Your firm is being asked to change faster than it can afford to guess.

We can quickly build a digital twin of your firm and test the hard decisions on it first — so you can find the moves that improve the outcomes you care about most, whether that's profit, margin, retention, or succession, before the strategy is real.

Your pricing, your leverage, your compensation, your practice mix, your clients, your partners — a few adaptive rounds isolate the best lever combination and the causal story, and Monte Carlo simulation1 shows how reliable the answer is.

The Question — from the managing partner's chair

If I give the engine my firm's actual numbers — not an archetype, my firm — which levers should I pull, and in what order, to move profit-per-partner with the least political cost?

Step 1 · Input

Describe your firm in six numbers

You are the input. Not an archetype.
Six observable proxies — not abstract self-ratings — capture how you price, how steep your pyramid is, who owns the book, what you practice, how concentrated your clients are, how hard the partners will fight change.

Step 2 · Mechanism

Build the digital twin, then optimize it

Your numbers become a full simulation: twelve psychological profiles, a fifteen-step litigation workflow, every partner↔associate seam and its tacitness, the AI Profit Paradox.
Then instead of testing every lever combination blindly, the engine runs a few rounds, each informed by the last — try a lever, see what moved, use that to pick the next experiment.
That is what “adaptive” means here: learn as you go, not test everything — each round is chosen from the result of the last, so the search narrows toward what actually moves profit-per-partner. All of this runs on the fast deterministic engine; the winning combination is then replayed once on real LLM agents to see it in lifelike action.

Step 3 · Output

Get the best combination — and the causal story

Not “pull this.” The best lever combination, in order — plus why: which levers interact, so you know comp only helps after you've converted to AFA2, not just “comp helps” — and what each move costs you politically.

Adaptive picks · Monte Carlo measures

Two different jobs, working together.

Adaptive picks what to try. Instead of blindly testing all 32 lever combinations, the optimizer runs a few rounds — main effects → interactions → refinement — each one chosen from the result of the last. It narrows toward what actually moves profit.

Monte Carlo measures how much to trust it. The firm engine is stochastic — attrition, market shocks, political events — so every experiment is averaged across many seeds, and the final recommendation carries a confidence interval.

Adaptive finds the answer. Monte Carlo tells you how reliable it is.

The answer is found cheaply, then pressure-tested for real.
Two different engines, run one after the other — the reason you can trust the recommendation without paying to search with an LLM.

1 · Find the levers — deterministic mock
A fast, deterministic engine runs the hundreds of simulations the search needs: your exact firm, one lever changed at a time, averaged across seeds.

Same inputs always give the same result — reproducible, and effectively free (no AI calls).

This is the stage that decides which levers to pull, and in what order, for your firm's specific numbers.
2 · Pressure-test it — probabilistic LLM run
With the levers now chosen, one run on a real LLM (DeepSeek) replays that exact configuration with the twelve agents actually reasoning — partners and associates making real judgment calls, matter by matter.

This run is probabilistic: real, non-deterministic behavior, so you see whether the recommended plan holds up when lifelike agents actually operate it.
Why this order
Searching with the LLM would mean hundreds of runs — thousands of dollars and hours.

The deterministic mock does that search for free; the LLM run then pays once to validate the winning combination on your firm's exact specs.

The magnitudes shift between the two engines — the deterministic mock gives the shape of the answer, the LLM run gives lifelike behavior at that setting.
Pricing
Hourly, partial fixed-fee, or already value-based.
Leverage
How steep the associate-to-partner pyramid is.
Origination
One dominant rainmaker, or a distributed book.
Practice mix
Litigation vs. transactional share.
Client concentration
One whale client, or many.
Partner power
How hard the partners will fight any change.

PPP is the AmLaw scoreboard, but it isn't every firm's goal — and chasing it blindly can mean shrinking the partnership to raise the average. You choose the objective; the engine ranks the levers against that, and shows the trade-offs against the rest.

Profit-per-partner (PPP)
The classic scoreboard — the default when the question is partner economics.
Margin
Matter profitability — when the worry is doing the work at a loss.
Revenue per lawyer
Top-line productivity — for growth-stage or merging firms.
Retention & succession
Associate attrition and a sellable book — for an aging partnership.
A weighted blend
e.g. 60% PPP, 40% retention — optimize several goals at once, with constraints (“without attrition rising, without de-equitizing partners”).
Pricing
Hourly → fixed fee. The master lever when it's still on the table.
Leverage
Reshape the associate pyramid.
Comp
Who gets credit when AI does the work.
Seams
Write down the unwritten know-how. When a partner hands work to an associate, a lot of what they “just know” never gets said — so detail is lost, work is redone, quality slips. Turning that tacit knowledge into a checklist or template closes the gap.
Latency
How fast you observe and act.

This is what one firm’s configuration file actually looks like — every field, in plain English. You hand the engine these numbers; it hands you back the levers to pull.

