AI & LLMs inside FileMaker - what actually works in 2026

Every client conversation in 2026 eventually turns to "can we add AI to our FileMaker system?" The answer is yes, it's easier than it was two years ago, and most of the demos you've seen online are technically accurate but commercially misleading. This guide is the honest version - the use cases that work in production, the ones that don't yet, what we deploy, what we charge, and the UK GDPR considerations you need to get right.

What Claris ships natively in 2026

Claris has invested meaningfully in AI-in-FileMaker since 2024. The native tooling as of the current FileMaker 2025+ releases:

• AI Model script step - configure a provider (OpenAI, Claude, Gemini, Ollama, Azure) once; call it from any script with a prompt + input.
• Embedding script steps - generate vector embeddings for any field or text input without leaving FileMaker.
• Vector search / semantic search - built-in semantic search against embedded fields, with a cosine-similarity score.
• Function calling / tool use - models can call scripts and return structured results for multi-step workflows.
• Privileged access controls - AI feature use can be gated per privilege set, so users with no AI budget don't rack up calls.

On top of this, any script can still call any LLM's REST API directly via Insert from URL - so if you need a model or feature Claris hasn't wrapped, you're not blocked.

Five AI use cases that actually work in production

1. Semantic search

This is the clearest win we see. Legal, clinical, CRM and knowledge-base use cases where users search across long-form notes fields ("find cases like this one") move from frustrating-literal search to genuinely useful. Typical delivery: 2–4 weeks on an existing system, £4,000–£10,000.

2. Document summarisation & extraction

Drop a PDF (contract, CV, inspection report, invoice) into FileMaker; an LLM extracts structured fields (counterparty, renewal date, total value, risk flags) with a confidence score. Human verifies. We see this live in procurement, legal, HR and insurance workflows. Typical delivery: 3–6 weeks, £8,000–£20,000.

3. Classification & routing

Inbound enquiries, tickets, referrals, leads - an LLM classifies them (sentiment, priority, routing team, likely resolution category) and updates the record. Cheap, high-volume, usually well-bounded enough to trust in production. Typical delivery: 2–3 weeks, £4,000–£10,000.

4. Draft generation (with human-in-the-loop)

First-draft quotes, response emails, case notes, executive summaries, meeting minutes. The operator always approves before send. This works; what doesn't work is removing the human from the loop - the quality and hallucination risk isn't there yet for autonomous send. Typical delivery: 2–4 weeks per workflow, £3,000–£8,000.

5. RAG-style Q&A over your FileMaker data

Chat interface inside FileMaker that answers natural-language questions grounded in a scoped subset of your data ("which suppliers missed SLA in Q1?", "what's open on the Millbrook account?"). We build these with embedded records + tool-use so the LLM cites the records it used. Typical delivery: 4–8 weeks, £12,000–£35,000.

What doesn't yet work reliably

• Fully autonomous agents that act on your database without human review. Too much hallucination risk, and the blast radius of a wrong write is high.
• High-stakes factual accuracy - compliance reports, medical advice, legal opinions - should not be LLM-authored without expert review.
• Long-context, multi-step orchestration - reliability drops sharply past 8–10 agentic hops in current models.
• Numerical analysis at scale - LLMs are poor calculators. Use Claris's SQL / aggregate functions for the maths; LLMs for the language layer on top.

Which LLM should you use?

AI + UK GDPR - the things you must get right

• DPA with the provider. No DPA, no sending personal data. Full stop.
• Zero-retention enterprise tier. Consumer ChatGPT / Claude accounts do log prompts for training; enterprise API tiers don't. Always use the enterprise tier.
• Lawful basis documented. Updating your ROPA (Record of Processing Activities) for AI processing is not optional.
• DPIA for special-category data. Health, legal, criminal, biometric → expect to do a DPIA.
• DSAR coverage. Log what went where; be able to answer "you held my data, who else did you send it to?".
• UK residency where required. NHS, regulated finance, some public-sector → Azure OpenAI (UK) or self-hosted.
• Privilege-gated AI features. Not every user should be able to rack up £££ of API calls.

For the broader compliance picture, see FileMaker Security & UK GDPR - 2026 checklist.

What a good AI-in-FileMaker deployment looks like

• One scoped, well-bounded use case delivered first (not "let's AI the whole system").
• Provider chosen with compliance in mind; DPA signed and on file.
• Prompt templates version-controlled in FileMaker (not pasted into scripts).
• Every AI call logged - input, output, model, cost, user, timestamp.
• Confidence thresholds enforced - low-confidence outputs go to a human queue, not into the record.
• Cost dashboard so the business sees what AI is costing, per use case, per month.
• Kill-switch: a privilege-gated flag to disable the AI layer globally if a provider has an outage or a compliance question.

The short answer

Adding AI to a FileMaker system in 2026 is a solved engineering problem for bounded, scoped use cases. Start with semantic search, document summarisation or classification - each delivers clear ROI in 2–4 weeks. Pick a provider with a UK-compatible compliance story. Log everything. Keep a human in the loop on anything high-stakes. Don't try to autonomise your whole business in one go - nobody is doing that successfully yet.

See also - FileMaker Security & UK GDPR, Claris Connect vs n8n vs Zapier, and Is FileMaker still worth using in 2026?