AI Agents Are Coming for Busywork: What 'Agentic AI' Means for Small Businesses

Why Everyone Suddenly Says 2025 Is the Year of AI Agents

Somewhere in Biñan this week, an owner scrolls holiday headlines about AI agents for business while, two desks away, the bookkeeper re-encodes December sales into the spreadsheet she has retyped every January for years. This post translates between them: what "agentic AI" means, which busywork it can honestly take over in 2025, and what deserves a line in the budget you approve this quarter.

The narrative does have footing. On 11 December 2024, Google introduced Gemini 2.0 as "our new AI model for the agentic era" — Gemini 2.0 Flash Experimental for developers in AI Studio and Vertex AI, plus a Deep Research feature. That is why "agentic" is in every January headline.

The number behind the hype needs a caveat. Deloitte Global's TMT Predictions 2025, published 19 November 2024, forecasts that 25% of enterprises already using generative AI will deploy AI agents in 2025, and 50% by 2027 — enterprises already running generative AI, not a quarter of Philippine SMEs.

Our view on AI agents for business: the direction is right, the timeline wrong for most SMEs here. The useful move this January is not "buy an agent" — it is finding which repeating task is a plain workflow you could have automated three years ago.

Agentic AI Explained: A Workflow Follows Steps, an Agent Decides Them

One distinction carries the topic. In "Building effective agents", published 19 December 2024, Anthropic separates them: workflows are "systems where LLMs and tools are orchestrated through predefined code paths"; agents are "systems where LLMs dynamically direct their own processes and tool usage, maintaining control over how they accomplish tasks." (An LLM, or large language model, is what powers ChatGPT, Gemini, and Claude.)

On your floor, a workflow is: a Viber order arrives, pull the item and quantity, add a row to the orders sheet, send a confirmation — same path, a thousand times. An agent is: "chase the five overdue invoices this week," and it picks who to email, when to follow up, when to escalate. That second one is where the risk lives.

	Chatbot	Workflow	Agent
Touches your systems	No	Yes, fixed rails	Yes, picks which
Multi-step, unprompted	No	Yes, set order	Yes
Chooses the order	No	No — you did	Yes

A tool that drafts a reply you copy and paste is not an agent. It is a useful assistant — see our practical do's and don'ts for using ChatGPT at work.

Most SME busywork is deterministic: same inputs, same outputs. That makes it a workflow problem, and workflows are cheaper, testable, and auditable. Reaching for an agent where business process automation services would do is how automation budgets get spent without anything changing.

The 'agent' label is about to get stretched

Through 2025, expect rules-based tools and single-prompt chatbots re-badged as agents in sales decks. Three questions cut through:

• What decisions does it make on its own, with no person in the loop?
• Which systems can it write to, not just read from?
• When it gets something wrong, who finds out, and how fast?

How Would an AI Agent Even Reach Your Business Data?

AI agents for business need your data, and a model cannot see your sales sheet, your inventory, or your inbox. Something has to connect it — and until recently every connection was custom-built, one at a time. That is much of why "AI automation" quotes came back expensive.

That plumbing is standardizing. On 25 November 2024, Anthropic open-sourced the Model Context Protocol (MCP), "a new standard for connecting AI assistants to the systems where data lives, including content repositories, business tools, and development environments," with the spec, SDKs, and pre-built connectors for Google Drive, Slack, GitHub, and PostgreSQL. Treat it as early plumbing that lowers integration cost over time, not something you install this month.

Now the audit. SME data here sits in Google Workspace or Microsoft 365, Viber and Messenger threads, an Excel or Sheets "system," a POS, sometimes QuickBooks or Xero — and plenty of it only in one senior staffer's head. Point a model at undocumented data and it will just be confidently wrong, faster.

So the pre-work is the work: document the process, get the data into one queryable place, fix access. That is also what makes custom software built around your actual process, or any later integration, cheaper to build.

Five Kinds of Philippine SME Busywork Worth Automating First

Here is the triage: each item is either a workflow to automate now with boring tools, or genuinely agent-shaped and worth a careful pilot.

1) Compliance data assembly — WORKFLOW. January is year-end and Q4 season: annual information returns on withholding, the annual inventory list, and the encoding and reconciliation feeding the annual income tax cycle. Most of that pain is gathering and reformatting numbers that already exist. Automate the assembly; filing and sign-off stay human.