Setting Value What it means
Pricing posturehourlyBills by the billable hour today (vs. partial or fully value-based / AFA).
Leverage ratio3.53.5 associates per partner — how steep the pyramid is.
Origination concentration0.4Fairly distributed book (0 = everyone originates, 1 = one rainmaker owns it all).
Practice mix (transactional)0.3Mostly litigation (0 = all litigation, 1 = all transactional).
Client concentration0.25Many clients, no single whale (0 = many, 1 = one client is most of revenue).
Partner power mix0.5Balanced governance (0 = cooperative, 1 = rainmakers hold veto power).
Baseline PPP$2.1MToday’s profit-per-partner — the starting point every lever moves from.
Baseline margin38%Today’s matter-profit margin.
Tech maturity0.4Some knowledge-management infra, not fully codified (0 = none, 1 = fully codified).

…and the levers the engine is allowed to try on that firm:

Pricinghourly → afa_nativeConvert the billable hour to a fixed-fee (AFA) model.
Compensation (comp)0.3 → 0.9How hard partners are paid to adopt AI (0 = not at all, 1 = fully rewarded).
Seamsfalse → trueCodify the tacit partner↔associate handoffs where work gets lost.
Latency2 → 1Sprints between observing a result and acting on it — lower is faster.
Leverage3.5 → 5.0Reshape the pyramid — more associates per partner.

Not “here are ten levers.” A written deliverable you can take into the partnership — your firm's numbers, the levers worth pulling and in what order, and the causal story behind them.

The firm we modeled
Your signature in four grouped tables — structural posture, market-facing work & comp, the books, and culture as observable proxies — every field tagged so you see what's anchored vs judgment.
Levers tested
The five levers, each with Δ PPP, Δ margin, and its own metric — plus the interactions: seams×pricing synergy, comp flips sign under AFA.
The recommendation
The best combination, the headline PPP, the Δ vs baseline with its ± confidence interval, and the causal story in one sentence.
Metric trajectories
Eight key metrics sprint-by-sprint — PPP, RPL, margin, realization, AI adoption, attrition, redline rework, utilization — the actual simulated path, not a snapshot.
Assumptions & caveats
The evidentiary-status legend, and the one honest caveat: the shape of the answer is defensible, the dollar magnitudes are archetype-calibrated until you feed real firm intake.
Raw data
The full trace — metrics, every LLM prompt and response, the state at each sprint — so nothing is a black box.
The headline it opens with
“Fix the seams first, then convert pricing to AFA. Don't start with comp — under your hourly billing, paying partners for AI raises adoption but costs profit.”

The adaptive optimizer's first round: pull each lever alone against the baseline and rank by Δ PPP. Later rounds take the top levers into a 2×2 factorial to find interactions, then refine toward the winning combination. The ± spread is the Monte Carlo confidence interval — what separates a reliable lever from noise.

Lever Δ PPP ± spread Read
Seams (codify tacit handoffs)+$872k±$162kthe partner-judgment wall
Pricing (hourly → AFA)+$567k±$39kflips AI from cost to benefit
Leverage (flatten pyramid)+$0±$16kmoves hours, not PPP
Latency (faster loop)−$23k±$16kmarginal
Comp (pay partners for AI)−$55k±$0raises use, hurts PPP under hourly

Feed a real firm's six numbers and the optimizer re-runs for your firm — that is the point. These numbers are the archetype proving the mechanism, not the answer.

The headline finding

Adoption is not the problem — pricing and comp are. The AI Profit Paradox: on the billable hour, AI is a net loss (fewer hours, less revenue). On fixed fees, the same AI is a net win. Same tool, same people, same adoption.

The whole point

The right lever depends on your firm. The billable-hour trap is the most common problem — but it is not universal. What works for one firm is wasted effort for another.

  1. Monte Carlo simulation. A method that estimates an uncertain outcome by running the model many times with different random draws (here, different seeds for attrition, market shocks, and political events) and reading the distribution of results — the average and the spread — rather than trusting any single run. The spread becomes the confidence interval on the recommendation. Named for the Monte Carlo casino, the technique was introduced by N. Metropolis & S. Ulam, “The Monte Carlo Method,” Journal of the American Statistical Association 44:247 (1949), 335–341.
  2. AFA — alternative fee arrangement. Any way of pricing legal work other than the billable hour: a flat or fixed fee for a defined scope, a capped fee, a subscription, or a success/contingency component. It shifts the firm's incentive from hours logged to work delivered — which is why AI (fewer hours, same result) becomes a win rather than a loss under it. AFAs were pushed into the mainstream by the Association of Corporate Counsel’s Value Challenge (2008); adoption is tracked in Altman Weil’s annual Law Firms in Transition survey.

your firm in · adaptive rounds, each smarter than the last · a causal story out