runbook, and a review cadence.

Explicitly out of scope

• No client-facing chatbot. Clients do not talk to this system.
• No automated decision-making about grants, eligibility or outcomes. The assistant drafts and retrieves; a human decides.
• No client personal data into consumer AI accounts, and no reliance on an advisor's personal subscription for work data.
• No replacement for the case ledger or any system of record. The fact base is an advisory aid, not the Centre's record of truth.
• No always-on listening, no recording of meetings without a stated notice, and no screen capture of client material.

None of what follows is a complaint about individual workload. The routine administration around every meeting is a real cost, and it is quantified in section 06. These five are different: they are gaps in how the Centre holds what it knows about its clients, uses it, and governs it.

1. The shared record of a case is thin, and it is written from the wrong end of the meeting

The only shared account of an engagement is what the advisor typed into the Partners' Engagement Portal (PEP) afterwards. That entry may not capture the SME's real interests, the issues raised in the room, or the solution the client actually needed. A client can sit with one advisor this month and a different advisor from the same Centre the next: until someone opens PEP, the second advisor starts blind, and PEP is habitually read after the meeting rather than before it. Preparation therefore happens backwards.

The writing is the other half of it. The PEP case summary is composed from scratch every time, tedious and repetitive, and so are the meeting minutes. This is the clearest single use case for an assistant at the Centre: draft the summary into the PEP format for the advisor to check and paste, rather than compose it.

2. The Centre has no view of its own client base in aggregate

Contact and interest information is captured case by case and never assembled: point-of-contact name, designation, phone and email; each company's stated interests and pain points; whether it has any appetite for overseas expansion. Across all of an advisor's meetings, that is the Centre's best market research. Left in individual files, it answers nothing.

Three decisions then rest on judgement instead of evidence:

• Business Community Engagement topics and themes for the coming year, which should be curated from what clients said they wanted.
• Group-based upgrading projects, which should be built where a pain point surfaces across many companies, not where a single case suggested it.
• EDM targeting, which should reach the companies that expressed interest in a given topic. Without that targeting, physical events pay for it in attrition.

3. Mid-meeting, an advisor cannot say what the sector is doing

Clients ask, in the room, what the rest of their sector is doing right now and what comparable companies have applied for. Answering well needs a view across every company in that sector and its scheme history. That view does not exist today, and nothing can assemble it inside a one-hour call.

4. AI is already in use, with no SOP and no boundary

There is no AI utilisation SOP, so no black-or-white line for advisors to work inside. Advisors are using personal chatbots for Centre work, which transmits confidential client information to a service outside the Centre's control. That is not a discipline failure; it is what happens when a useful tool arrives before a policy.

Those personal tools also cannot help the Centre as a whole. Each chat is private to one advisor, so what is learned in one is invisible to the rest. Making the knowledge shared requires a common harness in which each advisor is tied to a named bot, and every bot reads and writes the same Centre facts corpus, so one advisor's questions sharpen a record the next advisor can use.

5. Nothing collected is rolled up

Whatever the Centre accumulates, a cleaned and aggregated portion can be reported to SMF. Today there is no consolidation, so there is nothing to roll up. Any such reporting is aggregate and de-identified: individual client detail does not leave the Centre, which is a design rule rather than a courtesy.

Nothing here is exotic. It is ordinary software running on a laptop, with one deliberate design decision: the client material stays on that laptop.

Component	What it is	Where it lives
The assistant	Open-source agent software that runs scheduled jobs, reads files, searches the web, and answers in a chat app.	Local machine (free, open source)
The model	The AI that does the thinking, rented by usage from a provider.	Provider's servers; only the question and the answer cross the boundary
The memory	A folder of ordinary text notes: client facts, preferences, procedures.	Local machine, readable and editable by a human
Meeting notes	Local transcription and summarising of meetings, with no bot joining the call.	Local machine; audio never uploaded
The client record	One searchable note per client, assembled from existing case history.	Local machine, with an encrypted weekly archive
Reach	Responses delivered in the chat tools the Centre already uses.	A private connection, not a public web address

What leaves the machine, and what does not

Leaves the machine	Never leaves the machine
The question the advisor types. Web search queries. The specific text the model needs to answer that question.	Meeting audio and transcripts. The client fact base as a whole. Source documents and identity documents. Credentials, saved in an encrypted vault.

5.1 A consolidated fact corpus for the group-based upgrading work

The group-based upgrading project needs one reliable view of each participating company: registered name, UEN, industry, what they have applied for and when, the contact person, and the equipment or capability in scope. Today that view is assembled per advisor, per case.

