The three-year reference architecture for an agentic professional services firm.

This is the closing piece of the SALT insight series. The previous twenty-three pieces argued positions — Constrained Agency as the operating discipline, the auth layer as the moat, the 75% rule on operating-model spend, the Microsoft estate as the default stack, the manufacturing 2030 causal chain, the hybrid-by-regulation healthcare ceiling, the billable hour as a deflating commodity, agents as systems rather than teammates. This piece pulls those threads into a single working blueprint for one operator class — the professional services firm in the 100–2,000 person band — across a three-year horizon. The architecture is opinionated, dated, and explicitly falsifiable.

The thesis is one sentence. An agentic professional services firm is built in three sequenced years: Year 1 lays the platform foundation, Year 2 builds the constraint stack, Year 3 migrates the contracting model and captures the margin lift. Firms that compress the sequence fail; firms that follow it compound. The reference architecture below is what each year actually contains, what milestones the firm should hit, what failure modes to guard against, and what falsifiable indicators grade the result.

Before walking the three-year sequence, name the seven layers. Every agentic firm runs on these, whether the firm has built them deliberately or accumulated them by accident. The reference architecture builds them in deliberate order.

Year 1 is the foundation year. The firm picks the substrate, the partner, and the first three pilots. The temptation is to skip Year 1 and start with the high-margin work in Year 3; firms that try this consistently fail because the substrate gaps surface six months in and force a rework. Year 1 is unglamorous, predictable, and load-bearing.

Year 2 is where most firms lose the plot. The Year 1 pilots produced wins, the partner team is excited, and the temptation is to scale aggressively and skip the unsexy infrastructure work. Skipping the constraint stack is the dominant Year 2 failure mode. Firms that scale before they build the auth, governance, and talent layers produce a wider footprint of agents that all share the same brittle foundation. When a single agent misbehaves, the blast radius is the entire footprint.

Year 3 is where the strategic payoff arrives. The substrate is laid, the constraint stack is built, the talent is in place. The firm migrates the contracting model from hourly to outcome-based on the service lines where the constraint stack supports it, captures the margin lift the agentic decade offers to outcome-priced firms, and consolidates the firm’s competitive position for the next decade. Firms that arrive at Year 3 having built Years 1–2 properly compound. Firms that arrive having skipped layers find the contracting model migration produces revenue volatility and partner churn.

The strongest counter is that three years is too long — the agentic landscape will shift materially over the architecture’s horizon, and a firm that commits today to a 36-month plan locks in choices that will be wrong by Month 24. This is partially right. The architecture is not a 36-month commitment to specific products; it is a 36-month commitment to a sequence — substrate, constraint stack, contracting model — that is structurally robust to product churn underneath. If Microsoft Entra Agent ID is replaced by something better in 2027, the firm migrates the auth layer without redesigning the architecture. If the dominant agent runtime changes, L2 swaps without affecting L3–L7. The architecture is opinionated about sequence, not about specific vendors. The vendor decisions can be revisited as the market evolves; the sequence cannot be skipped.

A second counter is that three years is too short — the operating-model layer requires deeper change than 36 months allows for the typical firm’s political dynamics. This is right for some firms. For those firms, the architecture is a five-year program rather than a three-year program, with Year 3 stretched into Years 3–5. The sequence still holds; the calendar accommodates the firm’s actual change capacity.

01 · Place yourself on the architecture today. Walk the seven layers. Mark which are built, which are partial, which don’t exist. Most firms in the band will find L1 partial, L2 piloting, L3–L7 absent. That is the honest baseline. Stop comparing to vendor reference architectures; compare to the seven layers above.

02 · Commit to the three-year sequence explicitly, with the partner group, in writing. Year 1 is the substrate. Year 2 is the constraint stack. Year 3 is the contracting model. The commitment matters because compression is the dominant failure mode, and partner-level discipline is what prevents compression.

03 · Hire the first constraint engineer in the next 90 days. Whatever the partner group decides about platform, MSP, pilots — none of it works without the constraint engineer. Make this hire before any agentic deployment goes live, not after.

This is the twenty-fourth piece of the SALT insight series. Across the four pillars — Foundation, Manufacturing, Healthcare, Professional Services — the throughline has been the same: Autonomous Business is the destination, Constrained Agency is the operating discipline that gets you there, the Microsoft estate is the default substrate, and the operating-model layer is where the actual work compounds. The reference architecture in this piece is the working version of the thesis for the firm class SALT serves directly.

The shelf is now twenty-four pieces. Each carries a falsifiable position. Each will be graded against subsequent reality on the standing position-review rhythm. Where positions falsify, SALT publishes the correction explicitly. The intellectual estate is built; the work of operating against it begins.

Now go build.