Make every AI your team uses understand your
business.

Every assumption in this calculator, and where it comes from

The headline counts only what an organization pays today and what ACE removes of it, and it is anchored to the AI bill — yours if you set it, estimated at $2,000 per active AI user if not. It is the re-sent context ACE removes from that bill (about a quarter of it), plus the people-side costs — the hours spent re-explaining the business to each AI tool, the re-asks that are about missing context, and the cleanup of fabricated facts — tapered as the bill grows: where the bill is large, tooling, caching and process already absorb some of that friction (assumption). The alternative world — rereading the whole corpus on every request — is shown as the counterfactual and never added in: a firm either pays that in tokens and waiting, or pays the human costs. Published research is applied to the organization you described; the “with ACE” figures apply the same constants at a 4,800-token request — arithmetic, not measured deployment rates. Sources are dated; where a number is ours, it says “assumption”.

How the headline is built

Off the AI bill

$168K of $700K

Bill estimated from your inputs: $2,000 per active AI user × 350 users × 1.0 for 3 systems (assumption). 30% of a typical bill pays full price for re-sent context — system prompts, pasted documents, re-fetched retrieval, replayed history; cached repeats already bill at 0.1×, so the full-price share is not the whole bill (assumption). ACE assembles that context at 4.8K tokens against ~24K re-sent — an 80% reduction → ≈ 24% of the bill. The re-runs removed are not counted here

People-side taper

× 0.07

The people-side figures below are the full measured costs; the headline claims × 1 ÷ (1 + bill ÷ $50K) of them — full strength under a ≈ $15K bill, half at $50K, a tenth at $450K. The larger the bill, the more of this friction tooling, caching and process already absorb — and a claim is judged against the bill it sits beside. Assumption

Re-explaining the business

4.8K h → $595K

350 AI users × 1.3 h a week × 50 weeks, at $125/h loaded. The 4.1 h of weekly AI overhead (15K h in total) is not claimed beyond this part

Answers asked again

60K → 30K removed → $310K

37% of requests run again; ACE removes the 50% treated as context-attributable, at 5 minutes of the asker's time each (assumption). Priced as tokens instead, the re-runs would be ≈ $302K — the counterfactual, not counted

Cleanup of fabricated facts

$232K → $170K recoverable

$186 per employee a month (HBR / BetterUp) × 500 people; with ACE the tax scales with the per-answer rate (3.6% at 4.8K tokens vs 14% at ~24K, the typical request today). 30K answers per year would carry a fabricated fact without ACE, 6.9K with; the odds a given person meets one today are 83% without and 32% with

The counterfactual — reread everything

223B tokens → $1.1M

644 requests a day × (1 + 37% re-runs) × ~1.0M input tokens + ~800 output, at $5 in / $25 out per million; 78 s of prefill per answer, 4.8K h per year in total; with ACE 1.1B tokens → $8.4K. Reading the whole corpus once for 500 people would cost ≈ $85K. Shown, never summed

What one person’s context spans

Your organization, as set

People

500

Slider, exponential scale 1–100,000

Kind of business

Law firm

Fee-earners and staff preparing matters, calls and filings.

Share who ask AI business questions

70% → 350 people

Gallup (22,368 employed U.S. adults, 30 Oct–14 Nov 2025): 66% of employees in remote-capable roles use AI at work, 40% frequently; by Q2 2026, 52% of all U.S. workers. We take a lower share for requests that need the organization’s context, whole people, minimum one · source

Full-context requests per AI user per day

4

Derived from the 31 minutes a day knowledge workers lose switching between tools and re-entering context (Howdy, Apr 2026) at ~5 minutes of pasting per full-context request — about six a day at most; 2–5 by kind · source

AI systems in use

3 → ×2.2 load

Salesforce/YouGov (517 employed adults, 4–5 Dec 2025): workers who use AI inside and outside their work systems toggle between an average of four AI options. Each additional system adds 60% of a first system’s requests (assumption) · source

Scale discount

× 0.21

The per-head figure holds to 10 people, then thins × (people/10)^−0.4, to a floor of 5%, reached at about 18,000 people. So 10 people count 100% of the per-head figure, 100 people 40%, 500 people 21%, 3,300 people 10%, 10,000 people 6%, and 100,000 people 5%. Assumption. The reasons it thins: a larger organization has more staff outside the context-heavy core, more shared context, and a lower share of frequent AI users — Gallup finds 69% of leaders and 40% of individual contributors use AI at work; frequent use is 40% in remote-capable roles and 17% elsewhere · source

Annual AI spend

Auto — estimated $700K

Follows your inputs: $2,000 per active AI user a year (assumption: Copilot-class seats run $360–720, plus usage and AI tools) × 350 users, scaled × 1.0 for 3 systems — more systems, more seats. Drag the slider to your own figure; where it implies fewer active users than the headcount model, the people-side figures scale down to match

Loaded cost per hour

$125

lawyers’ mean wage $89.35/h (BLS OEWS, May 2025) × benefits. Benefits load 1.42× — wages are 70.3% of total compensation (BLS ECEC, March 2025, released 13 Jun 2025); BLS OEWS wages are the May 2025 estimates · source

Per request

Input tokens reread

1.0M · 4.8K with ACE

The homepage context race, one scroll up — ~120 documents × 8,420 tokens reread without Percidian; one assembled brief with ACE

A typical request today

~24K tokens

Assumption: partial context — a system prompt, pasted excerpts, re-fetched retrieval, replayed history; RAG contexts commonly run 4K–64K tokens. Today’s fabrication rate anchors here, not at the 1M-token reread of the counterfactual — that is a world nobody actually runs

