RESOURCES · EVIDENCE
Every number on this site, and where it comes from.
If a figure appears on this site, it is here with its source and date, so an owner, an advisor or a board can check it without asking us. Demo numbers are composite and say so on the demo; they are not in this table. When a source is updated, the row is updated and the change is noted at the bottom.
Updated 9 October 2026
95%
of enterprise generative-AI pilots showed no measurable effect on profit and loss; about 5% reached rapid revenue acceleration.
Where it appears: Home page, the three figures under the opening statement.
Source
MIT NANDA, The GenAI Divide: State of AI in Business 2025 (Aditya Challapally et al.), August 2025. 150 leader interviews, a 350-employee survey, 300 public deployments.
Report summary and PDF ↗6% · 88%
6% of companies are AI high performers, meaning AI contributes 5% or more of EBIT; 88% report regular AI use in at least one business function.
Where it appears: Home page, the three figures.
Source
McKinsey, The state of AI in 2025: Agents, innovation, and transformation, 5 November 2025. 1,993 respondents in 105 countries; self-reported.
mckinsey.com ↗39%
of Canadian small business owners say the business is too reliant on them for day-to-day operations.
Where it appears: Home page, the three figures.
Source
CFIB, Succession Tsunami: preparing for a decade of small business transitions, news release 10 January 2023.
cfib-fcei.ca ↗56%
The best model gets 55.8% of real financial-analyst tasks fully right (65.4% with partial credit). The May 2026 snapshot had the best model at about 52%; the figure moves as models are added.
Where it appears: Home FAQ, which models we use; the Way, why not let the model handle the numbers.
Source
Vals AI, Finance Agent v2 benchmark, snapshot 6 October 2026. Independent; Vals is a benchmarking company.
vals.ai ↗66%, none above 70%
Across 19 models on 101 task-oriented accounting questions (classification, journal entries, AP, AR, bank reconciliation, reporting, close), the best scored 66.0% and every model failed at least a third of its tasks.
Where it appears: Home FAQ, which models we use; the Way, rule 04 FAQ.
Source
DualEntry Labs, AI still fails one-third of real accounting tasks, updated 27 August 2026. DualEntry sells accounting software; read it as a vendor's test.
dualentry.com ↗Withdrawn
Microsoft retired the Excel COPILOT function on 14 September 2026 after a year in preview. Its launch guidance, as reported at the time, advised against using it for numerical calculations or for tasks with legal, regulatory or compliance implications; that wording is no longer on the current support page.
Where it appears: The Way, rule 04 FAQ.
Source
Microsoft Support, COPILOT function (retirement notice); Windows Central, 17 August 2026, for the launch guidance.
support.microsoft.com ↗~9 models · one month
In the eighteen months to mid-2026 one lab retired around nine models; in April 2026 it acknowledged that its own engineering errors had degraded a paid coding product for over a month.
Where it appears: Home FAQ, are we locked into one provider.
Source
Reported by Daniel Kornum, Varick Agents, 2026, citing the lab's public statements. We have not independently counted the retirements; the lock-in argument does not depend on the exact number.
varickagents.com ↗65% · 29%
Six months after starting fully agentic, 65% of one operator's workflow nodes ran as plain code and 29% of workflows made no model call at all.
Where it appears: The Way, others who say this.
Source
Tomasz Tunguz, Is AI doing less and less?, 25 February 2026. One practitioner's own stack, 14 workflows.
tomtunguz.com ↗20–30×
A hosted frontier-model agent mesh works, but at twenty to thirty times the cost and fragility the problem requires; narrow local models near the data do better on bounded tasks.
Where it appears: The Way, others who say this.
Source
David Linthicum, It's time to stop buying the consulting firms' version of agentic AI, October 2026. An opinion from thirty years of enterprise architecture, not a measurement.
linkedin.com ↗35.6% vs 18.0%
On the TravelPlanner benchmark (1,000 test trips), running the procedure as code with the model called only at fixed points passed 35.6% of trips against 18.0% for the best model-led method on the same model (Claude Sonnet 4), with 11 commonsense violations against 275 and about a quarter fewer steps. In the authors' controlled re-run with the same constraints given to every method, the margin narrowed but held (37.2 vs 24.5 on a 100-task satisfaction score).
Where it appears: The Way, others who say this.
