1. The Expensive Illusion of "AI Adoption"
Vendor dashboards flatter everyone involved. Active seats climb after a lunch-and-learn, prompt counts jump whenever a new model drops, and token volume makes for an encouraging slide at a board update. None of those numbers tell a CFO, a managing partner, or a general counsel whether any real work is getting done faster or more reliably than it did six months ago.
Where the math usually breaks down is in review and rework. Take an AI-generated contract summary that glosses over a survival clause or an indemnity carve-out. A senior associate who expected a twenty-minute head start ends up spending forty minutes reading the underlying agreement line by line to untangle a confident, plausible-sounding draft. A research memo with one invented authority taints every real case on the page until a person pulls each citation. When Stanford researchers ran a preregistered study of leading dedicated legal AI products in 2024, even tools marketed as "hallucination-free" still returned fabricated or misgrounded answers at substantial rates (Magesh et al.). Newer models have improved since then, yet engineering teams building with them keep running into the same constraint: in LangChain's latest industry survey, output quality remained the top barrier keeping agent workflows out of production (LangChain, State of Agent Engineering). When checking the machine takes as long as doing the job by hand, a paid seat is just another software subscription sitting on the shelf.
Unreliable tools rarely fail loudly, either. When an approved workspace truncates long PDFs, mangles table formatting, or hedges every answer, busy people seldom file a support ticket. Some quietly go back to working the old way while the organization keeps paying for their license. Others open a personal account in another browser tab and paste the work there instead, moving confidential contracts, financial models, or client drafts outside the company's vendor terms and audit logs because the sanctioned tool could not clear the bar for daily work.
At renewal time, login counts leave all of that hidden. A better starting question for any team lead is straightforward: which specific tasks leave someone's desk after a light review, and where are your people still redoing the output from scratch?
2. Why Defaulting Every Step to a Frontier Model Wastes Seats and API Spend
Look under the hood of almost any document workflow and you will find five or six modest jobs masquerading as one hard problem. Before anyone weighs a thorny legal or commercial question, something has to clean up text from an uneven scan, identify what kind of document arrived, pull out the parties, dates, and dollar figures, compare those terms against your house playbook, and assemble a clean note for a reviewer. Only one or two of those steps call for slow, heavy reasoning. The rest are structured sorting and extraction.
Off-the-shelf workspaces and early internal prototypes rarely draw that line. Pointing every prompt and background call at a provider's flagship model is the quickest way to get a demo working on Friday afternoon, so that is where most builds stay. Six months later, the team is still paying flagship token rates and waiting on flagship response times just to pull dates off an invoice or classify an intake attachment. On Anthropic's published price list, input tokens on its largest tier cost several multiples of its lightest tier (Anthropic, Pricing), and comparable spreads exist across OpenAI and Google. Research on model cascades and routing, including FrugalGPT and RouteLLM, backs up what practitioners see in production: letting smaller, faster models handle routine extraction while holding heavier reasoning models for genuinely ambiguous calls keeps answer quality steady without burning budget on clerical steps.
Choosing models step by step also spares you a quieter operational headache down the road: model retirement. Every major provider retires model versions on a published calendar. Anthropic returns errors once a deprecated snapshot passes its retirement date (Anthropic, Model Deprecations), OpenAI posts shutdown schedules across its API catalog (OpenAI, Deprecations), and Google assigns an explicit retirement timeline to each Gemini release on its agent platform (Google Cloud, Model Versions and Lifecycle). Workflows built around the habits of one pinned model version turn each deprecation notice into an unplanned scramble. Defining each step around the job it performs, backed by a small test set, lets an engineer qualify a successor model in an afternoon.
Where to split routine steps from heavy reasoning depends on your own documents, your tolerance for missed nuances, and which models your vendor contracts actually cover. What leadership can ask right away is simpler: did someone choose the model behind each step deliberately, or is the whole workflow still running on whatever default was used in the prototype?
3. Context Stuffing, Caching, and Retention Terms
Huge context windows invite an obvious shortcut: drag an entire closing binder or discovery folder into the chat window and ask your question. Because every file went into the upload, the process feels thorough, right up until the answer overlooks a decisive exhibit and bills for every file in the folder on each follow-up turn.