Output tokens per answer

800

Assumption: ~600 words, a brief. Enterprise chat answers run 200–500 tokens; models rarely write over 1,000 words (Epoch AI, LUQ) · source

Price per million tokens

$5 in · $25 out

Claude Opus 5 base rates (Anthropic pricing page, read Aug 2026). Sonnet 5 is $2 / $10; cache hits 0.1× · source

Prefill speed

13K tokens/s

1M tokens in 77 s on a well-provisioned Llama 3 405B cluster (Introl, 2025); a single A100 prefills 100K in ~21.7 s (TokenSelect, Nov 2024) — the wait shown is a floor · source

Words → tokens

1.33

Anthropic: 1 token ≈ 0.75 words. Claude 4.7+ tokenizers produce ~30% more; we keep the lower figure · source

How people work with AI today

Re-runs

37% · 19% with ACE

Workday, ‘Beyond Productivity: Measuring the Real Value of AI’ (3,200 leaders and employees, fielded by Hanover Research, Nov 2025; released 14 Jan 2026): 37% of the time AI saves is lost to correcting, clarifying or rewriting its output — about 4 hours per 10 gained. We apply it as the share of requests run again, and with ACE remove only the half we treat as context-attributable (assumption; Salesforce/YouGov: 76% say their AI tools lack company data or work context, 62% expect access to cut time spent searching) · source

Overhead hours per AI user per week

4.1

Reworking AI output ~1.5 h (Workday: 85% save 1–7 h a week; 37% of ~4 h lost to rework) plus 2.6 h switching between tools (Howdy, Apr 2026: 31 minutes a day). Only half of the switching time — re-explaining the business to the next tool — is claimed as recovered (assumption) · source

Users who treat AI output as a starting point

86%

Microsoft Work Trend Index 2026 (20,000 knowledge workers, Feb–Apr 2026): 86% treat AI output as a starting point, not a final answer — a share of users, not of outputs, so it is not used as a multiplier · source

Cost per hallucination that lands on a colleague

1 h 56 min → 241

HBR, ‘AI-Generated Workslop Is Destroying Productivity’, BetterUp Labs & Stanford Social Media Lab, 22 Sep 2025 (1,150 US workers): each piece costs the recipient 1 h 56 min; priced here at this kind’s loaded rate. Shown per incident, not multiplied by the count · source

Cleanup tax per employee

$186 a month

Same HBR / BetterUp study, from respondents’ own salaries — the figure the tile multiplies by your people; with ACE it scales with the per-answer hallucination rate, p at 4.8K assembled tokens against p at ~24K — the typical request today, not the 1M-token counterfactual. The tail is far worse: Deloitte refunded the final instalment of an A$440,000 report over fabricated citations (6 Oct 2025); a US federal court sanctioned two firms $31,100 for AI-invented authorities (Lacey v. State Farm, 6 May 2025); 1,668 court cases involving hallucinated citations by 2 Jul 2026 (Charlotin database) · source

Hallucination curve — odds per answer, by input length

Short, grounded input

< 2%

Vectara HHEM hallucination leaderboard, 2025: top models under 2% on grounded summarisation · source

Legal RAG tools

17–33%

Stanford RegLab / HAI, ‘Hallucination-Free?’, 30 May 2024: Lexis+ AI >17%, Westlaw AI-Assisted Research ~33% of queries; general chatbots 58–80% on legal questions · source

At 32K tokens

−30%

NoLiMa (Adobe Research, ICML 2025; arXiv Feb 2025): 11 of 12 models below half their short-context score at 32K; GPT-4o 99.3% → 69.7% · source

Degradation with length

all 18 models

Chroma ‘Context Rot’, 14 Jul 2025: every frontier model degrades as input grows; ~300 focused tokens beat the same question in ~113K. Liu et al., ‘Lost in the Middle’ (TACL, 2024): >30% loss when the answer sits mid-context · source

The curve

3.6% at 4.8K · 17% at 32K · 30% at 1M

p = 1% + 29% × (1 − e^(−(tokens − 1,000)/40,000)), a fit to the anchors above, capped at 30% because nothing measures fabrication at 1M tokens; Chroma says the true figure keeps rising

One person’s context — the sources on the bar

Emails a day

117

Microsoft Work Trend Index special report ‘Breaking down the infinite workday’, 17 Jun 2025 — trillions of M365 signals, 31,000 workers in 31 markets. ~110 words each: Boomerang’s 40M-email study puts business emails at 50–125 words · source

Chat messages a day

153

Same Microsoft report (Teams). Slack: 92 messages sent per person per day. ~20 words each is our assumption · source

Meeting hours a week

8

Microsoft WTI ‘Will AI Fix Work?’, 9 May 2023: heaviest quartile 7.5 h a week in Teams; meetings ×3 since Feb 2020. Speech at ~150 words a minute (NCVS) · source

Files reachable per employee

6M+

Varonis data risk reports (2019–2021): new hires can access 6M+ files on day one; 12M+ at firms over 1,500 staff; 10M in financial services. We count only ~300 documents touched a quarter · source

Clinical notes

5,002 chars · 359 per patient

Epic Research, 2024: 5,002 characters per note in 2023 (+8% since 2020). JAMIA ‘Call me Dr Ishmael’, 2024: median ED patient arrives with 359 notes, 58,662 words · source

Data rooms

5,000–50,000+ pages

M&A Community, 2025; Bloomberg Law’s standard diligence list runs to 174 document types · source

Specialized systems by kind

see the bar

Counts per person per quarter are our assumptions and say so on hover; document sizes are the homepage’s 8,420-token document or the source named

Working year: 250 days, 48 weeks, 21 days a month. Horizon for one person’s context: 90 working days.