The phantom savings
A managing partner does everything the vendor recommended for the shiny new AI tool that has “guaranteed ROI.” The tool is deployed across document summarization, document review, research, and contract analysis. The efficiency data looks great: the firm saved 10,000 attorney hours last year. At the firm’s blended rate of $400 per hour, that’s $4 million in value. The contract, of course, deserves to be renewed, because the vendor was right. Except they weren’t. When the CFO runs the year-end numbers, the $4 million isn’t there. Both revenue and profit are flat; they might have even declined a little. The hours saved didn’t translate to retained value.
This is the legal industry’s version of phantom income—the tax concept where you have income recognized for tax purposes without a corresponding cash distribution. The vendor’s ROI calculator said you earned $4 million; your P&L says otherwise, and you’re about to renew the contract on a tool that hasn’t banked you any savings.
The question isn’t whether AI saves time. The experimental evidence shows that it does on certain tasks. The question is: even when you use AI for the ideal tasks, what happens to the time after it’s saved?
The value of a saved hour
TL;DR: Time saved is not value captured. What an AI-saved hour is actually worth depends on billing economics and whether freed capacity finds new work. Deployment architecture affects both cost and privacy exposure. There is no universal “best” AI deployment—the answer depends on your firm.
In computer architecture, we have a fundamental concept called Amdahl’s Law used for analyzing the achievable speedup from an optimization. It states that the speedup from an optimization is limited by the fraction of the task that cannot be optimized. That is, when a computer vendor tells you that their new processor is faster than the old one by 10x, if your application spends 60% of its time on input/output operations, that 10x speed might only get you a 1.56x speedup. The vendor’s topline number addresses an issue that is not captured by your specific application scenario.
This same structural problem applies to AI in law firms. When a vendor tells you AI will make your attorneys 30% more efficient, that number—like the processor’s 10x—describes a task-level result under particular conditions, not a firm-level financial return you’re guaranteed. The actual gain depends on what fraction of your firm’s economics can absorb the improvement. Most importantly, you’re limited by the fraction that can’t be improved, and it might be larger than you realize.
Why is it that even though AI seems to work and your lawyers are getting more efficient, your firm is not seeing the effects, and might even be economically worse off? I argue here that the answer lies in the calculation of the value of the savings from AI efficiency. Simply put,
$$ \text{Value of saved hour} = \text{cost avoided} + \text{value of redeployment -} \\ \text{hourly revenue displaced.} $$For you nerds out there, we can express this value as \( \pi \):
$$ \pi = \underbrace{(1-f_h)C_a}_{\text{avoidable cost}} + \underbrace{V_d M}_{\text{redeployed capacity}} - \underbrace{f_h R_h q_h}_{\text{forgone hourly revenue}}, $$where \(f_h\) is the fraction of AI-eligible work performed under hourly billing, \(C_a\) is the avoidable cost per saved hour of non-hourly work, \(V_d\) is the probability that freed capacity is redeployed to additional revenue-generating work, \(M\) is the incremental contribution generated by one redeployed attorney hour, \(R_h\) is the nominal billing rate for hourly work, and \(q_h\) is the fraction of displaced hourly work that would otherwise generate collectible revenue.
In plain English: the value of one AI-saved hour equals the cost you actually avoid on non-hourly work, plus the incremental contribution from redeploying your freed attorneys’ capacity on new matters, minus the hourly revenue you can no longer bill.
So, you have three forces pulling in different directions. Two help firms capture value; one takes it away. Your ROI is driven by how these forces interact and your ability to mitigate the forgone hourly revenue.
To see how this works, consider a hypothetical small litigation firm that bills 95% of its hours at $275/hour, collects 90% of what it bills, its avoidable cost on non-hourly work is $55/hour, and only 25% of freed attorney time actually lands on new revenue-generating matters (at $220 contribution per redeployed hour). If we run the \( \pi \) formula, each AI-saved hour doesn’t generate value; it destroys $177 of it, even though the vendor’s calculator, which simply multiplied saved hours by the billing rate, said the same hour was worth $275.
These input numbers aren’t universal constants. Your billing mix, collection rate, avoidable cost, and demand pipeline are all firm-specific, meaning a different firm with different numbers will get a different answer.