Researchers have been measuring this failure mode since long-context models arrived. When Nelson Liu and his colleagues tested how language models use long inputs, accuracy peaked when the key fact sat near the very beginning or the very end of the prompt and dropped sharply when the same fact sat in the middle (Liu et al., Lost in the Middle). Larger windows have not erased that curve. In 2025, Chroma evaluated 18 newer models, including GPT-4.1, Claude 4, and Gemini 2.5, and found that reliability still degraded as input length grew, even on straightforward tasks (Chroma, Context Rot). Performance varies across models, of course, yet anyone who has watched an AI tool summarize twenty exhibits while missing the one side letter that altered an indemnity cap has seen context rot in the wild. Worse, token billing meters the noise right alongside the signal, so uploading twenty agreements to ask about one clause means paying to process nineteen irrelevant files on every turn.
Careful workflow design narrows the material before the model ever sees it, retrieving only the passages relevant to the current step and preserving document and page provenance, so that a human reviewer can verify the source in seconds. You pay for fewer tokens and give the model far less clutter to misread.
When a workflow repeatedly sends the same large reference files or system instructions, providers offer prompt caching to lower the bill: the service temporarily holds processed portions of an earlier prompt on its servers so follow-up calls can reuse them at a discounted rate. On public or low-sensitivity material, that is an easy savings, but when confidential corporate records or client files enter the workflow, a cache becomes yet another place where derived copies of your data sit on vendor infrastructure. Each major provider sets different defaults that have to be squared with confidentiality obligations and any Zero Data Retention (ZDR) terms in your contracts:
- Google Cloud turns automatic caching on by default on its Gemini Enterprise Agent Platform (formerly Vertex AI), keeps those cached entries for up to 24 hours, and does not support customer-managed encryption keys on its longer-lived cache storage. Google's documentation notes that keeping prompt data out of cache storage requires turning automatic ("implicit") caching off and leaving manual caches disabled (Google Cloud, Context Caching Overview; Google Cloud, Zero Data Retention).
- Anthropic leaves API prompt caching off until a developer enables it, clears cached entries after five minutes by default (with a paid one-hour option), and treats caching as eligible for ZDR except on its designated covered models (Anthropic, Prompt Caching).
- OpenAI caches repeated prompt prefixes automatically and offers an extended setting that holds cached prompt data for up to 24 hours, adjusting its default retention behavior based on whether ZDR is active on the organization (OpenAI, Prompt Caching).
Caching can still be well worth using once legal and engineering agree on where it belongs. The catch is that the cloud setting that trims your monthly API bill is often the same setting that governs how long derived copies of your prompts remain outside your environment, and the developer enabling it rarely has your enterprise retention addendum open on their desk. Our procurement guide to Zero Data Retention and enterprise AI subscriptions walks through how those settings line up with contract language, model by model.
4. Why "Vibe Checks" Fail: Defining Usable Work Before Scaling Spend
Watch how most AI tools win internal approval, and it is easy to see why enthusiasm fades over time. A vendor gives a polished thirty-minute walkthrough, a couple of champions on the team test five prompts on files they already know by heart, the prose looks crisp, and procurement signs the order form. Engineers have a blunt name for that kind of evaluation: a vibe check.
Testing by feel flatters almost any modern language model because people naturally try clean, familiar files, leaving out the crooked scans, fourth-round contract redlines, and tangled email chains that eat up real hours. Worse, an informal trial runs once. Nothing stays in place to catch drift three months later when a provider updates model weights behind an endpoint, or when someone on your team edits the system instructions to fix one complaint and quietly breaks two other tasks. And because nobody wrote down what an acceptable result required before the pilot started, two experienced reviewers can read the same AI draft and reach opposite conclusions about whether the tool is helping.
A workable standard takes an afternoon to set up. Pull a small packet of real, redacted documents from the workflow you plan to support, write down in plain English what a sound output has to capture and what errors disqualify it, and score each run on two practical measures: did it get the substance right, and how long did a qualified person spend turning the output into finished work? Re-running that same packet whenever a model version, prompt, or template changes is the everyday version of what the NIST AI Risk Management Framework and its Generative AI Profile describe under the Measure function.
