Document Intelligence is the authoring surface inside Kimbodo Intelligence, built for the documents a business cannot afford to get wrong: proposals, statements of work, agreements, credit memos, policies and the letters that go out over someone's signature. The agent lives inside the document rather than in a window beside it. It reads the draft, the delivery plan, the company record and the conversation that produced them — and when it wants to change your words, it asks, in redlines, one clause at a time.
Every frontier model, one editor, and a complete record of who changed what, when, and on what basis.
Almost every AI writing feature shipped in the last two years does the same thing: you paste, it rewrites, and you are left with no record of how the result was reached. That is fine for a blog post and unacceptable for anything with a counterparty. Document Intelligence was built the other way round, starting from how document control actually works.
It is not pasting text into a box and reading it back. It holds the document's structure, its delivery plan, the thread it came from and the company it is for, and it can be scoped down to the single paragraph you have highlighted.
The default is not to write. It is to propose. Changes arrive as tracked redlines with a rationale and a risk level attached, and the document does not move until a human accepts them.
Every proposal, acceptance, rejection, comment, review decision and version is an audit event with an actor and a timestamp. The agent is a participant in the record, not an invisible hand in the margins.
This is the part most tools gloss over. It is worth being specific, because the difference between a redline system you can rely on and one you cannot is entirely in these details.
The same three modes as the rest of Kimbodo Intelligence, enforced at the tool layer rather than requested politely in a prompt. In Ask and Plan the write tools are genuinely withheld from the model — it cannot touch the document even if it decides it would like to.
Most proposals carry a delivery schedule that was drawn in a slide, disagrees with the scope section, and is obsolete the day it is sent. Here the plan is structured task data attached to the document. Draft the proposal, tell the agent when the work starts and how it should be shaped, and it writes the plan — which you then edit like a project manager rather than like a designer.
Tracked change handles the words. Getting a document through an organisation needs the layer above it: who is reviewing, who has to approve, what state it is in, and what it said the last time somebody signed off.
Choose reviewers and assign the role each one holds — reviewer, approver or legal reviewer — and the document moves into review rather than relying on a chased email.
Approve, reject or request changes. The document status follows the decision through draft, in review, needs revision, approved, locked and archived.
Freeze the working draft as a numbered version with a change summary, then carry on editing. What you sent stays exactly as you sent it.
Bringing back an earlier version creates a new one from its content. Nothing in the record is overwritten to make the present look tidier.
Threaded comments anchored to the selected text, typed as questions, legal review, commercial review, approval requests or internal notes, with @mentions that resolve to project members and can carry an assignee.
A locked document stops being editable by the people who were drafting it. Executed paperwork should not stay as editable as a first draft.
Each agent run stores an immutable snapshot of the context it was given, along with the model, the token count and the cost. Every AI-assisted change can be traced back to the evidence behind it.
Live indication of who else is viewing or editing, and optimistic concurrency on save that refuses to let a stale tab quietly overwrite newer work.
Model choice is a property of the request, not of the product. The editor sits behind a provider-neutral interface, so a new frontier release is a catalog entry rather than a rebuild — and the document, the tracked changes and the audit trail are entirely unaffected by which model wrote a given draft.
Lending runs on documents. Term sheets, commitment letters, credit memos, participation agreements, servicing transfers, vendor contracts, policy manuals and the adverse action letter that has to say precisely the right thing. In that world the interesting question about an AI editor is not how good the prose is. It is whether you can prove, a year later and to a skeptical reviewer, exactly what changed, who agreed to it, and what they were looking at when they did.
Not "AI-assisted document". A specific hunk, with the original text, the replacement, the stated rationale, the risk level, the person who accepted it and the moment they did.
Write tools can be withheld entirely. A document held in Ask mode gets you drafting help with no capacity to alter the instrument, which is a very different risk conversation than a model with edit access.
Named approvers and legal reviewers, explicit approve, reject and request-changes decisions, and document states that reflect where the paper actually stands.
Frozen versions and restores that add rather than overwrite. The document as executed remains the document as executed.
Context snapshots record what the agent was shown for each run. When somebody asks why a covenant was worded that way, the evidence is attached to the document rather than lost in a chat log.
Import their marked-up Word file, work on it with full tracking, and send back a clean Word, Google Doc or PDF. Nobody outside has to adopt your tooling to close a deal.
Kimbodo Intelligence provides document control, attribution and approval workflow. It is not a compliance certification, a licensed legal opinion or a substitute for your own regulatory review — it is the record that makes that review possible.
Tracked authoring, approval workflow and a plan that is real data are not a financial-services feature. They matter anywhere a document carries obligation, and the shape of the problem barely changes from one sector to the next.
Policy wordings, endorsements, submissions and claims correspondence.
Agreements, NDAs, master services terms and playbook deviations.
Proposals, statements of work and delivery plans that have to agree with each other.
Tenders, bid responses, framework agreements and evaluation records.
Protocols, standard operating procedures, payer contracts and vendor agreements.
Architecture proposals, implementation plans, runbooks and migration schedules.
Leases, purchase agreements, contractor scopes and programme schedules.
Supplier agreements, quality documents, tenders and onboarding packs.
A document tool that cannot accept a Word file or produce a PDF is a document tool nobody outside your building can work with. Bring documents in, work on them with the full tracking and review apparatus, and send them back out clean.
Import from a Google Doc link, a Word DOCX, Markdown or plain text, converted into the editor's own structure so tracked change works on it from the first edit.
Export as a Google Doc, Word DOCX, PDF, Markdown, HTML or plain text. PDF and DOCX are export formats; PDFs are produced, not parsed back in.
Exports carry the work plan as an appendix when the document has one, so the schedule and the scope arrive in the same file.
Publish the plan separately as a Google Doc, Markdown or CSV, and the Gantt itself as a Google Slide, PNG, SVG or PowerPoint.
Every export is recorded against the document and its version, so the rendition that went to the customer is identifiable afterwards.
Any draft or reply in a project thread can become a tracked document in one action, so good work in a chat does not stay stranded there.
Document Intelligence is one surface of Kimbodo Intelligence, sharing its memory, its tools and its permission model with the projects, the pipeline and the mail. The proposal drafts from the deal, the plan generates from the proposal, the review happens where the words are, and the entire process leaves a record you would be comfortable handing to an auditor.
The fastest way to judge this is to watch it work on your own paperwork. Send us a note about the documents you live in, and we will set up a working session: we take one real proposal, agreement or memo, and redline it in front of you — tracked changes, review flow, delivery plan and all.
Tell us what you are building. We usually reply within one business day.