• One record per client, not one per advisor. Each advisor's questions sharpen the same record, so the tenth advisor starts where the ninth finished.
• Cohort assembly on evidence. Build a list of companies that match a cohort's profile by industry, scheme history and recency, rather than by recollection.
• Continuity when people move. The record survives leave, team changes and departure, because it is an asset of the Centre rather than of an individual.
• Consistency of advice. The same client question asked by two advisors returns the same underlying facts.

5.2 Client pain points and interests, for events and workshop curation

Every engagement contains a sentence where the client describes what is actually hurting. Aggregated, those sentences are the Centre's best market research.

• Pain-point mapping by industry. What F&B clients say about manpower, what logistics clients say about systems, in their own words, with counts.
• Interest signals for events. Which schemes, technologies and topics clients ask about, and which questions keep recurring without a good answer.
• Workshop curation on evidence. Topic selection and sequencing driven by observed demand, with the source counts attached, rather than by assumption.
• Follow-up targeting. Companies facing a known deadline or renewal become a defensible invitation list rather than a mail-merge of the whole database.

Boundary that applies here: aggregated insight is what the Centre uses. Individual client detail is never published, and any external output is de-identified. Aggregation is computed on the local machine, not by a third-party analytics service.

5.3 The advisor's day

• A client asks for a slot in the chat; the assistant reads both calendars and offers real availability, then creates the meeting once the client chooses.
• A brief is waiting on the phone before the day starts, and a wrap-up at the end of it.
• Before a call: "what do we know about this company", answered in seconds, from the record rather than from memory.
• Renewals, unconfirmed meetings and open follow-ups surface themselves instead of depending on recall.

The measured running cost of the working system is US$10 a month, about S$13 at the rate on 1 October 2026. No per-seat licence, no new hardware.

Item	Monthly, USD	Basis
AI model usage (assistant)	10.00	US$10 a month for a plan that includes US$70 of model credit.
Web search	0.00	Free monthly credit (US$5). Beyond it, US$5 per 1,000 searches.
Assistant software	0.00	Open source, self-hosted.
Meeting transcription	0.00	Free community edition. Optional Pro tier at US$10 per user per month, billed annually, adds speaker labels and integrations.
Hardware	0.00	A Mac already owned, left running. Electricity only, no capital purchase.
Total	10.00	About S$12.80 a month

Measured consumption, for credibility

From the provider's ledger on 1 October 2026, for the current billing period: 2,046 runs, 172.5 million tokens, 4.41 of the 70 included credits consumed, 100% completion. The working system is using about 6% of what US$10 buys in a month.

What happens as more advisors join

Deployment	Monthly, USD	Yearly, USD	Note
One advisor, one machine (pilot)	10	120	What is proposed now
Ten advisors, machine each	100	1,200	Full independence, per-person data boundary
Ten advisors, one shared machine	10 to 30	120 to 360	Marginal cost of the tenth advisor is near zero; requires a shared-data decision
Rented equivalent, ten seats	200 to 300	2,400 to 3,600	ChatGPT Business at US$25, Claude Team at US$25, Otter Pro at US$17, per user per month

Assumptions, and what would change the number

• Model usage is metered. If the Centre ran heavy automation across ten advisors on one account, usage would rise; the plan's own ceilings (US$14 in any 5 hours, US$35 in any 7 days, US$70 a month) bound the worst case at US$70 a month per account, still below ten rented seats.
• Prices are in US dollars and change without notice; these were verified on 1 October 2026. No allowance is made for staff time to set up and maintain the system, which is the real cost of the self-hosted route.

Client personal data (names, contact details, NRIC in some documents) means the Personal Data Protection Act applies to whatever we build. The design principle here is subtraction: keep client material on the machine so that most obligations are removed rather than managed.