Putting those two numbers on paper gives engineering, finance, and legal a shared baseline almost immediately. Teams can see which workflows save real review hours and which ones stall in the classic pilot trap: drafts that look nearly finished at first glance yet still force a lawyer, analyst, or operations lead to verify every line against the source. When a workflow stalls there, tightening retrieval, system instructions, and source citations usually moves the review clock far more than upgrading to a pricier model.
5. Unattributed Spend Is Both a Budget Problem and a Governance Problem
Ask most leadership teams what they spent on AI last quarter for a specific department, workflow, or client matter, and you will usually get an invoice total followed by a long pause. Seat licenses sit on one company-wide purchase order, internal tools share a single production API key, and individual subscriptions trickle through employee expense reports alongside travel and software renewals.
Inside a company, finance and operations leaders increasingly treat that blind spot as a primary control gap. In the FinOps Foundation's 2026 practitioner survey, AI cost management ranked first among skills cloud and finance teams said they needed to develop (FinOps Foundation, State of FinOps 2026). The Foundation's guidance for AI states the order of operations plainly: allocate spend to real owners first, right-size second, and "measure the use case, not the token" so leadership can see what a completed unit of work actually costs (FinOps Foundation, FinOps for AI). For a CFO, COO, or general counsel reporting to the board or an audit committee, that allocation puts AI spend on the same footing as any other auditable operational control, and it surfaces shadow AI whenever a busy group logs almost no activity on the company's approved tools.
Inside a law firm, from a solo practitioner carrying one paid seat to a multi-office partnership, knowing which tool ran on which matter also supports billing and supervisory duties that the lawyers themselves must apply. Under ABA Formal Opinion 512 (addressing Model Rules 1.4 and 1.5), a lawyer billing hourly may charge for the time actually spent prompting, reviewing, and revising AI-assisted work, never for hypothetical hours the tool saved. General-purpose AI workspaces resemble library or word-processing subscriptions and will ordinarily be treated as firm overhead, whereas identifiable third-party charges incurred for a specific client's matter may be billed through at actual cost, absent an advance agreement otherwise, when they are reasonable and properly disclosed to the client. (ABA opinions are persuasive guidance, and the ethics rules adopted in each lawyer's jurisdiction govern.)
Those billing, disclosure, and ethics calls always stay with the firm, or with an organization and its counsel; per-matter and per-workflow metering supplies the factual record needed to make them. Without logs showing which model, cache setting, and workflow touched a matter, a managing partner cannot separate firm overhead from a client-specific disbursement with confidence, and leadership lacks clean operational records for its own reviews of competence, confidentiality safeguards, and non-lawyer supervision under Model Rules 1.1, 1.6(c), and 5.3, duties we examine further in Can Lawyers Use Generative AI for Client Work?.
6. Where This Leaves Your Organization
When an AI budget grows faster than the finished work coming back, buying another block of seats or waiting on the next model release rarely changes the math. Stalled rollouts almost always trace back to four fixable gaps in how the work was wired together: routine steps running on oversized defaults, whole folders crowding the prompt, unwritten standards for acceptable drafts, and one blended invoice at month's end. Closing those gaps lowers what you spend per completed job while producing the operational record a client, a board, an auditor, or a regulator expects to see.
Before renewing a vendor agreement or approving the next rollout, pick two or three workflows your team actually relies on and put six questions on the table internally and with your providers:
- Across the seats and API keys we pay for today, which specific tasks actually leave someone's desk faster after human review, and where are people redoing the output or turning to personal accounts?
- In each workflow we run, who chose the model behind each step, and which routine sorting or extraction steps are still paying frontier rates out of habit?
- When a provider retires a model version we depend on, how do we find out, and can we test and swap in a successor without rewriting the workflow?
- What goes into the prompt on a typical run, and do our active prompt-caching, grounding, and memory settings match the retention commitments in our client and vendor contracts?
- Have we written down what acceptable output looks like on a packet of our own redacted documents, tracking both accuracy and human review time whenever models or instructions change?
- Could we show, today, what we spent on AI last quarter by department, workflow, or client matter, or can we point only to a top-line invoice?