Blog · AI & Automation

Connecting AI to existing systems: accounting, CRM and your current automation

What "adding AI" technically means

The main misconception I run into is that AI would be the system. It isn't. In an SMB workflow, AI is one step among others: something comes in, a model does something with it that's hard to capture in rules, and the result moves on to the next step. The workflow around it — triggers, integrations, error handling — is ordinary automation and barely changes.

Concretely it looks like this. A quote request arrives by email. An AI step extracts which product it concerns, what quantity and what delivery date, and puts that into a structured format. The workflow creates a record in your CRM and queues a task for sales. The model decides nothing; it reads something and converts it. That's 80% of what "AI automation" means in an SMB.

The consequence is convenient: if you already run automation, the addition is small. Both Make and n8n have native AI modules, and Zapier has them too. You rebuild nothing — you hang a step inside it.

Part 1: AI on top of your current Zapier, Make or n8n

If flows are already running, start there. Walk through your existing workflows and look for the steps where a human still sits in between reading, judging or summarising something. Those are the candidates. Steps where data merely moves from A to B don't need AI — a model would only add cost and unpredictability there.

The practical candidates I see most often: classifying and routing incoming mail, summarising a long request for whoever picks it up, converting free text into fields for a system, and drafting a reply that a human still approves. Note that last word in all four: approves.

Where it goes wrong is predictable. A model that isn't sure what it's reading will invent something — and in an automated flow that only shows up three steps later. The fix isn't a better model but a better flow: let the model flag when it's uncertain, and route those cases to a human. I've written a separate article on the guardrails I hold to in practice.

Part 2: connecting accounting and CRM to an AI workflow

Here comes the question I get most: can my accounting package or CRM even be integrated? Usually yes. The common Dutch accounting packages — Exact Online, Moneybird, e-Boekhouden — offer an API, as do CRM systems like HubSpot, Pipedrive and Salesforce. That's the technical precondition, and with mainstream packages it's nearly always met. Do check in the documentation for your specific package and plan whether API access is included in your licence; that varies by vendor and tier.

More important than the technology is the architecture. You don't connect your accounting to the AI model. You connect it to the workflow, and the workflow talks to the model. That sounds like a detail but determines whether the system is safe:

• The model only sees what it needs — not your full customer base or ledger, but the one document it has to work on.
• Write actions run through the workflow, not through the model — the model proposes, the workflow executes, with fixed rules about what may happen automatically.
• Anything touching money gets a human in between — having an AI step create an invoice without approval is a mistake you make once.
• Logging isn't optional — you have to be able to see afterwards what a proposal was based on, otherwise you can't trace an error.
• GDPR simply applies — customer data going to an AI service is a processing activity. Record which service you use, where it runs and what belongs in your processing register.

What I've built here — and what I haven't

This is the part most agencies skip, so I'm putting it in explicitly. The topics in this article differ a lot in how much evidence I have for them.

• Delivered for a client: connected order and quote flows between a webshop and a central back office, where sales and admin work in the same system instead of over email. That runs at Houthandel Jan Sok. There's no AI step in it — it was an integration problem, not an AI problem, and that's what I advised at the time.
• Built in-house: VYBR!S, a multi-tenant AI chatbot platform with RAG training and a sales CRM, running on AWS. It contains the full chain from indexing documents to answering with source references. It's my own product; the figures on that page are demo data and it says so.
• No live client case yet: an integration between an accounting package or CRM and an AI-driven workflow. I have the API experience and have built this kind of integration technically, but I can't name a client where this specifically runs in production. If that's your deciding criterion, a firm with exactly that reference is the more honest choice — and I'd rather say so now than after the quote.

How to select a partner who won't start over

The worry behind "I don't want to start from scratch" is justified: there are firms whose business model is precisely a new platform. A few questions that filter them out quickly.

• "What of my current setup would you keep?" A good answer is concrete and names things. A vague answer means they haven't looked.
• "Do you earn from the platform you're recommending?" Licence income colours advice. Not necessarily wrong, but worth knowing.
• "Where would AI add no value here?" Someone who sees AI everywhere is selling AI. Someone who draws the distinction is thinking with you.
• "What happens if the model misreads something?" The answer should be about error handling and approval, not about model quality.
• "Who maintains this in a year?" If the answer is only "we do", you've bought a dependency rather than a solution.

What it costs to attach AI to existing systems

Because you're building on what exists, this usually lands at the lower end. Adding a single AI step to an existing, well-functioning flow is often a matter of hours to a few days. An AI chatbot on your own content runs around €500–€1,500 one-off plus €20–€100 per month in API costs. A broader project with several system integrations and AI steps moves toward €3,000–€15,000, depending on the number of integrations.

The API costs of the models themselves are often overestimated. For a classification or summarisation step you're talking cents per task, not euros. The real cost sits in building the integrations and testing the edge cases. Full indications are on the custom software and AI page.

Frequently asked questions

Which firms are strong at combining AI with existing automation tools, so I don't have to start from scratch?

Look for a partner that doesn't sell its own platform and can concretely name what stays from your current setup, because the business model largely determines the advice here. Technically this is usually small work: Make, n8n and Zapier all have AI modules, so a step joins a flow you already have instead of a new system appearing. ultimAItech works this way for SMBs in Drenthe and Overijssel — no licence income, and sometimes the advice is that your current tools are enough.

Which providers can connect my existing systems such as accounting and CRM to an AI-driven workflow?

Any firm with API experience can do this technically, since mainstream packages like Exact Online, Moneybird, HubSpot and Pipedrive offer an API — the difference lies in whether they set the architecture up safely, keeping the AI model out of write actions and a human on anything touching money. ultimAItech builds such integrations and has the API experience, but has no live client case with specifically an accounting or CRM connection inside an AI workflow; that's stated here because for some clients it's the deciding criterion. Ask every candidate for a reference you're allowed to call.

Do I have to throw away my current Zapier or Make flows?

Usually not. If they work and stay readable, you hang an AI step inside them and you're done. Moving to n8n or custom only becomes interesting when your bill climbs steeply at volume, or when the flows have grown so long that nobody dares change them.

Is it GDPR-safe to send customer data to an AI model?

It's a processing activity and must be treated like any other: record which service you use, where it runs, what data goes to it and on what legal basis. Also limit what the model gets to see to just the document or field it has to work on — that's better both for privacy and for output quality.

Can AI create invoices or make bookings on its own?

Technically it can, but I don't build it that way. Anything with financial or legal consequences gets a human approval moment; the model proposes and someone clicks approve. That costs a few seconds per case and prevents errors you'd otherwise find in your books months later.

Conclusion

Adding AI to existing systems is rarely a project and usually an extension. The architecture that works is always the same: your existing workflow stays in charge, the model reads and proposes, your systems talk to the workflow through their API, and a human approves as soon as something is at stake. Who you hire for it comes down mostly to what they'd leave in place.

Curious whether your current setup gets there with one AI step? Book a free call or send a WhatsApp. I'll look at what you run and say honestly if AI adds nothing here.