Prove which AI touched which matter.
Matterwall reads your clients’ outside-counsel guidelines and turns them into a policy that governs the AI tools your firm already buys. It blocks what a client’s clause forbids, runs restricted work in its own interface, and logs every decision and every attorney review against the clause behind it. When a client asks what AI touched their work, you send a report instead of an assurance.
Free for the first 20 Am Law 51-200 firms. Engaged through your GC or outside ethics counsel, so the findings are work product.
Your clients are starting to ask what your AI touched
Fulkerson Advisors read 1,054 sets of outside-counsel guidelines written since ChatGPT shipped. Twenty mention AI. One bans it. Most of the sample is public bodies, so the corporate clients who pay your rates are underrepresented.
The rate is the part to watch. AI turned up in 0.9% of the guidelines dated 2024, 2.2% of those dated 2025, and 3.8% of those dated 2026. The 208 written before 2024 mention it zero times. That is roughly a doubling every year, from a base of nothing four years ago.
What those twenty ask for is the useful part. Twelve want to be told when AI was used, nine want their data kept out of public tools, eight want a lawyer to check the output, and three want written permission first. The same study quotes the FDIC telling its firms never to put its work product into a tool that runs in an open environment, and Zscaler asking that the AI be noted in the time entry. Every one of those is a question about a specific matter on a specific day.
Which tool saw this file? Did a lawyer review the draft? Was this client’s data ever sent to a public model? Most firms can answer with a policy everyone signed. Almost none can answer with a record.
Matterwall makes the record. It routes each prompt by the guidelines that govern the matter, and writes down what it did and which clause made it do that.
How the policy actually reaches the prompt
The question every CIO asks second, including the tools Matterwall cannot touch.
Where does Matterwall sit?
In three places, because no single one covers a firm. Your egress proxy or managed browser allows or blocks each AI vendor by domain, per client policy. The Matterwall interface is the only surface Matterwall itself routes, so restricted work happens there. Everything else comes from the vendor's own audit log, pulled on a schedule and matched to matters. Single sign-on tells us who the lawyer is. It does not carry the traffic, and any vendor who implies otherwise is guessing.
How does it know which matter this is?
From the document, not from the lawyer. The primary signal is iManage or NetDocuments context for the files in the prompt. A file dragged off a local disk carries no such metadata, so we match it against the DMS by near-duplicate comparison rather than an exact hash. A lawyer who cut two paragraphs before pasting would defeat exact matching. Heavy edits and short excerpts still fall through. The secondary signal is the matter open in time entry. A lawyer can correct a matter after the fact, and the correction is logged with their name on it.
What about pasted text with no file at all?
It has no matter, which means it has no client, so the policy cannot come from the prompt. It comes from the lawyer. Matterwall applies the strictest policy among the clients that lawyer is currently staffed on, read from your matter system. A litigator staffed on a no-public-tools client gets the restricted path for everything they paste until they attribute the prompt to a matter. That over-restricts, deliberately, and the one-click attribution is how a lawyer clears it.
What about Copilot for M365?
You cannot route Copilot, and we will not pretend otherwise. Microsoft tells you to exempt Copilot traffic from TLS inspection, so a proxy cannot read it either. Enforcement is a tenant policy in your Microsoft estate, scoping Copilot away from restricted matter sites with Restricted Access Control and sensitivity labels. The record comes from the Purview audit log, which writes one entry per Copilot interaction with the files it touched. We help you write the policy and we read the log.
Does this replace Harvey?
On unrestricted matters, no. Harvey calls its own backends, so Matterwall cannot reroute a Harvey prompt. What it can do is pull Harvey's audit-log API, match each entry to a matter, and block Harvey at the proxy for clients whose guidelines forbid it. Restricted work happens in the Matterwall interface. Your lawyers use two surfaces depending on the client, and that is a real cost of buying this.
Where does the gateway run, and what does the log hold?
In your cloud account or your colo, not ours. The log stores metadata by default, which is the matter, the clause, the route, the model, the reviewer and the timestamp. Prompt text capture is off unless you turn it on, and when you do, the text sits in your store, readable by your administrators and not by us. Retention is yours to set, and the store honors a legal hold.
