Governed AI for regulated finance
Put your firm on AI without ever running it ungoverned.
We build the automations your shop actually runs on, and we own the boundary they run inside. Every model call is routed by your policy, recorded on a chain that breaks if anyone edits it, and stopped for your signature when the stakes are high. If you have no AI yet, you adopt it already governed and your record starts clean. If you already run AI and cannot prove what it did, we put the boundary around what you have.
The boundary
Three mechanisms between a model and your book of business.
Not a dashboard bolted onto an API key. A boundary the calls actually cross, with three jobs it refuses to skip. Watch how each one works, then see how yours gets built.
Route
A policy written for your niche classifies every request before any model runs: what counts as routine, what touches regulated data, what is high-stakes. The classification picks the model tier and the controls, and it is itself recorded with a written rationale.
router_decision · rationale · model_tierRecord
Every decision lands in a hash-chained audit log as it happens. Each row seals the one before it, so an edit anywhere breaks the chain at the exact row. What the model saw, what it answered, and who approved it: one record, tamper-evident.
audit_log · sha256 chain · verify()Restrain
Actions your signed spec marks consequential stop at a consent gate and wait for a wallet signature from the person you named. No signature, no action. And when any check cannot complete, the system fails closed instead of guessing.
consent_gate · EIP-191 · fail-closedThe engagement
From first sit-down to an operating boundary.
The engagement is not a sales call. It runs inside CONSOHL itself, so the first thing you see is the product governing its own delivery. By the fourth step you are operating, and the shape of your niche decides what you operate.
A governed working session
We interview you through your own operation, live on the console, starting from what your niche usually runs rather than a blank page. The workflow classifies what your firm does into routine work and work that deserves a gate.
A spec you sign, not a deck you nod at
The session ends in a specification of your boundary: gated actions, data rules, model policy. You sign it with your own wallet, and that signature is the first entry on your audit chain.
Your own isolated stack
The signed spec generates your configuration, and your stack comes up on your own machine: your database, your vector store, your keys, your domain. No other client shares the box, because isolation you can point at beats isolation you are promised.
Operated, not handed off
Your team works through the console: runs light up node by node, approvals wait at the gate, the chain grows underneath. We operate the boundary under retainer, watch for drift, and keep the evidence ready before anyone asks.
What your staff actually get
One window onto your own operation.
Not a chatbot bolted to the side of your business. Your processes become workflows your people run by name, and the boundary holds around every one of them without anybody having to think about it.
They pick the job
Your workflows appear by name in one interface. Someone chooses the letter or the review they need, types the facts, and reads the reply. It works like every chat they have used, and it needs no training in anything underneath.
their own operation, by nameThe work runs governed
Underneath, a real workflow executes: it reads your own records, checks what it cites, screens the draft against the language that draws complaints, rewrites once if it has to, and refuses if it still cannot pass.
routed · screened · recordedIt waits for you
Anything consequential stops and asks for your signature before it exists. Close the laptop, go home, sign in the morning: the work picks up exactly where it paused, and the whole sequence is on the record.
paused · signed · sealedWhere it runs
Your machine, your rules, your choice of how far.
Every firm gets an isolated stack. What changes between these is where the reasoning happens, and moving between them changes nothing your staff sees.
Governed cloud
Frontier models reached through one recorded gateway, on your own server with your own keys. The fastest path to working automation, and every call is logged verbatim.
your server · governed egressHybrid
Sensitive work runs on hardware you rent and control, routine work keeps the reach of a frontier model. One policy decides which is which, and it is written for your niche.
split by policyEntirely in your building
An open-weight model on your own rack, specified and built as part of the engagement. No path out, no standing connection to us, nothing leaving the premises. Your staff operate it from the runbooks we deliver.
air-gapped · nothing leavesOne industry, niche by niche
Built for finance. Specialized per niche.
One governed core, then a template per financial niche: the router policy, the workflow library, and the evidence map that fit how that business is actually examined. Your firm gets the template specialized to you, not a generic tool with your logo on it.
Credit and collections
Dispute and consumer-communication work where one careless sentence draws a sanction. Consumer-facing output gates before it leaves, and every letter traces to the record that produced it.
Lending
Origination support where adverse outcomes must be explainable. Decisions carry written rationales, and the gate holds anything consequential for a named human.
Trading and fund administration
Books-and-records retention maps naturally onto a hash chain: trade rationale logs, exception reports, and best-execution memos as evidence workflows.
Insurance
Claims and underwriting support with the same spine: policy routing, recorded reasoning, and signature gates on what touches the customer.
Evidence, not adjectives
What the record looks like.
We never sell a magic word. We sell a record that holds up: every row sealed by the next, every gate crossing signed, every model choice explained in writing at the moment it was made.
| 09:41:07 | router_decision | dispute_letter → gated | rationale: consumer-facing output | chained ✓ |
| 09:41:19 | consent_requested | awaiting owner signature | EIP-191 · named signer | chained ✓ |
| 09:44:02 | consent_approved | 0x3f…c266 signed | verdict: approve | chained ✓ |
| 09:44:03 | gateway_call | model per policy tier | prompt + response recorded | chained ✓ |
| 09:44:09 | chain_verify | full pass | 0 breaks · row 8,412 | intact ✓ |
Questions firms actually ask
Straight answers.
Does CONSOHL satisfy my regulator for me?
No, and any vendor who says yes is selling you an adjective. CONSOHL produces the evidence: which model was called, under which policy, what it saw, what it answered, and who signed off. Your counsel maps that record to your obligations. We keep the record complete and tamper-evident; the claims stay yours to make.
Where does my data live?
On your own machine. Every client runs an isolated stack: your own server, your own database, your own vector store, your own encryption boundary. No shared infrastructure, no pooled tenancy, no other client on the box. Embeddings are computed locally and never leave it.
What stops the AI from doing something it shouldn't?
Three mechanisms, in order. A router classifies every request against your niche policy before any model is called. Everything that runs is written to a hash-chained audit log as it happens. And actions your signed spec marks high-stakes stop and wait for a wallet signature from the person you named. Fail-closed is the default: when a check can't complete, the action doesn't run.
What happens in the first engagement?
A structured working session, run inside CONSOHL itself. We walk the processes firms in your niche typically run, ask which ones you do and where yours differ, and sort your operation into routine work and work that deserves a gate. No documents required to start, and nothing you do share ever leaves your own machine. It ends with a specification you sign with your own wallet. That signed spec is the contract your stack is built from, and you watch the whole thing happen live on the console.
Which AI models does it run?
Frontier models over one governed gateway, so every call crosses a single recorded boundary. The model tier is picked per request by your policy: routine work runs lean, consequential work runs on stronger models, and the choice itself is logged with a written rationale.
Begin with the working session
Put a boundary around it before someone asks why there wasn't one.
AI is coming into your business either way, and so are the questions about it. One structured engagement: your processes in, a signed specification out, your own governed stack built from it, running on your own machine. Picture twelve months from now: every consequential decision your AI made sits in one intact chain, each one signed, each one explainable, and you can put your hands on any of them in seconds. Capacity stays limited to a few engagements per quarter, because every boundary is operated, not shipped.