Your vendor’s ROI calculation is not a forecast until you calibrate it to your firm’s billing mix, collection rate, avoidable cost, and redeployment demand.
The formula tells you the answer, but doesn’t tell you what to do about it, because each variable buries a strategic question that you must ask to pragmatically deploy AI.
What determines whether your billing mix helps or hurts? Can your firm actually redeploy freed capacity, or does it just become slack? Why does the same architecture decision that sets your deployment cost also determine your privacy exposure? What do plausible compliance events actually cost after you account for controls and insurance? And when you put it all together, how much of the vendor’s headline ROI number survives your firm’s actual economics and risk profile?
These five questions are the five lenses of a diagnostic framework I call PRISM. It does what a prism does: it separates the white light of a vendor projection into the spectrum of what’s actually happening inside your firm.
Pathway Economics: What is a saved hour actually worth?
When AI saves an attorney an hour, that hour does not automatically turn into retained value. The saved hour passes through one of three paths, each with a different effect on the bottom line, and your firm’s billing structure determines which path generates value.
An important caveat is that the saved hour must discount the time spent verifying, correcting, or otherwise reviewing the AI-generated work. If AI saves 30 minutes but the attorney spends 10 minutes checking citations and fixing the output, only 20 minutes of economically usable capacity was actually created. The framework accounts for this verification overhead before the saved capacity enters the three pathways below.
Path 1 is the avoided cost. On work that isn’t billed hourly, like flat-fee matters, subscriptions, value-based arrangements, finishing faster means you spent less labor delivering the same revenue. In theory, that can improve matter economics without reducing revenue. In practice, less labor doesn’t always mean less cost. Your associates and support staff still need to be paid, and the office still needs to be run. Rent still needs to be paid. The actual cost you avoid turns out to be a lot narrower than expected: contract-attorney expense, overtime, variable staffing, maybe a hire you no longer need to make. In the \( \pi \) formula, that’s \( C_a \), and it might be a much smaller number than anticipated.
Path 2 is the incremental contribution from redeployed capacity. Here’s where the freed attorney picks up new work, a client matter that was waitlisted gets staffed, or a pitch that was going to be declined gets resourced. Simple ‘hours \(\times\) billing rate’ calculators effectively assume that every saved hour becomes economic value. That assumption can be badly wrong, for reasons we’ll get to next. For now, the point is that this pathway (\( V_d \times M \) in the formula) requires two things to be true simultaneously: demand has to exist, and your firm has to be organized to route freed capacity toward the demand.
Path 3 is revenue you can no longer bill. For work billed by the hour, an AI-saved hour is an hour you cannot charge for. The matter still gets done, but you must now invoice fewer hours than before. ABA Formal Opinion 512 makes this explicit: efficiency savings from AI must be reflected in billing. This billing headwind is the term in the formula that subtracts \( f_h \times R_h \times q_h. \) For firms that bill most of their work hourly, this term can dominate the other two terms.
This path creates a contradiction in AI deployment economics: the firms with the strongest competitive pressure to adopt AI quickly—large, profitable, high-utilization, hourly-dominant practices—face the strongest billing headwind, and whether it becomes a net loss depends critically on whether freed capacity can be redeployed. Their billing model converts efficiency gains to revenue losses, and the very thing that made them successful can make AI adoption a financial headwind rather than a tailwind.
This contradiction shows up in industry data without showing much evidence of a way out. Clio’s 2024 Legal Trends Report estimated that 74% of hourly billing tasks are potentially automatable by AI, and that, for each hourly-billed lawyer, AI automation puts approximately $27,000 per year at risk. That $27,000 illustrates the Path 3 problem: revenue at risk from doing the same work faster. The obvious response is to shift toward alternative fee arrangements. In 2024, 39% of law firm lawyers told Bloomberg Law that they expected AI to accelerate the move to alternative fees. In 2025, however, only 9% of respondents observed any such shift. Bloomberg’s own read is that it may be too early to tell. Maybe. But clients aren’t waiting for the data to come in. Instead, Thomson Reuters’ 2026 data shows firms growing revenue through rate increases, not operational restructuring—working more, at higher rates, within the same billing model. While that might be a viable short-term strategy, it might be unsustainable, given that nearly 60% of in-house counsel report “no noticeable savings yet” from outside counsel’s AI use. Among those reporting no savings, 58% say firms haven’t adjusted pricing to reflect AI efficiencies. More broadly, 64% of respondents expect AI to let them rely less on outside counsel.