What we can do to each tool
Two of these will come up on your first call. The ChatGPT desktop app pins its certificates, which means your proxy cannot inspect it and the answer there is to block the app and keep people in the browser. Microsoft asks you to exempt Copilot from TLS inspection outright. The Westlaw and Lexis rows say what they say because we have not been able to confirm an enterprise audit export from public documentation, and we would rather tell you that than fill the cell in.
A lawyer who opens ChatGPT on a personal phone is outside all of this, and no routing layer catches that. Matterwall covers the tools the firm sanctions, and the first job of the review is to tell you what those are.
Why not just…
…write an AI policy and have everyone sign it?
A policy is what you tell the client. A log is what you show them. Of the twenty guidelines that mention AI, twelve ask for disclosure of what happened on their matters, and an attestation is not disclosure. You will need the record eventually, and reconstructing it after the question arrives is the expensive way.
…use the CASB we already have?
Netskope, Zscaler and Purview already inspect prompts, block uploads and log who used which AI tool, and they do it well. What none of them holds is the matter. They classify data, so they can tell you someone pasted something sensitive. They cannot tell you it belonged to the client whose §14.2 forbids public tools, because no CASB reads outside-counsel guidelines. Matterwall is the clause layer on top. If you already run one, it does the inspecting and we do the mapping.
…use Harvey / ChatGPT Enterprise?
Keep them. They serve unrestricted matters as they do now, and Matterwall reads their audit logs so that work still lands in the record. On a client whose guidelines forbid them, the proxy blocks them and the work moves to the Matterwall interface. You are buying the control and the record, not a replacement for either tool.
…wait for the labs to offer on-prem?
When they do, their model becomes one more route to choose between, which is the problem Matterwall already solves. More routes make the policy layer worth more. Nothing you buy now is stranded by that.
One policy, applied to every tool you sanction
Matter-aware policy engine
The engine checks each prompt's client, matter and OCG terms, then decides what that client's guidelines allow. Unrestricted work goes to whichever frontier tool your lawyers already use. Restricted work moves into the Matterwall interface. If you have not bought the private stack, restricted work is refused there, with the clause quoted and the lawyer told what they can do instead, which is usually to work without the AI or to ask the client for written permission. The refusal names the client and the clause, so the lawyer can take it to the partner or to the client instead of working around it. Where a clause asks for written permission first, the prompt is held rather than refused, and it runs once the permission is on file.
The record, and the report you send
Every prompt writes one row against the OCG clause behind it, with the route, the model and the timestamp. Review attaches to the work product, not the prompt. A research memo takes forty prompts and one lawyer reading the draft, so it produces one attestation against the draft. That attestation records what the lawyer read, what they changed and when, because a bare checkbox is not evidence. Work done in Harvey or ChatGPT lands in a review queue built from those vendors' audit logs, and the lawyer signs off there. A row still at Pending is work nobody has signed off, and that is the number to watch. The log exports as a per-client report your GC can attach to an audit response or an RFP answer. Writing an AI flag back into Intapp, Aderant or Elite time entries is on the roadmap and not built yet.
Matter walls at retrieval
Retrieval runs against iManage and NetDocuments and enforces matter walls as it searches, so the model can only see documents the lawyer asking could already open.
Private stack
The latest open-weight frontier model on hardware dedicated to your firm, in a colo cage, a dedicated cloud tenancy, or on-prem. It is for the clients whose guidelines say the data cannot leave your building. Open-weight models trail the frontier by a wide margin on legal reasoning, so this is a compliance purchase and not an upgrade. Clauses like that appear in roughly 1% of guidelines, which for most firms is a handful of client names or none, and that is why it is priced separately.
Security
- Policy engine runs in your cloud account or colo
- Prompt text capture off by default
- No training on your data, ever
- Matter walls at retrieval
- Full audit log per prompt, per matter, per OCG clause
- Retention and legal hold under your control
What it costs
We do not charge per lawyer. A platform fee covers every lawyer and every tool behind the policy, and after that you pay for the clients that actually have AI clauses. If the review finds six, you pay for six. A firm should not pay a percentage of its AI budget to govern a fraction of its work.
