AI summary in the CRM
The whole relationship, read in ten seconds.
Conversations become a summary, structured fields and a proposed next step, confirmed by the salesperson.
Put AI inside your business. Not beside it.
We connect models from OpenAI and Claude to your CRM, documents and workflows, so AI works on real business events. Each assistant sees only what its user may see, logs every step, and waits for a person before anything reaches a customer.
- Models
- OpenAI · Claude
- Control
- Human approval
- Every step
- Logged
What we build
Twelve capabilities. Four jobs AI actually does.
Every capability runs inside a system you already use, a CRM, an inbox, a website or an internal tool, not in a separate chat window.
Assist your teamAI assistants · Knowledge assistants · Summaries
Less searching, less retyping.
AI assistants
Inside the tools your team already uses, not a separate app.
“Draft a quote from this thread”
Internal knowledge assistants
Answers from your policies, manuals and past work, with sources.
“What's our group refund window?”
Summaries
Long threads, calls and documents reduced to what matters.
“Summarise this client's last month”
Serve your customersSupport replies · AI search · Recommendations
Faster answers that stay accurate.
AI customer support
Replies drafted from your knowledge base; agents review and send.
Suggested reply · 2 sources
AI search
Site and catalogue search that understands what people mean.
“quiet hotel by the sea, two kids”
Recommendations
Relevant products, packages or next steps from real history.
Customers on this route also booked…
Automate the workAI agents · Document handling · Workflow steps
Repetitive steps handled, with approval.
AI agents
Multi-step tasks across systems, within the permissions you set.
Check stock → update order → notify
Document handling
Emails, forms and documents classified, read into fields and routed.
Invoice PDF → fields → ERP
Workflow steps
AI steps added to the workflows you already run.
New lead → score → assign
Understand your dataAI in the CRM · AI analytics · Data analysis
Questions answered in plain language.
AI in the CRM
Lead summaries, enriched records and suggested next steps.
Next step: send family quote
AI analytics
Ask your reports questions instead of building new ones.
“Which branch grew most this quarter?”
Data analysis with AI
Patterns, anomalies and categories across large datasets.
Tag every ticket by topic
Where AI plugs in
One AI layer. Every system can use it.
Instead of a separate bot per department, we build one governed layer between your data and your applications, so permissions, logs and models are managed in one place.
01 · Company data
- CRM & customer records
- Documents & policies
- Conversations · WhatsApp, email
- Orders & bookings
- Product & service catalogue
- Reports
02 · AI layerprivate · permissioned · logged
Shared by every system
- Connectors & retrievalsearch · embeddings
- Model routingOpenAI · Claude · other
- Prompts, tools & actionsAPI · webhooks
- Permissions & guardrailspolicy
- Evaluation & audit logquality · cost
A person approves customer-facing and irreversible actions
03 · Where it works
- CRMlead summaries
- WebsiteAI search
- Support inboxdraft replies
- Salesnext step
- Reportsplain questions
- Internal toolsknowledge assistant
Connect, don't copy
Data is read from the source system when a request comes in, under that system's permissions.
One audit trail
Every question, source, answer and action is recorded in one log you can review.
Swap models, keep the system
Applications talk to the layer, not to a vendor, so changing a model is a configuration change.
In practice
What it looks like on a Tuesday morning.
Three examples of AI working inside everyday screens. Illustrative interfaces with sample data.
01 / 03 · swipe
Models & platforms
Model-agnostic. Chosen per task.
We integrate OpenAI and Claude through their APIs, and other API-based models where a project needs them. Each task gets the model that does best on your own test cases, measured rather than assumed.
OpenAIOpenAI API
GPT models for drafting, extraction, classification and summaries.
ClaudeAnthropic API
Claude models for assistants, analysis and work over long documents.
