What Forge does with a document.
Three ways to work with it, one set of rules over all of them.
Three ways to work.
Which one you use depends on how often the work comes back. The governance does not change between them.
Open chat
Ask anything, with live web search. Attach a file for a single conversation and it is deleted along with it. Personal details are masked before the message reaches the model, and restored in the answer you read.
Projects
Put a set of documents in one place for your team. Ask questions answered from that material with the source named, or generate new work from it: a brochure in your brand voice, a summary a client can read, a first draft of a response.
Workflows
When the same kind of document keeps arriving, define the path once and every one after that follows it. Vendor onboarding and bank statement reconciliation ship ready to use.
Three things do the work, not one.
A workflow here is built out of three parts, because most real work needs all three and a chat box only has one of them.
AI, for understanding
Reading a document written for a person, classifying it, summarising it, reasoning about what it says.
Software, for certainty
Calculation, validation, data processing and rules. The things that must give the same answer twice.
People, for judgement
The decisions somebody has to be accountable for, and the exceptions no rule anticipated.
Forge brings the three together in one governed workflow and records which of them did what. On a bank statement the rules sort every row they recognise, the model is asked only about the remainder, and each record says which of the two decided it.
Why it is built this way rather than as an agent that works out its own steps is the argument on the thesis page.
What a piece of work goes through.
A contract arrives. Rather than one person reading it and typing the results somewhere else, it follows a path your organisation defined once.
- ArrivesA PDF or a spreadsheet, uploaded or forwarded.
- AIThe fields you asked for are read out of it.
- SoftwareFormats, dates and totals are checked against your rules.
- AIWhat the rules could not decide is reasoned about, and marked as such.
- A personReviews what needs reviewing and moves it on.
- SoftwareThe output is produced in the format you need.
- RecordedEvery stage carries who moved it and when.
Then talk to the records in plain language.
Ask what you would ask a colleague. Which vendors are still missing bank details. What came in from this customer last quarter. No filters, no formulas, no report to request.
Workflows you set up from a template.
Vendor onboarding and bank statement reconciliation ship ready to use. Start from one, change the fields to match how you work, and the process is repeatable from then on instead of being re-invented by whoever picks it up.
SoonAnd when the document type is particular to your business, describe the fields you want out of it and Forge reads them. A new process does not wait on us to build it.
The model will change. Your workflow should not.
Forge sits between your people and the model. Every request passes through it: sensitive details are removed, the call is made, the activity is recorded, the answer comes back with the real values restored. The governance lives in that layer, not in the model.
Which means the model underneath is a choice rather than a commitment. Point Forge at the provider your organisation has already approved, or at your own account with your own key, and change it when something better arrives. The rules you wrote, the approvals you set and the record you keep do not move, because none of them were ever in the model.
Delivering work to clients.
One more thing the same seal makes possible, if you do client work. Register your method as a program and your client runs it on their own data and gets the analysis back, without ever seeing the source. You sell the method instead of licensing it away.
Getting started, in three steps.
- 1
Set up the workspace and the rules, before anyone uses it.
Invite your people, decide who can reach which projects, add your organisation’s own AI provider key, and turn on the privacy policies you want enforced, including redaction that no member can switch off.
- 2
Bring the documents in.
Your people ask questions of them and generate work from them, in open chat and in projects. Sensitive details never reach the model and everything is recorded. This is the part that protects the data.
- 3
Turn the work that repeats into a workflow.
Start from a template, and at each step say what should do the work: an AI model, one of your own programs, a rule, or a person. This is the part that controls the process.
Governance first, then the work, then the workflows. Most software is adopted the other way round.
Questions we are asked.
- How is this different from Microsoft Copilot?
- Copilot is an assistant. Forge is a process.
- Copilot sits inside Word, Excel, Outlook and Teams, answers from what is already in your Microsoft estate, and works for whoever holds the licence. It is good at that: draft this, summarise that, what did we agree on Tuesday. The output is a draft, and the session ends.
- Forge is for the work that comes back every week. Vendor onboarding, statement reconciliation, contract checks. You define how it should be handled once, and every document after that takes the same path, and what is left behind is a record rather than a conversation.
- Most organisations should have both. Ask Copilot to help you write something. Ask Forge to handle two hundred invoices the same way twice.
- Does Forge sit in front of the AI?
- Two questions get asked in one there, and the answers are different.
- In front of the model, yes. That is what Forge is. Nothing goes straight from your person to a provider. The message arrives here first, sensitive details are swapped for placeholders, the call goes out on your key or ours, the answer comes back, the real values are put back into what your person reads, and the whole exchange is recorded.
- In front of another company's product, no. When one of your people types into another vendor's app, that product owns the path from their keyboard to its own model, and there is no place in it for us to stand. Intercepting that would mean inspecting traffic on your employees' devices, which is a different kind of product and not one we are.
- How is this different from an AI agent that does tasks for me?
- An agent works out its own steps from the goal you describe. That is the right design for work you cannot specify in advance, and the wrong one for work that has to come out the same way every time, because the steps can differ on the next run and nothing records which were taken.
- Forge asks you to define the steps once. After that the model is one step among several rather than the thing choosing them, and each record carries which step decided what. Use an agent to explore. Use a process for the work you would have to explain to an auditor.
- How does the pricing work?
- Two parts. The model cost, which you can cover with your own provider key so it is between you and the provider. And a platform fee for the governance, the privacy and the workflows, charged to the organisation and scaled to how large it is and how much it uses. There is no per-person licence, so you never have to decide who gets access.
- Why does Forge not use AI for everything?
- Because not every task needs a model. AI is useful for understanding, classification and reasoning. Software is better at calculation, validation and rules. People are essential where judgement and accountability matter. Using a model for arithmetic is how organisations end up with confident wrong numbers.
Bring the work you cannot put anywhere else.
Thirty minutes. Bring the work you cannot currently put into an AI tool, and we will tell you whether Forge helps. If it does not, we will say so.
A full list of what is implemented, marked built, coming or not offered, is on the technical reference. The security overview covers the same ground for a procurement review.
