An in-house legal AI copilot is a chat-and-document assistant that a corporate legal department points at its own material — its executed contracts, its playbooks, its policies, its prior positions — and uses across three jobs: reviewing inbound documents, answering research questions, and producing first drafts. The copilot is the department’s generalist. You hand it a document or a question, and a lawyer checks what comes back before it leaves the building.
What it is not
It is not a CLM. A CLM owns the contract record — the repository, the approval routing, the signature, the renewal date. A copilot has opinions about a contract and no authority over it. Teams that buy a copilot expecting workflow enforcement end up with neither.
It is not a legal research platform. Westlaw and Lexis sell curated, citation-linked authority; a copilot grounded in your Google Drive does not. It is not matter management or e-billing, which track spend and outside-counsel work rather than produce it. And it is not an autonomous agent — in the LegalOn and In-House Connect 2026 State of AI for In-House Legal survey of 452 in-house professionals, the majority preferred supervised, human-in-the-loop automation over systems that act alone.
The three jobs
Review. You drop in a counterparty NDA, DPA, or MSA and get a redline against your playbook, plus a list of the positions that fall outside it. This is the highest-volume use in most departments, and it is what the LegalOn survey measured at 52% of in-house teams either using or evaluating AI for contract review.
Research. Not case law — internal research. What did we agree to with this customer in 2023? Which of our contracts carry an uncapped indemnity? What does our own data-processing policy say about sub-processors in the EU? The answers live in the department’s own corpus, and nobody outside it can retrieve them.
Drafting. A first-pass DPA amendment, a board memo, a response to a customer’s security questionnaire. The copilot writes the version a lawyer edits rather than the version that ships.
The FTI Consulting and Relativity General Counsel Report released 11 March 2026 puts the same three jobs in the same order by adoption: summarization at 83%, clause identification at 63%, transcription at 53%.
Why “in-house” is a real category, not positioning
Harvey and Legora were both built inside law-firm workflows first — M&A diligence, cross-jurisdictional advisory, litigation-grade document analysis at scale — and both are now selling into corporate departments. Harvey reported 500+ in-house legal teams among 1,300 organizations when it raised $200M at an $11B valuation on 25 March 2026. Legora reported 800+ law firms and in-house teams across 50+ markets alongside its $550M Series D at a $5.55B valuation on 10 March 2026.
Those are real deployments, and the origin still shows in the product. Firm-first tools are built for depth on a single high-stakes matter, priced against billable recovery, and staffed by lawyers whose full-time job is that matter. An in-house team of four handling 200 contracts a quarter has the opposite shape: shallow work, enormous breadth, no billable hour to charge the tool against, and a queue that never empties.
That gap is what in-house-first vendors sell against. GC AI was built by a three-time general counsel and organizes around reusable “skills” for recurring department requests. Eudia pairs software with acquired legal-services capacity — it bought Johnson Hana and its 300+ legal professionals — and sells the combination to Fortune 500 departments. Spellbook runs inside Microsoft Word, where in-house lawyers already redline, and includes a playbook-build service in its in-house packaging.
How it shows up in the work
The LegalOn survey found in-house teams spend an average of 3.1 hours reviewing a single contract. A four-lawyer department clearing 200 contracts a quarter is spending roughly 620 hours on review alone — about 1.2 full-time lawyers of capacity, before a single novel question gets answered.
The copilot’s job is the first 80% of that pass: extract the terms, flag the deviations from playbook, propose the redline. What it does not do is decide whether the deviation is acceptable for this counterparty at this revenue. That is the part you are paying lawyers for, and the reason the department headcount does not drop when the copilot lands.
Buying criteria that separate vendors
| Criterion | The question that exposes it | What a real answer looks like |
|---|---|---|
| Corpus access | Which repositories does it read, and how does the connector handle permissions? | Named connectors for your DMS, Drive, or CLM, honoring source-system permissions per user |
| Grounding | When it cites our own contract, can I click through to the clause? | Every assertion links to a document and page, not a paraphrase |
| Playbook fidelity | Whose playbook — yours or the vendor’s default? | You upload positions and fallbacks; the tool reproduces them without re-prompting |
| Word surface | Do lawyers work in the tool or in Word? | Redlines round-trip through Word with tracked changes intact |
| Data terms | Is zero data retention contractual or aspirational? | A signed ZDR agreement plus SOC 2 Type II, both in the MSA |
| Seat economics | What happens when a non-lawyer in Sales wants access? | A viewer or requester tier that does not cost a full lawyer seat |
The corpus and grounding rows matter most, and they are the two vendors resist answering precisely. A copilot that cannot retrieve your prior positions is a general-purpose model with a legal skin, which you can buy from Claude for a fraction of the price.
Cost reality
Published pricing is rare in this category, and the spread is wide. GC AI lists an Individual plan at $500 per month for solo and fractional GCs, with Team priced on request and annual billing that adds enterprise SSO, a shared skill library, and US case law. Spellbook publishes no number at all — pricing scales with the number of team members on the license, behind a 7-day trial. Harvey and Legora both quote through sales only.
Secondary reports put Harvey’s in-house seats in the four-figures-per-month range with a seat minimum and an annual term; treat that as an estimate rather than a planning number, because none of it is vendor-published. The planning number you can defend is the one you build from your own volume: hours of review per quarter, blended lawyer cost, and the share of that pass the tool actually removes.
Common pitfalls
Buying the firm-grade tool for a volume problem. Depth pricing against breadth work produces the worst cost-per-document in the category. Guard: score the shortlist on your five highest-volume request types, not on your hardest matter.
Treating the copilot as a research platform. The Stanford RegLab study of leading AI legal research tools (Magesh et al., 202 queries, published in the Journal of Empirical Legal Studies) found hallucination rates of 17% to 33% even in purpose-built, citation-linked products. A general copilot over your Drive has no citation layer at all. Guard: route external-authority questions to a research product and keep the copilot on internal corpus questions. See legal AI grounding vs hallucination.
Skipping the playbook work. A copilot without your positions loaded reviews against a generic standard and produces redlines your team rejects. Guard: budget two to three weeks of a senior lawyer’s part-time attention to encode positions and fallbacks before measuring the pilot. The contract review SOP is the input.
No usage floor in the contract. The FTI report found formalized technology roadmaps at an all-time high of 53%, more than double the prior year’s 25% — which means nearly half of departments still buy without one. Guard: set a per-seat weekly usage threshold at 30 days and a reallocation rule for seats that miss it.
Do you need one?
Under about three lawyers, an enterprise general-purpose assistant plus a written AI policy covers the ground at a tenth of the cost. From three to roughly fifteen lawyers with a recurring contract queue, an in-house-first copilot is the default, and the volume math above is the business case. Above fifteen, or where the work is litigation- and diligence-heavy rather than contract-heavy, the firm-first platforms earn their premium.
If contract workflow is the actual complaint — approvals, versions, renewals — buy a CLM first. A copilot on top of a broken intake process makes bad routing faster.
Related
- In-house vs outside counsel — what stays inside determines what the copilot has to cover
- Legal AI vs legaltech — where copilots sit in the wider vendor landscape
- AI policy for legal teams — the authorization layer that has to exist before rollout
- Legal Ops maturity model — the stage at which copilot investment compounds
- Legal intake — the front door the copilot triages