Other API-based modelssame interface
Added behind the AI layer when cost, language or data rules call for it.
ai-layer / routingSample configuration
| Task | Model | Chosen by |
|---|---|---|
| support.draft_reply | Claude | Chosen by: Tests on past tickets |
| crm.summarise | OpenAI | Chosen by: Quality against cost per call |
| kb.answer_cited | Claude | Chosen by: Citation accuracy on a test set |
| inbox.classify | Small, fast model | Chosen by: Speed on live traffic |
| fallback | Second provider | Chosen by: Availability |
Trust & control
AI you can switch on, and account for.
The controls are part of the architecture from the first workshop. We design them with your team and document them, so you know what the system can see and do.
Audit log · run 4821Sample
| 09:42:08 | retrieve | crm.booking 20417 |
|---|---|---|
| 09:42:09 | model | draft_reply |
| 09:42:10 | check | price_check · pass |
| 09:43:31 | approve | sales · edited 1 line |
| 09:43:32 | action | crm.note.create |
Permissions
The AI sees only what the person asking is allowed to see. Access follows the roles already defined in your systems.
role-based · per source
Data boundaries
We define which data may reach a model, what is masked before it leaves your systems, and which provider processes it.
masking · per field
Logging & audit
Questions, sources, outputs and actions are logged, so any answer or change can be traced back.
per-request trail
A person approves
Customer-facing output always waits for a person. Approval on internal steps, like CRM notes and tags, is dropped only where the log shows it is reliable.
approval rules per action
Evaluation
Test sets built from your real cases measure quality before launch, and again after every prompt or model change.
baseline · regression
Permissions
The AI sees only what the person asking is allowed to see. Access follows the roles already defined in your systems.
role-based · per source
Data boundaries
We define which data may reach a model, what is masked before it leaves your systems, and which provider processes it.
masking · per field
Logging & audit
Questions, sources, outputs and actions are logged, so any answer or change can be traced back.
per-request trail
A person approves
Customer-facing output always waits for a person. Approval on internal steps, like CRM notes and tags, is dropped only where the log shows it is reliable.
approval rules per action
Evaluation
Test sets built from your real cases measure quality before launch, and again after every prompt or model change.
baseline · regression
Audit log · run 4821Sample
| 09:42:08 | retrieve | crm.booking 20417 |
|---|---|---|
| 09:42:09 | model | draft_reply |
| 09:42:10 | check | price_check · pass |
| 09:43:31 | approve | sales · edited 1 line |
| 09:43:32 | action | crm.note.create |
How we engage
Start small. Prove it. Then scale it.
Each phase ends with a decision you make with evidence, so you only invest further when the pilot has earned it.
01 / 04 · swipe
Case study
AI inside a travel agency's CRM.
Nazarene Tours · CRM
An assistant and a document pipeline on the agency's own data
- Assistant
- A tool-using assistant over the CRM's customer, sales, support, pricing and marketing data.
- Documents
- Monthly airport statistics reports in Hebrew, read page by page by a model against a strict schema.
- Checks
- Every extracted table is validated against the totals printed in the report before it is stored.
FAQ
Questions about AI projects.
Will our data be used to train the models?
Not by us, and under the providers' current terms, the business APIs we use don't train models on the data sent to them by default. We confirm those terms with you before launch, send only the fields a task needs, and can mask personal details before they leave your systems.
What happens when the AI gets something wrong?
It will, sometimes. That is why customer-facing output waits for a person, why answers point to the source they came from, and why every run is logged. Mistakes found in review go into the test set we check every change against.
Is this just a chatbot on our website?
Usually not. Most of the value is inside the work: drafts for your support team, summaries in the CRM, answers for staff across internal documents. A public chatbot is possible, built on the same rules and data access.
Which model do you use?
Models from OpenAI and Claude, chosen per task by testing them on your examples. Cost, speed and answer quality differ from task to task, so one project can use more than one model, and the integration lets us swap a model later.
How do we know it works?
Before launch we agree on a set of real cases and the correct answer for each. Every change runs against that set, and after launch the review log shows how often people edit or reject an output.
Start with one process
Bring one process. We'll show you where AI fits.
A discovery workshop with our engineers, and an honest view of where AI will help and where it won't.
Useful to bring
- A process that eats your team's time
- A few real examples: emails, tickets, documents
- The systems involved and who approves what