Vas (Varick Agents): "Where to Begin With AI Transformation"
A playbook for where a large company should start with AI agents, built around one anonymized client case study. Ray's verdict: skim, file. The method is sound and concrete. The proof is one vendor's own client, so treat the numbers as an illustration, not evidence.
Why this is in the vault
This is the second Varick playbook in the vault (the first is [[2026-05-26-varick-how-to-transform-a-company-with-ai]]). The new piece adds a specific deployment sequence: pick a department, map the as-is process from event logs, delete steps, then run agents in shadow mode before they act. That sequence is the clearest public description so far of what "agents in production" looks like inside a finance function. The founder's credibility-for-phData thread is built on that exact domain.
Key claims (paraphrased)
Start with a department, not a workflow. Most of the elapsed time in a process such as procure-to-pay sits in the handoffs between departments (procurement, receiving, payroll), not inside the one team you optimize. Each department meets its own service-level agreement (SLA) while the end-to-end process still takes three weeks. Vas calls these upstream and downstream dependencies. He concludes that "department-specific agents" do not exist: a finance agent has to read from and write to procurement and payroll.
Choose the department by volume, cost, and reach. Volume means how often its processes run each month. Cost means what the department costs (the CFO usually knows). Reach means how many other departments its processes touch. In his experience the usual order is finance, then sales ops, then HR, procurement, or ops, with legal and marketing last. He says about 40% of Varick projects land in the CFO's office and 25% in sales ops (self-reported).
Three starting states.
- Not started. Usually blocked by security, legal, or an enterprise resource planning (ERP) system migration. He argues that "get our data in order first" is part of the transformation, not a prerequisite, and that a late starter at least carries no unused seats and no gatekeeping "center of excellence."
- Copilot plateau (the most common). Usage spikes and then flattens, and the gains never reach the profit and loss statement (P&L). A copilot speeds a person up 10-20%, but only when that person thinks to prompt it. His comparison: Copilot that saves 1,200 of 12,000 staff 15% of a third of their day changes no cost line, while automating 75% of a 40-person accounts payable (AP) team removes about 30 roles' worth of cost.
- Stuck pilots (the worst). The pilots were built inside one team, with no process map and no metric set in advance. They worked on the happy path and failed on the first exception. The result is company-wide fatigue.
The 7 steps (illustrated with a global distributor that had hit the copilot plateau, had 3 stuck pilots, and a CEO who needed board results in 16 weeks):
- Score the departments and pick the top one. Finance won: a 20-day close, 700 AP exceptions a month, and 4 departments touched by the close.
- Agree on the value and write it down. The CEO, CFO, and controller set key performance indicators (KPIs) for time/cost, revenue net of token spend, and risk. The baselines included a true cost of $31 per invoice, against the $5 that staff assumed.
- Map the as-is process. The documentation said 11 steps. The ERP event logs plus interviews showed 20 steps in the ERP and 4 more in email. Half the extra steps came from intercompany invoices sent without a purchase order. He says this discovery step is where most partners fail.
- Design the to-be process, and delete steps before you automate. Of the 20 steps, 7 were deleted, 4 went to plain code, 6 went to agents, and 3 stayed with a human because the auditors require it. The most common mistake is to automate the as-is process unchanged.
- Build on existing systems. The "agents" are mostly deterministic code. A model runs only where the input is unpredictable, such as a new invoice layout or general-ledger coding learned from the previous 1,000 invoices. What the agents learn goes into a shared "company brain," so the second department starts with that context already in place.
- Shadow mode, then go live. For 2 weeks the agents run at full volume but take no action, and every case where they disagree with a human gets reviewed. From week 3 they act on the cases where all checks are coded and confidence is 99%+ (normally about 50% of cases). By week 12 they handle about 90%, and people keep the rest.
- Find the power users (5-10% of staff). Give them a day of training and run a leaderboard on time saved with their skill files, not on tokens used. Those people then build the agents for the next department.
Reported outcome: after 4 months the close takes 8 days instead of 20, AP exceptions are below 50 a month (from 700), and a person does 3 of the 20 close steps. Varick quotes about 4 weeks per agent from build to production.