What is your \( f_h \)—the fraction of your AI-eligible work that is billed hourly? If 90% or more is billed hourly, every saved hour creates a substantial billing headwind. That doesn’t mean AI is a bad investment; it means you face a billing headwind under your current model, and whether you overcome it depends heavily on redeployment capacity. You either need to change the model, build redeployment capacity, or accept that you’re adopting AI for competitive positioning rather than near-term profit.
The firms with the strongest perceived incentive to adopt AI quickly face the strongest billing headwind that can create a net loss if freed capacity can’t be effectively redeployed.
Redeployment Capacity: Can freed capacity actually find revenue-producing work?
Think of the last time an attorney on your team became unexpectedly available. Perhaps a deal fell through, a trial settled, or a project finished early. \(V_d\) defines the probability that freed capacity is redeployed to additional revenue-generating work. This depends on three things. First, demand pipeline depth: does your firm have work waiting for capacity, or are you already below full utilization? A firm turning away matters for capacity reasons has the demand-side conditions for a high \(V_d\), giving freed hours somewhere to go. A firm operating below capacity with no backlog has the demand-side conditions for a low \(V_d\), such that saved hours are more likely to evaporate. Second, staffing fungibility: can a freed litigation associate be redeployed to a corporate matter, or does your work require specialization that makes redeployment impractical? Here, full-service firms have more flexibility than boutiques. Third, organizational responsiveness: even with demand and fungibility, someone has to actually route the freed capacity to the waiting work. Firms with active matter-assignment processes redeploy faster than firms where staffing is partner-driven and ad hoc.
The financial sensitivity of \(V_d\) can be enormous. When I modeled an 85-attorney full-service firm and varied \(V_d\) from 0.25 to 0.65, the five-year risk-adjusted NPV went from -$7.6 million to +$3.6 million. That is not a rounding error. It is the difference between AI being a strong investment and AI being a write-off or even a loss, driven entirely by a behavioral variable that has nothing to do with the technology itself. The only thing that changed was whether freed hours were effectively redeployed.
When the attorney on your team becomes suddenly available, if they were on a new matter within a day or two, your \(V_d\) is probably healthy. If they spent time waiting for something to come in, or if their time was consumed by lower-value triage work or administrative catch-up, your \(V_d\) is probably weak. Scale that response across your entire firm, and you have a rough sense of where you sit.
The implication of this question is that you might not need better AI tools to improve your AI ROI, but better staffing systems. The return on investing in a matter-assignment process that matches freed capacity to waiting work may dwarf the return on the AI tool that freed the capacity in the first place.
The return on investing in a staffing process that routes freed capacity to waiting work may be higher than the return on the AI tool that freed the capacity.
Integrated Architecture: The same architecture decision changes both economics and privacy exposure.
Most firms evaluate their AI deployment via two independent decision systems. The technology team evaluates cost: cloud might be cheapest per seat, hybrid costs more, private infrastructure costs the most upfront. The compliance team evaluates risk: what data flows where, what the vendor’s terms say, what bar ethics opinions require. Both teams come up with their recommendations, which may or may not agree.
This is like buying a house by having your accountant check the price and your lawyer check the title independently, without comparing notes. Then being surprised when the cheapest house turns out to be in a flood zone.
Deployment architecture is the variable that links ROI and compliance exposure
Your deployment cost and your privacy and compliance exposure are linked by the same variable. Where inference happens, what data gets transmitted, who has access, what gets retained, what contractual terms govern the vendor’s use of your data—these are all properties of the deployment architecture. Cloud, hybrid, and private deployment architectures aren’t just cost tiers; they create different exposure profiles depending on how each configuration actually handles data. The cheapest option may carry exposure that, once priced, makes it the most expensive.
This isn’t an argument that cloud deployment is riskier. Nor should firms assume that private deployment is automatically safer; security depends on the controls and capabilities surrounding the architecture. A firm running its own infrastructure without adequate cybersecurity expertise may be a softer target than an enterprise cloud provider with a dedicated security team. The relevant question isn’t “cloud or private?” but “what are the actual data flows, controls, and contractual terms of this specific configuration, and what do they cost on both the financial and compliance sides?”