Source
Libin Qiu, Yuhang Ye et al., Blueprint First, Model Second: A Framework for Deterministic LLM Workflow, arXiv 2508.02721 (v2). Academic preprint, not peer-reviewed; one benchmark, travel planning, not business workflows.
arxiv.org ↗30 min → 8 h a day
Overseeing a couple of agents took SaaStr about 30 minutes a day a year ago and eight hours a day in October 2026, because agents that make decisions return work that looks finished instead of failing loudly.
Where it appears: The Way, others who say this.
Source
Jason Lemkin, The latest SaaStr AI Agent Playbook for GTM is out, SaaStr, 12 October 2026. One operator describing his own company; sales and marketing agents, not finance or operations.
saastr.com ↗~100 days
A QuickBooks Online connection stays signed in only if it is used at least once every 100 days or so (each use renews it); since late 2025 every connection also has a five-year ceiling. A job that runs less often than that finds its connection expired.
Where it appears: Blog, how AI should connect to your accounting systems.
Source
Intuit Developer, Important changes to refresh token policy, 12 November 2025. Vendor documentation; Intuit can change it.
Intuit Developer blog ↗142 tools · half
The community-maintained QuickBooks connector we tested exposes 142 tools, about half of which change records; it only blocks changes when three settings are switched on.
Where it appears: Blog, how AI should connect to your accounting systems.
Source
Agent Forge Works, our own test against QuickBooks Online and Xero sandbox companies, October 2026. Our test, not an independent one; counted from the connector's published tool list at the time.
The post →~93%
of permission prompts in Claude Code, Anthropic's coding agent, were approved by users; the more approvals people saw, the less attention they paid to each.
Where it appears: Blog, your ERP isn't going away.
Source
Anthropic Engineering, How we contain Claude across products (Max McGuinness et al.), 25 May 2026. The vendor's telemetry on its own product; no sample size or period given.
anthropic.com ↗$240,000 a year
The top of one estimate SaaStr received for its agents to keep using one software system through its API the way they already did. The vendor is not named.
Where it appears: Blog, your ERP isn't going away.
Source
Jason Lemkin, We Got a $240,000 Estimate for Agent API Access. Our Agent Suggested a $5 Postgres Instance., SaaStr, 3 October 2026. One company's estimate, not a published price.
saastr.com ↗10–20 min a day
Marketo's API limits took SaaStr's agent offline after 10 or 20 minutes each day, so SaaStr moved off Marketo.
Where it appears: Blog, your ERP isn't going away.
Source
Jason Lemkin, We Got a $240,000 Estimate for Agent API Access. Our Agent Suggested a $5 Postgres Instance., SaaStr, 3 October 2026. One operator's account of one vendor.
saastr.com ↗C$500M · 2.25%
BDC's LIFT program: C$500 million, loans of C$25,000 to C$5 million, at 2.25% when the project uses a Canadian AI solution or firm; the AI track needs C$1 million in revenue and a Digital Plan built with a BDC advisor first.
Where it appears: BDC LIFT guide; home FAQ, can this be financed; acquirers page.
Source
BDC, LIFT program page. Terms are BDC's and change; confirm with BDC before committing.
bdc.ca ↗
How to read this table
- A survey figure is what respondents said, not what was measured. McKinsey's and CFIB's are surveys. MIT's combines interviews with a review of deployments.
- A benchmark figure is a snapshot. Vals re-runs its leaderboard as models are released; the row carries the snapshot date and will be updated.
- A vendor's test is still a test, but the vendor had a reason to run it. DualEntry's row says so.
- An opinion is labelled as one. Linthicum, Tunguz and Lemkin are practitioners describing their own experience, cited because they reached the same rules independently.
- A preprint is research that has not been through peer review. The Blueprint First row says so, and says which benchmark it measured.
- Nothing here comes from our own engagements yet. When a client agrees to be named, the measured before-and-after goes on this page with their sign-off.
Changes to this page
- 15 Oct 2026Added three rows for the post on systems of record: Anthropic’s 93% approval rate (the vendor’s own telemetry, labelled as such) and SaaStr’s $240,000 estimate and 10–20 minute API cut-off (one operator’s account).
- 14 Oct 2026Added two rows behind new voices on the Way: Blueprint First, Model Second (a preprint, labelled as one) and SaaStr’s oversight hours (one operator’s account).
- 14 Oct 2026Added two rows for the first blog post: the QuickBooks connection lifetime (Intuit's documentation) and the community connector's tool count (our own test, labelled as ours).
- 9 Oct 2026First version. Vals AI row uses the 6 October snapshot; the May figure is noted for anyone who saw it quoted elsewhere.
30 minutes · no deck · no cost. Bring the work that piles up.