Indian law firms are already using artificial intelligence (AI) and clients have started asking for their bills to reflect that. At a panel in New Delhi on August 17, Hyundai Motor India's General Counsel Amitabh Lal Das said that law firms should bill fewer hours once AI lets them finish research faster. Das said he had, in his own experience, been charged research fees disproportionate to the advice produced and that he expected AI adoption to bring down companies' external legal spending generally.
The instinct behind cutting the bill because the work got faster is understandable. It is also the wrong place to start. A blanket discount treats every hour of AI-assisted work as equally cheapened, when the truth is closer to what Senior Advocate Chander Uday Singh raised at the same panel: generative tools occasionally invent case law that was never delivered and shape answers around what the questioner wants to hear rather than what the law actually holds, a tendency juniors are least equipped to catch. The real answer to Das's demand does not sit in a percentage knocked off in an invoice. It sits inside every firm's own time-sheets.
Das's concern is the saving created by faster research. Singh's concern is what happens between an AI-generated answer and the advice an advocate is actually prepared to sign. Neither position is wrong, but taken alone, each creates a problem: Das's view gives a client no basis for sizing the discount it wants and Singh's view leaves a firm treating every AI output with the same suspicion, whether or not it deserves it.
A legal fee has traditionally bundled 3 things into one rate: legal research, the professional judgment involved in choosing which defensible position to advise on and the risk of signing that advice. These were hard to price separately, so the billable hour stood in for all three. Generative tools can now perform a substantial part of the research function and have left judgment largely untouched, since telling a sound conclusion apart from a merely fluent one is exactly what these systems do least reliably. They have also created a new task: checking the machine's output before it reaches a client or a court. That task does not appear on any Indian invoice as a line of its own. The firm's experience has not become less valuable; what has changed is the cost of producing the first draft.
Three large corporate clients abroad have already published outside-counsel billing rules addressing this exact question. Zscaler, the American cybersecurity company, requires any AI-generated work product to be reviewed by a human attorney and says that the time and cost associated with generating it should not be passed on to the client. UBS, the Swiss bank, went the other way: its outside lawyers need client consent before using AI at all, learning time is non-billable and AI use must be disclosed on the invoice. The Hartford, the American insurer, requires the programme's name in the time entry and pays only for actual prompting and review time. Meta has no comparable public policy, but its legal operations chief has told Law.com that firms should stop charging 8 hours for work a model finishes in 30 minutes.
Each of these regimes assumes billing infrastructure that is unevenly developed in India: electronic billing, task codes and invoices capable of being audited line by line. Rate surveys show a market split between firms that still bill mostly by the hour and firms whose partners call the hour close to obsolete, with retainers and blended rates covering the rest.
That split pushes the real question up a level, to the retainer itself. A retainer negotiated in 2023 may now cover work that costs materially less to produce, with nothing prompting either side to revisit its terms. The provision that constrains fee structures here is Rule 20 of the Bar Council of India Rules, which bars a fee contingent on the outcome of litigation or a share of the recovery. AI disturbs none of that. Fixed fees, capped fees, staged fees, blended fees and retainers all remain fully permissible. A productivity-based review clause can be built into any of these structures, subject to terms agreed between client and counsel.
Indian regulators are not entirely silent on AI, though the silence on fee disclosure persists. The Bar Council of India told parliament in June 2026 that responsibility for AI-generated errors, including fabricated citations, rests with the advocate and not with the technology used. However, it didn't say anything about how a firm should price or disclose AI use on a client's bill. This stays private matter for the engagement letter, including what the Digital Personal Data Protection Act, 2023 implies about client material feeding a model. That is a data fiduciary question now, not merely a confidentiality one.
A blanket refusal to pay for anything a machine touched cannot realistically be enforced. No client can read a finished opinion and tell whether a human or a program produced its first draft. There is a substantive objection too: Khaitan and Co has built a proprietary platform on its own research. Cyril Amarchand Mangaldas adopted Legora firm-wide after piloting three platforms across 380 lawyers. A system grounded in a firm's own matter history can produce work that a generic subscription tool cannot reproduce without access to that institutional knowledge. Refusing to pay for it penalises the firms with the deepest knowledge base.
No market has fully resolved this either. The 2026 CLM Litigation Management Study, surveying more than 70 American claims executives, found that 87.9 per cent had billing guidelines silent on AI.
What can genuinely be verified is the cost of production. Many firms already hold much of this data, since time is recorded internally even where the client is billed through a retainer. Ask for the firm's baseline: matters closed before these tools arrived, by work type, with median hours and seniority mix, set against today's figures.
One possible allocation would follow how predictable the work is. Vendor agreements, employment paperwork, routine filings and first pass diligence have a knowable scope, so the firm bears little risk and could pass most of the saving on. Work within a firm-absorbed cap would split between the parties. Open-ended work leaves the firm carrying real exposure, which gives it the stronger claim to what it saves. Under this model, the client gets a verifiable figure instead of a pressured discount and the firm keeps getting paid for the judgment and institutional memory AI has not displaced.
Assume AI has made legal work interchangeable with software output and a filing eventually gets built around a citation that does not exist. The deeper risk sits inside the firm, not just in that assumption: treating verification as unpaid, invisible labour lets juniors skip the exact exercise that slowly turns them, over years, into the partners whose judgment the firm later sells - catching a fabricated citation before it reaches a court.
The fix is verification as a named, budgeted step: a specific lawyer accountable for review, a time entry distinct from drafting, a sign off that is recorded rather than assumed. Put it on the invoice as its own line, as The Hartford already requires, and the client pays for what actually keeps a fabricated citation out of a filing, rather than for a vague entry called research.
UBS and Zscaler now give Indian in-house counsel concrete templates if they choose to introduce similar provisions at the next renewal, though most will need to rebuild them around a productivity ratio, since a retainer does not expose the productivity change in the same way an hourly account does. Retainers agreed before widespread adoption of these tools may now embody production assumptions that no longer hold. This gets settled letter by letter, well ahead of any regulation.
An engagement letter that keeps pace with this should cover five items:
Disclosure of material AI use in a clear, searchable form
A bar on client data training any model
Learning time treated as firm overhead rather than billed to the client
Verification recorded under its own time code with a named reviewer and
A scheduled retainer review at a fixed interval.
The ACC's sample AI guidelines and Singapore's sector guide are usable starting drafts, adapted to Rule 20 and to a retainer-based practice.
Clients should be careful what they ask for. In the same CLM study, executives ranked cost third from the bottom in what makes a firm valuable - behind communication, resolution focus, legal skill and relationship. Hourly billing scored 52 out of 100 for aligning firm behaviour with client interests, down from 66 three years earlier. Asked who should absorb AI tool costs, half said the firm, half had no policy and none said their own organisation should.
The advantage of a verifiable number is practical, not rhetorical. It gives both sides something to examine instead of an assertion about what AI ought to save and it makes verification payable rather than buried inside a vague entry called research. It also gives the firm a clearer measure of what different classes of work cost to produce. Without that information, a fixed fee rests on an estimate and an arbitrary discount says little about the saving AI has actually created. What is on offer instead is a disclosed method, institutional knowledge priced for what it still is and a figure both sides can verify.
Firms have always kept cost data in their own accounts. AI gives firms a reason to turn those internal cost records into a pricing tool, rather than filing them away once a matter closes. That conversation is likely to start well before any regulator requires it.
Prasanth Raju is an advocate of the Bombay High Court.