7.1 How the design answers each PDPA obligation

Obligation	What it requires	How this system answers it
Consent	Consent before collecting, using or disclosing personal data for a purpose.	Client data is used for the advisory purpose it was collected for. No use for model training. Cloud models are used under zero-data-retention terms.
Purpose limitation	Use only for purposes a reasonable person would consider appropriate.	The corpus serves advisory support, programme design and reporting. Access is scoped to that.
Notification	Tell individuals the purposes of collection and use.	Existing advisory communications cover the purpose. An AI-processing notice is added to the client-facing privacy statement before any centre-wide rollout.
Access and correction	Individuals may access and correct their data.	The record is plain files on disk, readable and correctable by hand. No request needs a vendor's cooperation.
Accuracy	Reasonable effort to keep data accurate.	Derived facts are traceable to source case records; a human reviews anything relied upon.
Protection	Reasonable security arrangements against unauthorised access, use, disclosure or loss.	Full-disk encryption, a login on the machine, no public exposure (a private tunnel only), an access key on services, credentials in an encrypted vault, no client detail in group channels.
Retention limitation	Stop retaining when the purpose is served and retention is no longer necessary.	Source submissions are purged after extraction. The derived corpus carries a defined retention period in the SOP. Weekly encrypted archive, with a stated retention horizon.
Transfer limitation	Comparable protection for personal data transferred overseas.	Local-first design keeps client material in Singapore on the machine. Where a cloud model is used, the transfer is limited to the question text and is governed by the vendor's zero-data-retention terms, documented in the SOP.
Breach notification	Assess whether a breach is notifiable (typically within 30 days of awareness) and notify the Commission within 3 calendar days of that assessment.	A single-machine surface is bounded and auditable. An incident runbook names who assesses, who notifies and within what clock.
Accountability	Designate a data protection officer and put policies and practices in place.	The AI SOP, a named system owner, and a quarterly review of logs, incidents and the approved tool list.

7.2 The Generative AI guidance

The PDPC, with IMDA, issued its Advisory Guidelines on the use of Personal Data in Generative AI on 20 July 2026. Three points matter here:

• System deployers carry primary responsibility for PDPA compliance, including defining clear purposes, safeguarding data flowing through the system, and regularly reviewing safeguards, with explicit attention to agentic AI.
• Where the position is unclear, document the assessment. A Data Protection Impact Assessment (DPIA) is the expected record. This proposal treats a DPIA as a gate before any client data enters the pilot.
• If personal data is used to train or fine-tune a model, individuals must be told specifically. This system does not train or fine-tune on client data, which is why zero-data-retention model terms are a hard requirement rather than a preference.

Every control leaves a residue. These are the honest ones, with what we do about each.

Risk	Why it exists	Mitigation
The model provider sees the question	Whatever the advisor types is sent to a model endpoint. A question containing a client name is a disclosure.	Zero-data-retention terms on every endpoint; a written rule that client identifiers are not typed into prompts; the client record answers by reference to the local note, not by pasting client detail into a prompt.
The machine is the single point of failure	One laptop holds the working record.	Encrypted weekly archive, an off-machine backup of text, and a rehearsed restore. The record is designed to be reconstructible.
The machine is the single point of exposure	Everything sits on one device.	Disk encryption, login required, no public address, services behind an access key, credentials in a vault, automatic re-closing of any temporary open lane.
Advisor error in what is asked	An advisor may paste client detail into a prompt.	Training at onboarding, a visible prompt convention, and periodic spot checks in the quarterly review.
Derived facts are wrong	Extraction from documents can misread.	Facts are traceable to sources; nothing client-facing is sent without human review; corrections are made in the note itself.
Vendor change of terms	Model providers change pricing, retention and training policies.	The provider is a swappable component by design. Annual review of terms, with a stated fallback provider.
Scope creep	Capability invites mission creep toward client-facing automation.	Written out-of-scope list, revisited only by management decision, not by default.

The technology is the easy part. What determines whether this ends well is whether the Centre writes down the rules before, not after. Ten items, all of them short documents.

Policy	What it must say	Owner
1. Approved tools	The list of AI tools permitted for work, and which data class each may handle. Everything else is unapproved by default.	Management
2. Data classification	Four buckets (public, internal, client-confidential, restricted) with an explicit rule for each: what may enter a prompt, what may be stored, what may never leave the machine.	DPO + Management
3. Acceptable use	No client personal data into consumer accounts; no personal subscriptions used for Centre work; no client material in unsupervised tools.	Management
4. Human in the loop	AI output is a draft. A named human approves anything client-facing, financial, or interpretive. The advisor remains accountable for the advice given.	Team leads
5. Meeting capture	Participants are told when a meeting is recorded or transcribed, how the transcript is stored, and how long it is kept. Recording is opt-out-able on request.	DPO
6. Retention and deletion	A retention period per data type, a deletion method, and a rule that source documents are purged once extracted.	DPO
7. Access and credentials	Who may access the system and its record, how keys are held (encrypted vault), device encryption mandatory, and how access is revoked on team change.	IT + DPO
8. Incident response	How a suspected breach is assessed, who decides whether it is notifiable, the 3-day notification clock, and the notification template for affected individuals.	DPO
9. Vendor and model assessment	A checklist before any model or tool is approved: zero-data-retention terms, training opt-out, data residency, sub-processors, and contract terms.	DPO + IT
10. Training and review	Every advisor signs the AI use policy at onboarding; annual refresher; quarterly review of the approved list, logs and incidents.	Management