Why more companies don't start this way, in his view: seats are easy to sell and easy to show a board, consultancies run use-case scoring workshops without knowing what to ask, and the labs now sell embedded engineers. He says a client fired Deloitte after 3 months over poor discovery (an unverifiable anecdote).
Cited stats: secondhand, unverified
Ray did not check any of these against the primary source. Do not reuse them without a check. Source links as given in the article: McKinsey State of AI, SailPoint press release, OpenAI deployment company, Microsoft Frontier Company, CIO.com.
- McKinsey, 2026 State of AI: 80% of AI users say it made them more productive, but only 37% of respondents report any earnings impact. Almost 75% of companies making money with AI fundamentally changed their processes, against about 25% of the others.
- SailPoint (as of Oct 2026): 79% of organizations have AI agents in production, and only 2% use identity-security tools built for agents.
- OpenAI deployment company: launched May 2026 with $4B from 19 partners.
- Microsoft Frontier Company: launched July 2026 with $2.5B and about 6,000 embedded engineers and industry experts.
- CIO.com, 2026-10-13: a quote attributed to the chief information officer (CIO) of Vistance Networks: "We are spending millions sending AI experts to learn business. We should be sending business operators to learn AI." Vas calls it half right, because operators alone cannot run shadow mode or pass a Sarbanes-Oxley (SOX) audit.
Bias flags
- Vendor content marketing. Varick sells this exact engagement, and the piece ends with "book a call." The argument against seats, the big consultancies, and the lab deployment arms is also an argument against Varick's competitors.
- n=1 and self-reported. The case study is one anonymized client of his own firm, with no independent check of the 20-to-8-day close or the 700-to-50 exception figures.
- "Fewer than 1,000 people in the world can do this correctly" is a scarcity claim made for marketing, not a measurement.
- The 60-90% efficiency range for background agents (versus 10-20% for copilots) is asserted without a source.
Mapping against Ray Data Co
- phData main bet and credibility (
project_credibility_for_phdata_sales). The founder's chosen domain is agents in production, measured by pipeline. This piece gives him a vocabulary a buyer will recognize: department scoring, value one-pagers with baselines, as-is maps built from event logs, delete-before-automate, and shadow mode with a confidence threshold. phData's position on Snowflake is close to step 5 ("build on the systems you already have"). The ERP event logs that discovery depends on usually end up in the warehouse anyway. That makes "process mining on Snowflake, then shadow-mode agents" a plausible framing for him to test in his own words. It is Ray-selected and not yet tested with him. - "Agents are mostly code" matches the posture already in the vault: typed judgments go to a narrow model call (
feedback_jev_for_typed_set_problems), and everything else stays deterministic. - Shadow mode is a verification gate. It is the production-deployment version of Ray's own independent-verifier rule (
feedback_verification_independent_worker_pattern): the agent's answer is compared with a human's before it is trusted to act. - Ship 30 hook. The founder's Day 11 draft argues that agents can run a small business's back office ([[2026-10-19-week2-drafts]]). This article is the enterprise version of the same idea, and the clearest hooks are "20 steps became 3" and "delete steps before you automate." A future piece could reframe it in his voice rather than summarize it (per
feedback_no_derivative_sanity_check_pieces, the source is evidence, not the topic). This is a Ray suggestion and is not on his idea bank.
Related
- [[2026-05-26-varick-how-to-transform-a-company-with-ai]]: Varick's earlier playbook (4-trait workflow selection, self-improving agents). Read the two together to see the method develop from workflow-first to department-first.
- [[2026-10-19-data-engineering-weekly-292-zalando-rag-flywheel-netflix-ontology]]: a named-enterprise agents-in-production data point (Zalando, 85% live acceptance), a useful counterweight to this anonymized case.
- [[2026-10-20-commoncog-roger-williams-thinking-tools-org-transformation]]: Commoncog on organizational transformation (its issue also points to a members' thread on process mining as an academic field). It is a non-vendor lens on the same "redesign the process, not the tool" thesis.
- [[2026-10-19-week2-drafts]]: Ship 30 week 2, which includes the Day 11 back-office piece.