AI integration services: connected to your systems, kept in scope
AI integration services connect a language model through an API to a business's CRM, inbox, WhatsApp and document systems. ALTAI Digital works with the models behind ChatGPT and with Anthropic's Claude models. Existing systems stay in place; the model is added as a layer on top. It answers only from the sources it is given and says so when a source is missing. Sending, paying and deleting each keep a separate human approval step. Discovery takes one week; the first live flow follows in four to eight weeks.
A business runs its CRM, inbox, WhatsApp line and accounting software as separate tools. Data moves between them by hand, and some steps get missed along the way. AI integration services place a language model on top of these systems and close that gap. ALTAI Digital connects the models behind OpenAI's ChatGPT and Anthropic's Claude models through an API. Which model is used follows the task and the data rules, and the reasoning is written down. Existing systems stay in place; the model is added as a layer above them, not a replacement for them.
The work starts by agreeing which system is the source of truth for each type of data. Your CRM, ERP, accounting software, inbox, WhatsApp and Drive stay in place; none of them is replaced or switched off. Where an API exists, data moves through it. Where it does not, a bridge is built instead: an export file, a database read or a screen level flow. A transfer can run one way or both ways; a two way link only opens once a conflict rule is written down. Every transfer is logged, and a failed one is queued, reported and retried.
The model works only from the sources it is given. When a question has no answer in those sources, it says so instead of guessing. Only the question itself and the matching passage leave the source system; the full document is never sent out. Documents are not used to train the model, and access follows the same permissions as the source system. Human approval is a separate step for sending, paying and deleting, since none of these can be undone. A draft sits in an approval queue first and goes out only once a person has signed off on it.
Typical jobs include classifying an incoming enquiry and drafting a reply, summarising a document, and drafting a quote or a contract. An internal assistant that answers from your documents, and extracting data from a form or email into the CRM, fall under the same work. Every run is written to a log, and a failed step is queued, reported and retried. Discovery takes one week, and the first live flow is built in four to eight weeks. ALTAI Digital works from Rotterdam and Istanbul.
What goes wrong today
- 01
Data moves by hand and a step gets missed
Details from a form, an email or WhatsApp are typed into the CRM by hand. On a busy day a line is skipped or saved in the wrong field. The error surfaces weeks later, and fixing it takes time. Integration moves that step into an automated flow.
- 02
No agreed source of truth
The CRM shows one detail for a client, and a spreadsheet shows another. Which system counts as correct is never written down. Updates go both ways, and the two records drift apart. The source of truth and the conflict rule are agreed before anything is built.
- 03
A model guesses when it lacks a source
A general purpose language model fills a gap with a guess when it lacks the right information. A client ends up with the wrong price or the wrong delivery date. Without a source limit, that risk sits there every time the model runs. ALTAI Digital sets the model up to work only from the sources it is given.
- 04
A message goes out without approval
A flow built without a review step sends a draft reply straight to a client. A wrong sentence or a wrong figure goes out before anyone sees it, and it cannot be recalled. Sending, paying and deleting each keep a separate human approval step for that reason.
- 05
A failed transfer goes unnoticed
A transfer breaks partway through, and no one is told. A record stays missing from the CRM for weeks. Without a queue and a retry, the failure is never picked up again. Once every transfer is logged, a gap shows up immediately.
- 06
ChatGPT integration is treated as one chat box
A search for ChatGPT integration brings to mind only a chat box on a website. The actual work is connecting the model to the CRM, the inbox and the document system behind it. A chat interface is the visible end of that connection, not the whole of it. The build sits on top of your existing systems and does not replace them.
Scope of the build
Source of truth and transfer rules
For each type of data, the systems agree which one is the source of truth. A transfer runs one way or both ways; a two way link only opens once a conflict rule is written down. Where an API exists, data moves through it; otherwise a bridge is built. The rule is written before the build starts.
Existing systems stay in place
Your CRM, ERP, accounting software, inbox, WhatsApp and Drive are not replaced. The AI step is added on top of these systems. No record is moved to a new system. Your team keeps working in the tools it already knows.
The model is kept inside its sources
The model answers only from the documents and records it is given. When a question has no match in those sources, it says so instead of guessing. Only the question and the matching passage leave the source, not the whole document. Documents are not used to train the model.
