7UNIT / AI INTEGRATION
AI INTEGRATION
AI, integrated into the systems you already operate.
For teams that have a working technology stack and need AI inside it, not a new product beside it.
You already have software. 7Unit integrates language models, document intelligence and assistants into that estate: your CRM, ERP, product or internal tools. The existing system stays in control. The model is a component with a boundary, a fallback and a log.
INTEGRATION EVIDENCE
Integration evidence
Shipped patterns. We do not invent a client logo or a performance claim for your stack.
Practical AI assistants
LLM assistants in production, including WhatsApp inquiry agents and document pipelines, with session memory, cost attribution and a handoff when the model should stop.
Read the AI assistants caseBloom and Odoo
A learner portal integrated with Odoo as the source of truth. Educators kept their system. The new surface did not replace the back office. The same discipline applies when the new surface is a model.
Read the Bloom caseAgentOps
Agents connected to the tools a team already uses, with approvals and audit rather than an unbounded assistant.
Read about AgentOpsDocument and compliance workflows
Where a value must be exact, we use deterministic generation rather than a model. The AML document engine refuses to guess. That boundary is part of how we integrate AI: some steps must not be probabilistic.
Read the compliance engine case
THE OPERATIONAL PROBLEM
The stack is not the problem. The model has nowhere safe to live.
A demo chatbot does not become part of the operation. Integration means the model can read the right context, take only the actions you allow, and fail in a way your team can see.
A pilot with no system of record
The model answered in a sandbox. It cannot update the CRM, open a ticket, or show an operator what it did.
Unstructured work inside structured software
Documents, messages and notes still need a person to turn them into fields your ERP or product understands.
Provider risk
A single model API is now on the critical path, with no fallback, no cost attribution and no owner.
Knowledge the team cannot find
Procedures exist in drives and wikis. People ask a colleague because the software does not know them.
CAPABILITIES
What we integrate
Integration work is scoped to the estate you have. We do not insist on a new platform in order to use a model.
LLM and model APIs
Claude, OpenAI and similar providers, chosen for the task. The provider is a dependency we can observe, not a brand commitment.
AI inside an existing product
Assistant or classification features built into a SaaS or internal tool your users already open.
CRM and ERP connections
The model reads and writes through the system of record. People do not re-key the result into the real system afterwards.
Document intelligence
Classification and extraction pipelines that hand structured output to the workflow, with review on exception.
Knowledge and retrieval
Answers grounded in your material, with a path back to the source, for teams that already have the documents.
Observability and fallback
Latency, errors, token cost and a defined behaviour when the model is unavailable. Live conversations escalate. Async work can queue.
HOW 7UNIT WORKS
How an integration is engineered
We start from the system that already holds the operation, then decide where a model is allowed to participate.
- 01Inventory the estateWhich system is the source of truth, which APIs exist, and where data is not allowed to leave.
- 02Name the jobOne feature or workflow: extract this document, answer from this knowledge, draft this update. Not “add AI”.
- 03Draw the boundaryInputs, outputs, tools the model may call, and actions that require a person.
- 04IntegrateThe model sits behind your application boundary. Identity, tenancy and audit stay with the existing system.
- 05InstrumentCost, errors and handoffs are visible per tenant or per workflow before the feature is called production.
- 06Prove the fallbackWe test the path where the provider times out or returns nonsense. That path is part of the design, not a later patch.
- 07Hand overYour team gets the runbook, the provider decision, and a way to change the boundary without a rewrite.
HOW THE SYSTEM IS BUILT
The model is a component. The system remains yours.
7Unit integrates AI into an existing technology estate. If the right fix is workflow engineering without a model, we will say so and point you at business automation instead of forcing a provider into the design.
- Your system of record
- Model API
- Retrieval or extraction
- Allowed actions
- Human review
- Fallback and logs
ENGAGEMENT
Start from the system you have
An integration engagement is scoped to one job inside the current estate. We do not ask you to replatform in order to run a pilot. We also do not promise a saving before the workflow and the data boundary are known.
The current system stays primary
AI writes through the product, CRM or ERP you already trust. It does not become a second database.
A visible failure mode
When the model or provider fails, the workflow degrades on purpose: queue, retry, or handoff.
Cost and access you can explain
Calls are attributable. What the model is allowed to see and do is written down.
What we identify before we scope
- Current process, including the tools people actually use.
- A baseline: volume, delay, handoffs, or where work gets lost.
- The operational outcome you want, written before a model or a build is chosen.
- Systems involved, and which one remains the system of record.
- Risk and control boundaries: what may be automatic, what needs a person.
- Success criteria the engineering can be judged against.
USE CASES
Integration jobs we recognise
A CRM that never hears the conversation
The customer talks on WhatsApp or email. The CRM is updated later, or not at all. The model and the integration exist to close that gap, not to replace the CRM.
Documents in, fields out
Contracts, compliance packs or operational paperwork need to become structured data in a system that already runs the process.
An assistant inside the product
Users should not leave your software to ask a general chatbot. The assistant uses your tenancy, your permissions and your data.
A provider you can change
The first model is not the architecture. Calls, prompts and tools sit behind an interface you can observe and, if needed, swap.
INTEGRATIONS
Typical integration surfaces
We integrate with the APIs and products in the workflow. Common ones:
- CRM APIs
- ERP and Odoo
- WhatsApp Business API
- Internal product APIs
- Document stores
- Claude and OpenAI
- PostgreSQL
- Existing auth and tenancy
Questions buyers ask
Can you add AI to software we already use?
Yes. That is this page. We integrate models into a CRM, ERP, messaging surface or product you already operate. The first question is which system remains the source of truth, not which model is fashionable.
Will our CRM or ERP remain the system of record?
That is the default. Bloom kept Odoo as the educator source of truth while a new portal sat on top. The same rule applies to a model: it should read and write the system you already trust, not fork the data.
Do we need to replace our existing software?
No. Replacement is a product-engineering decision, described separately. AI integration assumes the estate stays, unless discovery shows a system cannot safely hold the workflow.
What happens when the model API fails?
We design that path. On assistants we have shipped, async work can queue and retry, and a live conversation hands off to a person. A feature that fails closed with no operator path is not finished.
How do you handle data sent to a model provider?
We document what crosses the boundary, minimise it, and do not treat client data as training material. Access, tenancy and audit stay in your system. If a step cannot tolerate a probabilistic answer, we keep it deterministic. Legal and regulatory sign-off stays with your counsel.
Can you integrate document intelligence?
Yes, where classification or extraction is the job: documents in, structured output into the workflow, human review when the model flags an exception. Where a generated value must be exact, we prefer a deterministic engine, as in the AML compliance work.
Do you only connect an API, or do you build the product around it?
Both, as the job requires. Connecting a model without identity, logging and a user-facing path is not an integration we will call production. Full product engineering remains available when the surface itself has to be built.
How is this different from an AI automation project?
AI automation starts from an operational workflow and may introduce new routing, agents and rules. AI integration starts from software you already run and puts a model inside it. Many engagements touch both. If you are unsure, we will say which problem we are actually solving before we scope.
CONTINUE