The four rules that apply from day one

• Client identifiers do not go into a prompt. Ask by reference to the local record instead.
• Nothing client-facing is sent without a human reading it first.
• Meeting capture is announced to participants, and local capture is preferred over any cloud service.
• If in doubt, it stays on the machine. Escalate the question, not the data.

Six weeks, one machine, four gates. Each gate is a decision to continue, not a formality.

Phase	Duration	Activity	Gate to pass
0. Foundation	Week 1	Management approves scope and the out-of-scope list. Draft SOP v0.1 and the data classification rule. Name the system owner and the DPO contact.	SOP and classification signed. No client data involved yet.
1. Dry run	Week 2	Install on one machine. Run with public and internal material only. Time the meeting-administration workflow and record a baseline.	System stable; baseline measured; zero incidents.
2. Client data, bounded	Weeks 3 to 5	Complete a DPIA covering the intended processing. Introduce one grant programme's client corpus. Local meeting capture under the stated notice. Produce one aggregated insight pack (industry and pain-point counts).	DPIA completed; no notifiable incident; corpus accuracy spot-check passed; aggregation produced without exposing individual client detail.
3. Decision	Week 6	Evidence pack to management: time saved per meeting, corpus quality, incidents, cost against budget, and the recommendation on per-advisor machines versus a shared one.	Management decision on rollout scope.

What success looks like, stated so it can be measured

• Time: meeting administration per client falls from ten to fifteen minutes to under three, measured over at least twenty meetings.
• Knowledge: a pre-meeting client brief takes under a minute to retrieve, and contains the scheme history correctly for at least 90% of a sampled set.
• Insight: one aggregated pack produced from existing history that the team could not have produced by hand in the same time.
• Governance: zero notifiable incidents, and every meeting captured was announced.
• Cost: total running cost within the US$10 to US$30 a month band.

What we need from management

• Approval to run the pilot on one advisor's machine, with the SOP in place before client data is introduced.
• A decision in principle on the eventual model: a machine per advisor (per-person boundary, roughly the cost of one rented seat each) or a shared Centre machine (near-zero marginal cost, shared record, one governed boundary).
• Nomination of the DPO contact and the system owner.
• Agreement that the pilot's output is a single evidence pack, and that no capability expands beyond the written scope without a fresh decision.

Sources

Subject	Source
PDPA obligations	PDPC Advisory Guidelines on Key Concepts in the PDPA (consent, purpose limitation, notification, retention, protection, transfer, breach notification); PDPA Parts 3 to 6A.
Generative AI and personal data	PDPC and IMDA, Advisory Guidelines on the use of Personal Data in Generative AI, issued 20 July 2026; PDPC Advisory Guidelines on use of Personal Data in AI Recommendation and Decision Systems.
Breach notification clocks	PDPA Part 6A: assess as soon as practicable (typically within 30 days of awareness); notify the Commission no later than 3 calendar days after the assessment that a breach is notifiable.
Assistant software	Hermes Agent documentation: installation, messaging platforms, Bot Mode. Open source.
Meeting transcription	Meetily (MIT licence): local Whisper and Parakeet transcription, local or bring-your-own-key summarising, community edition free, Pro at US$10 per user per month billed annually. Compare with local-first alternatives and with cloud services including Otter, Fireflies and HappyScribe.
Cost figures	The model provider's own plan and usage ledger, read 1 October 2026. Exchange rate US$1 = S$1.279 on the same date.
Technical guardrails	Full-disk encryption on the host machine; private tunnel instead of public exposure; access key on hosted services; credentials in an encrypted vault; a scheduled exposure check that asserts the surface every 15 minutes.

Companion material

A separate Business Advisor session covers the practical half: the seven assistants compared, the four bot platforms compared, six transcription tools compared, and copy-paste setup guides for the assistant, Telegram, Slack and the meeting transcriber.

Prepared by Lewis Ang, Senior Business Advisor, 1 October 2026. Prices, features and legal references verified on that date. This document contains no client names and no client data.