2) Order and inquiry intake from chat — AGENT-SHAPED, start as a workflow. Pulling item, quantity, and address out of Viber and Messenger threads is high-volume, high-error work. Start with extraction plus a human confirm step.

3) Quotation and proposal drafting — WORKFLOW with AI drafting. Pull numbers from a real price list, let the model draft, have a person review and send. Never let a model invent prices. It will.

4) Payroll and timekeeping reconciliation — WORKFLOW. Matching biometric logs against payroll and flagging exceptions is rule-based, and the exceptions are what humans should see. You already pay overtime for this by hand.

5) Supplier and inventory follow-ups — AGENT-SHAPED. Sequence-heavy but light on judgment — a reasonable first pilot, as long as nothing it does is irreversible.

The selection rule: high-frequency, low-judgment, reversible. If a mistake costs money you cannot claw back, it is not your pilot.

Where to Draw the Line: Payments, Personal Data, and Human Sign-Off

The hard rule first: nothing automated touches money movement without a human approval step — no disbursements, no supplier payments, no edits to payout details. For era context, Republic Act No. 12010, the Anti-Financial Account Scamming Act, was approved 20 July 2024. It does not regulate AI, but it shows the direction.

Personal data is the second line. Under Republic Act No. 10173, the Data Privacy Act of 2012, you remain the personal information controller — pasting customer lists, employee records, or ID scans into a third-party AI tool does not move that obligation to the vendor. Know where the tool stores data, keep personal data out of prompts, and check your privacy notice covers this.

There is no AI statute here as of early 2025, only policy direction. The DTI's National AI Strategy Roadmap 2.0 and Center for AI Research, launched 3 July 2024 with Asian Development Bank support, set seven strategic imperatives including upskilling and reskilling — a signal about training money, not a compliance checklist.

Write four guardrails into any pilot before it starts:

• A hard scope — which systems it may read, and which it may write to.
• An approval gate on anything external-facing or financial.
• A full log of every action taken, timestamped.
• A named human owner, not a committee.

An agent you cannot audit is an agent you cannot defend — to a client, to the National Privacy Commission, or to your own board.

What AI Agents Should Cost You in 2025 — and Where They Fit the IT Budget

You are inside a Q1 approval cycle; our take on where Philippine SMEs should put their 2025 IT budget covers the rest of the sheet. The cost stack has four parts: subscription seats, integration work, internal time to document processes, and someone checking outputs — the last is what most quotes leave out. No verified Philippine benchmarks exist, so treat any peso figure you are shown this year as illustrative.

One cost few owners price in: Republic Act No. 12023, approved 2 October 2024 and effective 18 October 2024, imposes VAT equivalent to 12% of gross sales on services including digital services, and covers nonresident digital service providers serving Philippine consumers. Foreign AI and SaaS subscriptions get more expensive. The implementing rules were still pending as we wrote this: the law allows the Department of Finance 90 days to issue them, with nonresident providers covered 120 days after they take effect (Grant Thornton Philippines). We covered RA 12023 and who actually pays the new digital services VAT in October.

So fund one pilot, not a program: a single process, a fixed time box, and an agreed "what would make us stop" condition before money is approved. The failure pattern is buying seats for everyone in January, then finding in March that nothing changed.

The cheapest win costs nothing. Time the task for two weeks as it is done today, then measure again after automating. Without that baseline you cannot tell whether the subscription paid for itself.

AI Agents for Business: Quick Answers Before You Commit

Is an AI agent the same as the chatbot on my website?

No. A chatbot answers questions; an agent takes multi-step action in your systems and decides the order of those steps. Most "AI" on Philippine SME websites in early 2025 is a chatbot, and that is fine.

Do I need to replace staff to see a return?

No. The realistic 2025 return is removing re-encoding and copy-paste from roles you already have, so the same people handle more volume — which is where the roadmap's upskilling emphasis points.

Will this work if my records live in Excel and Viber?

Partly. Extraction works today, but reliability tracks how structured your data is. Cleaning up where your data lives is the prerequisite, and it pays off either way.

What is the safest first project?

One high-frequency, low-judgment, fully reversible task, with a human approving anything that leaves the building.

The highest-return move this quarter is not buying an agent. It is picking one high-frequency, low-judgment, reversible task and learning what it costs to run by hand today. If that would help, we sit down with Philippine SMEs to map which repeating processes are workflow problems and which are genuinely agent-shaped, and what each would take to build. Book a call whenever you are ready.