In my modeled 12-attorney boutique handling low-sensitivity commercial work, the distinction barely matters. The privacy exposure difference between cloud and hybrid is small in absolute terms, and the deployment decision is straightforwardly financial.
For a 500-attorney firm with government, financial-services, and healthcare clients, the picture is completely different. In my modeling, cloud is the cheapest option by roughly $1.8 million over five years. But incorporating expected privacy and compliance losses from the firm’s sensitive client mix reverses the ranking. Cloud’s financial advantage disappears, and hybrid wins on risk-adjusted NPV. For some high-sensitivity practice groups, private deployment can become the preferred option.
This ranking reversal is the whole point of integrating the two analyses. Reducing the modeled privacy and compliance losses by just 4% flips the ranking, with cloud outperforming hybrid. In this example, the firm’s architecture decision sits on a knife’s edge, and neither the technology team nor the compliance team can see the edge from its own silo.
A single firm-wide architecture may be wrong for diversified firms
The practical implication for diversified firms goes even further. A single firm-wide architecture may be the wrong answer entirely. Practice groups handling routine commercial work might belong on a cost-optimized cloud deployment, while groups handling government contracts or healthcare data might need hybrid or private infrastructure. The architecture decision is a portfolio of decisions matched to the risk profile of the work.
Your AI deployment ROI and compliance exposure are inextricably linked by the same variable: the deployment architecture.
Scenario-Based Risk: What do plausible privacy/compliance events cost after controls and insurance?
As I’ve written before, a vendor’s SOC 2 report is not a risk model. The important question is not just whether your firm has controls, but what those controls actually do, and what happens when they fail. This can be modeled in three layers: what’s the probability of an adverse event, what are the consequences if it happens, and how much of those consequences does your firm retain after controls and insurance?
This decomposition matters because not all controls do the same thing. Most law firms also don’t have reliable actuarial data on AI incidents, so it helps to think in terms of realistic scenarios and the range of potential outcomes. For modeling purposes, I separate controls into preventive, mitigation, and risk transfer. Preventive controls, like data classification, vendor security review, access restrictions, qualified IT governance, reduce the probability that something goes wrong. Mitigation controls, like incident response plans, containment protocols, reduce the severity when something goes wrong. Risk transfer, like cyber insurance, shifts a portion of the financial loss off your books. Treating all three as interchangeable produces a false sense of protection. A firm with excellent incident response preparation but no data classification is well-prepared to manage a breach it could have prevented.
A useful way to think about it is that your expected privacy loss for each adverse event type is the mitigated probability times the retained loss. The mitigated probability is your baseline risk (driven by your architecture and client sensitivity) reduced by your preventive controls. The retained loss is the potential damage (remediation costs, client attrition, regulatory penalties, contractual liability) reduced by your mitigation controls, then further reduced by what insurance covers after deductibles and exclusions.
In practice, this leads to asymmetric stakes. In my modeled small firm profile with low-sensitivity clients, expected privacy losses are modest across all architectures, and the governance investment needed to manage them is proportionally small. In the modeled large firm profile with a sensitive client base, the stakes increase dramatically. A single unauthorized disclosure involving government or healthcare data can produce remediation costs, contractual liability, regulatory penalties, and client losses that run into seven figures.
The counterintuitive insight here is about the ROI of governance. A firm that spends $50,000 on data classification and vendor review before deployment is buying down the expected value of a tail risk that could be orders of magnitude larger. Governance is not merely overhead in this framework. Rather, it’s an investment with a quantifiable expected return that may end up being worth more than just getting the architecture right.
Governance has a quantifiable risk-adjusted return. Getting the controls right may be worth more than getting the architecture right.
Measured Value Capture: How much of the headline ROI survives the firm’s actual economics and risk?
If you’ve made it this far, you probably see where this is going. Start with the vendor’s headline number (hours saved times billing rate); pass it through your billing mix (some of the value disappears through Path 3); pass it through your redeployment capacity (some more leaks out because freed hours become slack); adjust for your actual adoption pace, which depends on your infrastructure and your partners’ willingness to use the tools; subtract your deployment costs; then subtract your expected privacy and compliance losses, which depend on your architecture, your client mix, and your controls.
