Automate work smartly
Most of what can be automated needs little or no AI. We build the code that does the work, add AI only where it earns its place, and run four such systems in production.
Most of it is not an AI problem
Ask around for automation right now and you will be offered AI, because that is what sells. But the work that costs your people the most hours is usually rule-based: an order that has to become an invoice, stock that has to be right, a driver who has to see the correct route. Those rules are known, so plain code does it better. It is cheaper, it is faster, it does the same thing every time, and you can test whether it is correct.
AI earns its place at the edges of that, where the input is language or judgement rather than a rule. A creditor email that has to be read and answered. A bank statement that has to be interpreted. Ten thousand documents nobody has time to go through. Put AI there and it is worth the money. Put it in the middle of a process that had clear rules and you have swapped something predictable for something that is usually right.
- Code, when the rules are known: orders, stock, invoicing, planning, routes, permissions.
- AI, when the input is language or judgement: correspondence, documents, classification, summarising, pre-qualifying.
- Neither, when a well-configured existing package already does it. We will tell you that too.
- AI systems of ours running in production
- 4AI systems of ours running in production
- Less processing time per application at BalansBuddy
- 98%Less processing time per application at BalansBuddy
- Lines of domain knowledge inside Alfred AI
- 71,000+Lines of domain knowledge inside Alfred AI
What an AI assistant actually does
Most things sold as an AI assistant are a chat window sitting on top of a list of frequently asked questions. It demonstrates well and changes very little, because answering was never the bottleneck. The work was.
An assistant that helps does three things in sequence: it reads what your company already knows, it answers on that material rather than on the open internet, and then it performs an action in a system. Reply to a creditor. Update a case file. Flag a payment that is about to go wrong. The value is in that third step, and that is also the step most demos skip.
How one gets built
Every business needs something different, so the order matters more than the recipe. This is the order we work in.
- Start at the work
- Which knowledge is stuck in people’s heads here, and which task comes back every day and follows rules someone can explain? Those two questions decide what gets built, and whether it needs AI at all.
- Open up your own data
- Documents, procedures, mail, the things in people’s heads. Without this an assistant answers in generalities, and generalities are what makes people stop trusting it.
- Connect to the systems
- The assistant has to be able to act, so it needs access to the place where the work lives: your mail, your administration, your case files.
- Test it against being wrong
- Not whether it sounds good, but what happens when it is unsure. What it refuses to do, what it escalates, and to whom.
- Go live and keep measuring
- A system like this is never finished on delivery day. What matters is whether the numbers still hold three months later.
Eight weeks from first conversation to something in production. You get one fixed price after that first call, and it does not change afterwards.
Which AI is best for business
The honest answer is that the model is the least interesting choice you will make. It changes every few months, and whichever one is ahead today will not be in a year. What decides whether the thing is any good is how well your own material is opened up, how reliably the system finds the right piece of it, and what limits sit around what it is allowed to do.
So we work with whichever models perform best at the time and replace them when something better arrives. Because you own the code, the data and the documentation, that swap is ours to make and yours to keep. Nothing about it locks you to us.
Why that judgement is worth anything
Deciding what should be code, what should be AI and what should be left alone is the whole job. That judgement does not come from a course. Behind 2B Global sits twenty years of work in industries that rarely overlap: luxury hospitality, streaming and broadcast, festivals and events, oil and gas, and music.
What that leaves behind is a habit of asking a different first question. Not "where could we fit a model in", but "which knowledge is stuck in people’s heads here and could be written down". Each of those worlds hands you one question you keep asking everywhere else:
- Hospitality
- Who is actually being served here, and how well? A process that is efficient on paper but leaves the customer waiting has not been automated, only moved.
- Streaming and platforms
- Do the costs rise with every extra customer, or does the same work serve a thousand? That difference decides what is worth building.
- Events and festivals
- Where is physical presence the limit? That is usually the place where one person is the bottleneck for everyone else.
- Oil and gas
- Who are all the parties involved and what does each of them want? Automation fails on the stakeholder nobody mapped.
- Music
- What is the equivalent of reading the room? Some judgements should stay human, and knowing which ones is part of the design.
How to pick who builds it
There are a lot of parties offering this now. Three questions separate them faster than any sales conversation.
- Does anything of theirs actually run in production, with people depending on it, or is the site all demos?
- Whose infrastructure will your data sit on, and can you move it out?
- Do you own the code after delivery, or are you renting something you cannot take with you?
For Meet-Encrypt we built Abby, an AI note-taker that runs entirely inside the client’s own encrypted infrastructure and never leaves it. That is the shape we build in wherever the situation allows it.
The four that run in production
Not pilots and not prototypes. These are in use, with the numbers that came out of them.
- BalansBuddy and Ember AI
- Debt assistance. Ember handles creditor correspondence for people in debt restructuring: processing time per application went from fifty hours to one, and the rejection rate from 55% to 5%.
- Amber Debito and Amber AI
- Accounts receivable for smaller businesses. Automated payment reminders, escalation that knows when to stop being polite, and a link into the accounting system.
- Trade Fair Control Room
- Event intelligence. Attendees are pre-qualified through automated interviews, routed to the right person and written straight into the CRM.
- Alfred AI
- An enterprise knowledge system built on retrieval: more than 71,000 lines of domain knowledge, answered conversationally, used to support sales.
Questions people ask
- Does it have to be AI?
- Usually only in part. Work with known rules — orders, stock, invoicing, planning, routes — is better served by plain code: cheaper, faster, the same every time, and testable for correctness. AI earns its place where the input is language or judgement, such as reading correspondence, interpreting documents or pre-qualifying applications. And sometimes the answer is that a well-configured existing package already does it, in which case we will say so.
- What does an AI assistant actually do?
- That depends on the work you give it. An assistant that only talks is a search box with better manners. An assistant that takes work off your hands reads your documents, answers on your own data and then performs an action in a system: reply to an email, update a case file, raise a flag. That last part is where the time is saved.
- How long does it take to have an AI assistant built?
- Eight weeks from first conversation to something running in production. Week one is the conversation and the plan, then design, building with weekly reviews, testing with the people who will use it, a security check, and go-live with training. You get a fixed price after that first call, locked before we start.
- Which AI is best for business?
- The model is the least interesting choice, because it changes every few months. What decides quality is how well your own data is opened up, how the system finds what it needs, and what guardrails sit around it. We work with whichever models perform best at the time and swap when something better arrives. Because you own the code, you are not locked in.
- Does my data stay mine?
- Yes, and that is a design decision rather than a promise made afterwards. For Meet-Encrypt we built Abby, an AI note-taker that runs entirely inside the client’s own encrypted infrastructure and never leaves it. We build that way wherever we can. All code, data and documentation are one hundred percent yours.
- What does it cost to have an AI assistant built?
- It depends on what it has to do, because every system is custom. What decides where you land: how many sources have to be opened up, how many systems it has to talk to, and how strict the requirements around privacy and control are. You get one fixed amount after the first conversation, and that amount does not change afterwards.
- Where can I have an AI agent built?
- There are many options. Three questions separate them quickly. Does anything of theirs actually run in production, or are there only demos on the site? Whose infrastructure will your data sit on? And do you own the code after delivery, or are you renting it? We run four AI systems in production, build on your own environment where we can, and you own everything.
Wondering what this would look like for your work?
Thirty minutes, free, and you will know whether it is worth building. No obligation and no sales trajectory.
Book a free 30-minute call