$114K, against $168K to $240K if we charged $35 to $50 a lawyer. Implementation is quoted separately and is not included in that figure.
$60K at this size. Every lawyer, every tool behind the policy, the log and the client reporting.
$9K. Clause mapping, policy maintenance and that client's report. Quoted only for the clients the review actually finds.
Hardware your firm buys and owns, $250K to $350K at this size, plus a managed fee per GPU-month for refresh, monitoring and on-call. We quote it only if the review finds clients whose guidelines require it.
You pay for the clauses you have, not for the lawyers you employ.
Stack and integrations
The policy engine runs in your own cloud account or colo. Retrieval integrates with iManage and NetDocuments, and the private stack runs on NVIDIA hardware. Matterwall is not an authorized reseller or badged partner of any of them.
What the review finds
We read the outside-counsel guidelines for your top 100 clients by revenue, map each AI clause to the matters it governs, and check them against the AI tools your lawyers already use. We propose every clause-to-policy mapping and your GC approves it, because whether ChatGPT Enterprise counts as a public tool is a legal question and not ours to answer. You get one table your GC and your CIO can both work from. Six clauses in a hundred is well above the 1.9% in the published research, and that is what we expect: these are the corporate clients who pay your rates, not the public bodies that fill the public sample.
Matterwall can enforce some of these by blocking the tool or refusing the prompt. Others it can only report, because the clause asks for something we have not built yet, such as a flag in the time entry. A few it can only detect, because the tool sits behind no control we hold, and Copilot for M365 is the usual example. The table says which is which.
Who you would be buying this from
Matterwall is built by Ilya Subkhankulov, in the Bay Area. He worked in fixed income pricing at Thomson Reuters and in client coverage at REDI and Goldman Sachs, which is where he first met the controls a regulated firm has to satisfy. From 2018 to 2022 he was in self-driving, as an engineer at Voyager and then a program manager at Cruise. He founded Celery before that.
- The policy engine and its log run in your boundary, never in shared tenancy
- No training on your data, ever
- Matter walls enforced at retrieval
- Audit log per prompt, per matter, per OCG clause
- Deploys inside your boundary. Your colo, your cloud account, or your building, so your controls govern the compute
- Every subprocessor we use holds a current SOC 2 report. The list goes out with your questionnaire
- Client guidelines are read by people under NDA, never fed to a public model
- No SOC 2 report of our own, and no date for one yet. Your data stays in your boundary, our subprocessors are all SOC 2, and we will sign the audit-rights clause in your MSA. That is the closest we can get today.
- No named client references yet.
- No published eval suite of our own. We point you at third-party benchmarks instead.
Security questionnaires and architecture review go to audit@megalabs.co, answered by the person named above rather than a form.
Find out what your AI is touching before a client asks.
Addressed to your general counsel or risk partner, because they own the guidelines and the privilege. A $25K engagement, free for the first 20 Am Law 51-200 firms while we build our first cohort. Two weeks, run entirely under your counsel's direction.
- Week oneGuidelines in, clauses mapped
Engaged through your GC or your outside ethics counsel and delivered at their direction. We take the guidelines for your top 100 clients by revenue and pull out every clause that touches AI, disclosure, or data handling. A hundred sets of guidelines is more than one person can read in a week, so a model running inside your dedicated store does the first pass and named people check every clause it flags and sample the ones it did not.
- Week twoMatched against what your lawyers use
We map each clause to the matters it governs and check it against the AI tools already in the building. We propose the mapping, your GC approves it. You get the matter-by-matter table, as attorney work product. Your GC decides who sees it.
- Client guidelines are read inside a store dedicated to your firm, by named people under NDA and by a model running in that same store. Nothing is sent to a public AI model, ours or anyone's, and no prompt containing your clients' paper leaves your boundary.
- Documents stay encrypted at rest in a dedicated store for your firm, and are deleted on request or at the end of the engagement, whichever comes first.
- No copies leave that store. We do not train on them, benchmark on them, or quote them to anyone else.
- Week two runs under your counsel's engagement letter and at their direction. Whether the findings are privileged is your counsel's determination, not ours, and we will not tell you otherwise.
- If your GC wants different terms, say so on the call. We would rather change the terms than lose the review.