AI automation and data extraction
Two Clouds builds AI into the systems you already run. We take unstructured input - inbound email, phone calls, support tickets, forms - extract the structured fields your business actually needs, and push them into your ERP, CRM or job management system through its API. Integration is the job; AI is one of the tools.
Most AI projects fail at the last mile.
The model is rarely the hard part. Getting its output into the system your team works in every day - correctly typed, validated, matched to the right customer and job, and auditable when it gets something wrong - is where these projects stall.
That last mile is our ordinary work. We have spent fifteen years connecting business systems by API, and an extraction step in front of that is a new input, not a new discipline. If the output has nowhere useful to land, the model is a demo.
This is a companion to our systems integration work rather than a separate discipline.
What goes in, what comes out, and where it lands
An extraction pipeline has three parts: the unstructured source, the fields pulled out of it, and the system of record it is written to. The third part is the one that decides whether the project is useful.
| What goes in | What is extracted | Where it lands |
|---|---|---|
| Inbound email and attachments | Job details, customer and site references, dates, quantities, priority | Job management system, by API |
| Phone calls and voicemail | Transcript, call summary, actions and follow-ups | CRM record against the right contact |
| Support tickets and issue reports | Category, severity, affected asset, duplicate detection | Ticketing or job queue |
| Forms and scanned documents | Typed fields, validated against your existing reference data | ERP or line-of-business database |
Every one of those destinations is a system we would already be integrating with. The extraction step sits in front of an API call, and the API call is the part that has to be right.
Six practical applications.
Inbound Job Capture
Email, calls and issue reports read automatically, with job data extracted and created in your job management system rather than rekeyed by an administrator.
Call Summarisation
Calls transcribed and summarised into a short, structured record - what was agreed, what happens next, and who owns it - filed against the right job or customer.
Document & Form Extraction
Typed fields pulled from PDFs, scans and forms, validated against reference data you already hold so that a bad read is caught before it reaches your database.
Internal Chat Assistants
A chat interface over your own documentation and data, answering from your content rather than from the open internet, with the sources it used shown alongside.
Reporting & Summaries
Recurring operational summaries generated from your own systems - what changed, what is overdue, what needs a decision - delivered where your team already looks.
API Delivery & Fallbacks
The unglamorous half: retries, validation, duplicate handling, and a human review queue for anything the model is not confident about.
What we are building right now.
A call-to-process pipeline. For one client we are turning recorded calls into a structured form submission, a written process summary and a reporting layer, with a chat assistant over the resulting records. The point is not the transcript - it is that the output is a record their team can search, report on and act from.
Inbound job extraction. For another we are analysing inbound mail, calls and reported issues, extracting the job data buried in them, and pushing it into their job management system through its API. Work that arrives as prose becomes a job record without anyone retyping it.
Both are in build rather than live, and we would rather say so than imply a track record we have not earned yet. If you want detail on approach, architecture or where the pitfalls are, we are happy to talk through either in full.
- AI wired into your existing systems, not a separate tool to log into
- Built by the engineers who already do your integration work
- Human review queues wherever confidence is low
- Your data stays in systems you control
- Costed as ordinary development, not as an AI premium
- Honest about what the technology cannot reliably do yet
How an AI engagement runs - from a narrow proof to something you depend on.
Frequently asked
What kind of AI work does Two Clouds actually do?
We build AI into existing business systems rather than building models. In practice that means extracting structured data from unstructured input - email, calls, tickets, documents - validating it, and writing it into an ERP, CRM or job management system through its API. The integration is the substantial part of the work.
Can AI push data into our existing ERP or job management system?
Yes, and that is the part we would spend most of the project on. Extraction produces fields; those fields then have to be typed correctly, validated against your reference data, matched to the right customer or job, deduplicated, and written through your system API with retries and error handling. We are currently building exactly this for a client, pushing extracted job data into their job management system.
Do you build chatbots?
We build chat interfaces over a business's own documents and data, where answers come from your content and the sources are shown. We are less enthusiastic about public-facing chatbots that answer general questions, because the failure modes are harder to contain and the business value is usually lower than automating a repetitive internal task.
What happens when the AI gets something wrong?
It will, so the pipeline is designed around it. Low-confidence extractions go to a human review queue rather than straight into your system, validation rules catch values that cannot be right, and everything is logged so a bad record can be traced back to the input that produced it. A pipeline with no review path is not finished.
Does our data get sent to an AI provider?
That depends on the approach chosen, and it is a decision we make with you at the start rather than a default. Options range from models run in your own infrastructure to commercial APIs with data-processing agreements in place. If your data cannot leave your environment, say so early and it shapes the architecture.
Have you delivered AI projects already?
We have two in build and none live yet, and we would rather be straight about that. One turns recorded calls into structured records, process summaries and reporting with a chat layer over the results. The other extracts job data from inbound mail, calls and issues and pushes it into a job management system by API. What we do bring is fifteen years of getting data into business systems reliably, which is where these projects usually fail.
Is AI automation worth it for a smaller business?
It depends entirely on volume and rekeying. If someone spends hours each week reading messages and retyping their contents into another system, extraction usually pays for itself. If the work is varied, low-volume or needs judgement at every step, it will not, and we will tell you that before quoting.
Let's build something together.
Get in touch to discuss your project and find out how we can help.
Contact us