What’s left at the per-hour level is what I call the value capture rate—\(V_{cap}\): the real value of a saved hour (\(\pi\)) divided by the value the vendor assumed (\(R_{naive}\)):
\[ V_{cap} = \frac{\pi}{R_{naive}} \]It tells you how much survives your billing mix and redeployment capacity, before adoption lag, deployment cost, and privacy risk are priced in. Those come out in the risk-adjusted NPV.
For the most favorable profile I modeled—a 500-attorney Am Law 100 firm with a deep matter backlog, 75% redeployment probability, supportive leadership, and above-average technology infrastructure—\(V_{cap}\) is roughly 10%. That is, the real per-hour value is roughly one dime for every dollar implied by a simple hours-saved-times-billing calculation. Although the firm still makes money, the naive calculation was off by roughly a factor of ten. For a small litigation boutique billing almost entirely hourly with a thin demand pipeline, \(V_{cap}\) is negative.
Both of these figures are outputs under specific assumptions I chose for the analysis, not universal laws. When you change the inputs, the outputs change, which is the point of this framework. Vendor-supplied ROI projections are meaningless without firm-specific calibration.
This framework discourages what I call the FOMO-first approach to AI deployment. Rather, it’s a “know before you go” framework.
For the most favorable firm profile modeled, only 10 cents of every dollar in a naive saved-hour valuation survives contact with the firm’s actual economics. For an unfavorable profile, the number is negative.
Five diagnostic questions before you sign another AI contract
-
What fraction of your AI-eligible work is billed hourly? If more than 90% of your AI-eligible work is billed hourly, every saved hour creates a substantial billing headwind. Whether that headwind becomes a net loss depends heavily on whether freed capacity can be redeployed.
-
The last time an attorney became unexpectedly available, what fraction of their freed capacity moved to additional revenue-generating work, and how quickly? This is a practical indicator of your redeployment probability (\(V_d\)), and you shouldn’t assume it away. Start measuring it by tracking what happens when attorneys finish work early. Where does freed capacity actually go? If the answer is “they spend time catching up on admin,” your redeployment capacity is probably weak, and no AI tool will fix that. Understanding your \(V_d\) will tell you more about your AI ROI potential than any vendor benchmark would.
-
Did your technology team and your compliance team sign off on the same configuration, using the same numbers? If they evaluated it separately, they’ve probably not priced the trade-off between cost and exposure.
-
For your most sensitive client, can you name the control that lowers the odds of a disclosure, the control that limits the damage, and what your policy actually pays after the deductible? If you can’t, you are flying blind on the risk side.
-
If your vendor’s ROI projection were cut to a tenth, would you still sign the contract? For the most favorable firm I modeled, that’s roughly what survives at the per-hour level.
If you can’t answer two or more of these, you’re not ready to deploy. You’re ready to pilot. Run the \(\pi\) formula under your firm’s parameters; if it comes back negative, that’s not necessarily a reason to avoid AI, but a reason to fix your billing structure or staffing processes before you deploy.
The stakes
The gap between evidence-based deployment and vendor- or FOMO-driven deployment isn’t a minor calibration difference. My modeling puts it at millions of dollars over a five-year horizon, with the variance driven primarily by parameters (billing mix, redeployment capacity, and realization) that generic ROI calculators don’t model at all.
If the opening hypothetical happened to capture only 6.5% of its naive saved-hour valuation—the rate produced by my modeled mid-size profile—the $4 million headline would fall to roughly $260,000 before license fees. At the boutique’s rate, they’re at a loss.
There are two versions of the next five years. In one, your firm deploys AI calibrated to its actual economics—its billing structure, its demand pipeline, its client sensitivity, its risk tolerance—and captures real, measurable value. In the other, your firm deploys based on competitive anxiety and vendor projections, chases phantom savings that never reach the P&L, and discovers the gap between the pitch and reality only after the money has been spent.
The framework tells you which path you’re on. The question is whether you run the numbers before you sign another contract.
The full quantitative framework, including the detailed illustrative analysis and sensitivity results, is in my working paper, Modeling AI Adoption in Law Firms: An Integrated Quantitative Framework (SSRN).