Human approval is a separate step
Sending, paying and deleting each stay with a person, since none of these steps can be undone. A draft enters an approval queue first and goes out only once approved. This step is never skipped.
Every run is logged
Each run of the model, the source it read and the output it produced are written to a log. A failed transfer is queued, reported and retried. Transfer success rate, delay and any pending error are reported monthly. A problem does not sit unnoticed.
Typical jobs are built as set flows
Classifying an incoming enquiry and drafting a reply, summarising a document, and drafting a quote or a contract are built as flows. An internal assistant that answers from your documents, and extracting data from a form or email into the CRM, follow the same pattern. Work starts with the two systems where the most data is moved by hand.
Access follows the source system
The model's access to a document is limited to whoever has access to it in the source system. A person without rights there sees nothing new through the model either.
How the build runs
- 01
Discovery and source selection
Systems, sources of truth and transfer rules are agreed together. The first job to build is chosen. Discovery takes one week.
- 02
Connection and source limits
The API or the bridge is built. The model is limited to the sources that were agreed. Approval steps are added to the flow.
- 03
First live flow
One job goes live first. The first live flow is built in four to eight weeks.
- 04
Logging and expansion
Every run is written to a log, and a failed step is queued. New jobs are added based on the results. Expansion happens one step at a time.
Products used in this build
Integrations
Your existing systems get connected; data is not carried by hand and not kept in two places.
Agent deployment
The written method for moving a job into a system: discovery, priority order, first live workflow, measurement and an audit log.
Knowledge platform
Scattered documents become one searchable source, and every answer shows its source with a page number.
Request handling
Requests from five channels collect in one list, the draft reply arrives ready, and the decision to send stays with you.
Document and report generation
Proposals, contracts, listing copy and monthly reports are produced from your template, with your data.
Terms used on this page
- API
- An API is a predefined connection rule that lets two pieces of software reach each other's data and functions directly.
- Bridge
- A bridge is the connection built between two systems when no direct API exists. It moves data through an export file, a database read or a screen level flow.
- Source of truth
- A source of truth is the written rule that decides which system's record counts as correct for a given type of data.
- Approval queue
- An approval queue is the logged step where a draft prepared by the model waits for a person's sign off before sending, paying or deleting.
- Run log
- A run log is the record that stores every run of the model, the source it read and the result it produced.
AI integration servicesFrequently asked questions
AI integration services connect a language model through an API to a business's existing CRM, inbox, WhatsApp and document systems. The model does not replace these systems; it is added as a layer above them. Which model is used follows the task and the data rules. ALTAI Digital can connect the models behind OpenAI's ChatGPT as well as Anthropic's Claude models.
ChatGPT integration is AI integration built with OpenAI's models. The work is the same; the model connects to your existing systems through an API or a bridge. ALTAI Digital chooses the models behind ChatGPT or Anthropic's Claude models based on the task and the data rules, and writes down the reasoning.
No. Your CRM, ERP, accounting software, inbox, WhatsApp and Drive stay in place. The AI step is added on top of these systems, and none of them is switched off. Your team keeps working in the tools it already uses, and no record moves to a new system.
The model is set up to work only from the sources it is given. When a question has no match in those sources, it says so instead of guessing. Only the question and the matching passage leave the source system; the full document is never sent out.
No. Documents are read only to answer that specific question and are not used to train the model. Access follows the source system; a person without rights to a document there sees nothing new through the model either.
No. Sending, paying and deleting each keep a separate human approval step, since none of these can be undone. A draft enters an approval queue first and goes out only once a person has approved it.
Discovery takes one week; systems, sources of truth and transfer rules are agreed during that time. The first live flow is built in four to eight weeks. The timeframe depends on the scope of the job that is chosen.
Classifying an incoming enquiry and drafting a reply, summarising a document, and drafting a quote or a contract are typical jobs. An internal assistant that answers from your documents, and extracting data from a form or email into the CRM, fall under the same work. Each job is built as its own flow.
The process first, the proposal after
In a short discovery call we map the process together and tell you plainly whether it is worth automating. If it is not, we say that too.
30 minutes · no preparation needed
