The integration problem, not the intelligence problem: an assistant is only as useful as the systems it is allowed to reach.
The short version: To connect AI to business systems you already run — ERP, CRM, accounting — the interesting question is no longer "can AI write a decent email". It is "can AI see our stock levels". The answer is yes — through a standard called MCP — but the work is mostly permissions and testing, not the AI. And a good third of the requests we get are better solved without AI at all.
The question that comes up in every meeting now
A finance manager in Petaling Jaya asked us this in June, and it is close to word-for-word what we hear most weeks:
"We already pay for Claude. Everyone says it is clever. So why do I still have someone opening AutoCount, checking an invoice number, then typing the answer back into WhatsApp?"
It is a fair question, and the answer has nothing to do with how clever the model is. The model is not the constraint. Access is. A chat assistant with no connection to your systems is a very articulate colleague who has never been given a login.
Closing that gap is an integration problem. It is the same category of work as connecting your e-commerce store to your warehouse system — just with a newer standard, and a few governance steps that conventional integrations never needed.
What MCP is, in business terms
MCP — the Model Context Protocol — is an open standard published by Anthropic, the company behind Claude. Anchor Sprint is a member of the Anthropic Claude Partner Network, so treat what follows as informed but not neutral; we have a commercial interest in you deciding this is worth doing.
Strip out the engineering and it is a plug shape.
Before it existed, every connection between an AI and a business system was hand-built. Connect Claude to your CRM, then decide to try a different model, and you rebuilt the connection. Your vendor changed their API, and you rebuilt it again. Each connection was bespoke, and none of it was reusable — which is why so many 2024-era AI pilots quietly died at the integration step rather than the model step.
MCP standardises the shape of that connection. One agreed format, so any AI that speaks it can talk to any system that speaks it.
Three consequences matter commercially:
Integration work stops being throwaway. A connector built to the standard survives a change of model. What you paid for stays yours.
It reduces lock-in rather than adding it. This is the counter-intuitive part. Adopting a vendor's standard normally ties you to that vendor. Here, because the standard is open and other AI providers have adopted it, a connector to your ERP keeps working if you switch away from Claude entirely.
Permissions stay in your systems. The AI does not get its own set of keys. It acts within the access your existing roles already define — a salesperson's assistant sees what that salesperson can see. That single design decision is what makes this reviewable under Malaysia's PDPA and Singapore's PDPA as administered by the PDPC, rather than a compliance argument you would rather not have. Both regimes turn on purpose and consent for how personal data is used — far easier to evidence when the AI inherits an existing, already-audited access model than when it holds credentials of its own.
Nothing here is legal advice — treat it as a starting point for a conversation with your own counsel or DPO.
What connecting AI to business systems looks like in practice
Three patterns cover most of what Malaysian and Singaporean businesses actually deploy.
Answering from live data. A customer asks where their order is. The assistant checks the warehouse system and answers with today's status instead of yesterday's export. Read-only access, no ability to change anything, and the fastest of the three to justify.
Drafting work for a human to approve. A supplier invoice arrives as a PDF. The assistant reads it, matches it to the right purchase order, flags the two line items that do not reconcile, and prepares the entry. A person approves it. This is where most of the hours actually go in a finance team, and where the honest return is.
Acting inside a boundary you set. The assistant creates the delivery order once payment clears, because that rule is unambiguous. Anything outside the boundary stops and asks. Fewer businesses need this than think they do, and it is the pattern that most deserves a slow rollout.
Notice that two of the three keep a human in the decision. That is not caution for its own sake — it is what makes the system reviewable when someone asks why an entry was made.
The costs nobody puts on the slide
The model is the cheap part. Say that again, because vendor pitches consistently bury it: the model is the cheap part.
For a typical mid-sized deployment:
- The connector itself — a few weeks of engineering per system, less if the system has a decent API, considerably more if it is an on-premise install from 2011 with no API at all. Malaysian businesses running older AutoCount or SQL Account installs should budget for the second case.
- Permissions and governance — deciding who the AI may act as, what needs approval, what gets logged. This is usually the largest line, and it is organisational work as much as technical.
- Evaluation — testing against your real data, not a demo. An assistant that is right 90% of the time on invoices is not ready to touch a ledger, and you only discover that by measuring.
- The model usage — genuinely the smallest number for most businesses. Our Claude cost calculator will model it, and the figure usually surprises people by being lower than expected.
At Anchor Sprint, scoped projects run from RM 5,000 for a single connected workflow to RM 50,000 for a deployment spanning several systems with approval steps and audit requirements. The spread is almost entirely governance, not intelligence.
When you should not do this
Three situations where we tell people not to.
The step in the middle needs no judgement. If the requirement is moving the same fields between two systems on a schedule, a conventional API integration is cheaper, faster, and will not surprise you. AI earns its place when something in the middle requires reading, matching, or noticing — not when it requires copying. A scheduled sync is not a lesser solution; for that job it is the correct one.
Nobody owns the data quality problem. If your stock levels are wrong in the system today, connecting an AI to them produces confidently wrong answers faster. Fix the source first. This is the most common reason we tell a business to come back in six months.
The volume does not justify the governance. Forty invoices a month handled well by someone who knows the suppliers is not a problem worth RM 30,000 and a quarter of change management. Automate the thing that happens four hundred times, not four hundred times a year.
How to work out whether you are ready to connect AI to business systems
Four questions, in this order:
- Which system holds the answer people keep asking for? If you cannot name it in one sentence, that is the finding.
- Does that system have an API? If yes, this is weeks. If no, the first phase is getting data out of it, and the AI conversation comes later.
- Who is allowed to see this data today? The AI inherits that answer. If the honest reply is "everyone, because it is a shared login", fix that before adding AI — you would be adding a fast interface to an existing problem.
- How often does the task happen? Under about fifty times a month, be sceptical of the business case.
Where this is heading
The businesses that successfully connect AI to business systems in Malaysia and Singapore are not the ones with the best prompts. They are the ones that did the boring integration work — connected the systems, set the permissions, tested against real data — and then pointed a capable model at the result.
That is unglamorous, and it is also why the gap between companies that talk about AI and companies that use it keeps widening. The talking part got easy in 2023. The connecting part is the work.
If you want to know whether a specific workflow in your business is worth connecting, we do that assessment before any commitment — including telling you when the answer is no.
免费咨询
Which of your systems is worth connecting first?
Tell us the workflow that wastes the most time. We will tell you whether AI is the right answer, whether a conventional integration would do it better, and what it would cost either way.
Related reading: our system integration service for how we deliver this work · AI agents and automation for what runs on top of it · how AI agents get built in Malaysia · why we build on Claude. Or tell us the workflow and we will scope it.
References
- Model Context Protocol — specification and documentation (official)
- Anthropic, "Introducing the Model Context Protocol" (primary announcement)
- Protection of Personal Data — MyGovernment portal (Malaysia PDPA)
- Personal Data Protection Act — PDPC Singapore (Singapore)