The short answer
A shared mailbox is well suited to preparation and badly suited to being left alone. The workflow sorts what arrives, extracts the fixed data, flags what is missing and prepares a draft. An employee checks and sends. Building this one process costs 1,500 euros excluding VAT and takes ten working days.
The concrete problem
The office team or administration opens the same kinds of email: fault report, maintenance, quote, invoice, spam. Every email means searching the customer database, checking a rate or balance, and only then replying. The retyping is slow. A commitment made in haste is expensive.
Who this is for
Decision-maker
You want response times down without extra headcount and without AI committing to a date, price or credit note on its own.
The person handling the inbox
You want to see per email: category, what is missing, what the fixed rule says, and a draft you are still allowed to reject.
Now versus after
| Now | After delivery |
|---|
| Open every email, look up the customer, only then reply | Category, missing fields and a draft are ready |
| A date, price or credit note sometimes goes out the door already | The draft does not promise that. You send or reject |
| Spam and real questions are mixed together | Irrelevant email goes to archive, not to a reply |
What you see below
Every enquiry is turned into a clear file, a useful draft and a next step. AI prepares the work so your team can respond faster and give customers better advice.
How the workflow works
Email arrives
A mailbox or shared inbox receives the question, quote request, complaint or fault report.
The language model reads and proposes
The model reads, classifies by category and urgency, extracts data and drafts a text. It only does the language work: reading, classifying, summarising, drafting.
Fixed rules are checked
Business rules in code check the case against customer status, rate agreements, outstanding balance and missing data. That is not asked of the model but looked up, so the same documented rule applies to every case.
Employee approves, adjusts or rejects
Nothing goes out, is booked or is committed to without approval. Exceptions and deviating cases always go to a human.
Every step is logged
Timestamp, category, customer, confidence and the applied regime are recorded, so what happened can be traced afterwards.
What you receive
- Baseline measurement of volume, minutes, lead time and errors.
- One workflow on one mailbox, at most two integrations.
- Categories, escalation path and logging.
- Test set of real emails and a final measurement.
- Handover to whoever handles the inbox afterwards.
Systems
Microsoft 365 or Google mailbox, plus a customer database. CRM or ERP only if that integration is in scope. Check Copilot, Zapier or the AI in Exact or AFAS first; if they cover the whole route, I build nothing. That is also on processes.
The draft reply
The language model prepares a reply. Fixed rules in code keep price, date and credit note out until you allow them. You adjust, approve or reject. Nothing goes out the door by itself. The full explanation of that step is on draft replies.
Data and human control
Only the fields this process needs go into the workflow. Master data is looked up, not asked of the model. Every external action waits for an employee. How that boundary is designed is on security.
One process working in ten working days
From the agreed start and complete information: baseline and design, followed by build, test on real emails, final measurement and handover within ten working days. The full route is on the how it works page.
What this means for the people doing the work now
The employee remains owner of the outcome and handles the cases the workflow does not understand. The goal is less time on the same retyping and sorting, not automatically fewer people. The full explanation is on the how it works page.
When this does not fit
This demonstration does not read real attachments: PDF or image processing requires separate work in a real implementation. What that process pattern looks like is on document checks. It does not connect to an existing CRM or ERP; that integration too must be designed and tested for the specific system. And it has not been measured: a baseline and a quality measurement on a test set belong to the paid implementation, not to the demonstration. See what you get for free.
Price
€1,500 excluding VAT for this one process, ten working days. Calculate with your own inbox volume in the calculator.
Why you get this demonstration
Because you want to see for yourself how the separation between language work, fixed rules and human approval works in practice, before you sign anything. At one of the largest banks in the Netherlands a production workflow runs with that same separation. That is not an email process or a Tamboer client case; the demonstration here shows the control principle, not the bank file. The production story is on the about page.
Frequently asked questions
Does the workflow send email itself?
No. The workflow drafts a reply. An employee approves, adjusts or rejects before anything goes out, is booked or is committed to.
What if the workflow misjudges a case?
Uncertain or deviating cases go to a human via an agreed escalation path. Every step is logged, so it is visible afterwards what happened and which regime was applied.
Does this work with my own mailbox and customer database?
AI turns separate emails into a useful customer file, a reply and a next step. I configure the workflow for your mailbox, customer database and business rules.
