The 7·point chasm.

AI‑first services firms print 20.5% EBITDA. Chat‑assisted firms print 13.3%.

That 7.2‑point gap is the largest operating‑model split in professional services in a decade — and it opened in a single survey wave. It isn't about who uses AI. 32% of their peers already do. It's about who rebuilt the workflow around agents and who bolted a chatbot onto a human one.

src · SPI Research "Professional Services AI Benchmark" · published 2026-10-20 (EIN Presswire, 17:01 UTC) · n = 400+ PS firms

Don't drink the 20.5% straight from the can.

• Definitional slippage. "AI‑first operating model" is self‑reported. Firms running one Cursor instance and a Notion AI seat are coding themselves into the hi‑margin bucket.
• Survivorship. The firms that bolted AI on, got no lift, and quietly closed are not in the survey. The real 13.3% zone is a graveyard with its lights still on.
• Inference reversal risk. ICONIQ: AI gross margins are still only 53%. If GPT‑6 Sol pricing doubles — or your agent loops — the EBITDA chart inverts. The chasm is also a trapdoor.
• Capex tail. SPI measures today's P&L. The AI‑first firms are also carrying unamortised rebuild cost that doesn't hit margin until 2027.

cross-ref · SPI 2026-10 · ICONIQ 2026-10-01 · Apollo/Slok 2026-08-10 · Sacra GPT-6 pricing 2026-10-12

The Copilot‑as‑Seat‑License trap.

Through 2025–26 the modal mid‑market services firm — brokerages, MSPs, boutique consultancies — did the same thing: bought Copilot or ChatGPT Enterprise at $30/user, rolled it out org‑wide, measured "productivity minutes" in a people‑ops dashboard, and declared the project done. Those are the 13.3% firms. The lift was individual, not architectural. Nothing in the P&L moved because nothing in the workflow moved.

Transferable lesson for the HLX / Hyro path: if you can't name the process boundary that an agent now owns end‑to‑end — doc intake, serviceability pre‑check, outbound reactivation, FBA re‑order — you aren't AI‑first. You're paying for seats.

cf · Service Leadership MSP 2026-10-20 · MIT NANDA 95% pilot zero‑P&L · Orgvue 2026-08: 55% of AI-driven cuts now regretted

The thing it's great at isn't the thing most firms are using it for.

heuristic · if removing the agent breaks a measurable business step, you're in the 20.5. if it only slows a person down, you're in the 13.3.

Which side of the chasm is HLX on, today?

0–1 yes · chat‑assisted · 13.3 zone  ·  2–3 yes · in transit  ·  4–5 yes · ai‑first · 20.5 zone

Mortgage pipeline, five stages. Pick the first rail.

Pick one rail. Ship it in 90 days with a verifier and a named human owner on the 6%. That one rail is the whole investment thesis for a brokerage in 2027: not "we use AI," but "our pipeline's doc‑intake unit cost dropped 71% and the broker now sees three times the files without touching intake."

cf · Alphabet AlphaProof narrow‑SI issue №21 · Ford's 6% / IBM's 6% issue №14

"Software company, or a services business pretending to be a product?"

The SPI chasm is the valuation chasm. Chat‑assisted = services multiples (6–10× EBITDA at 13.3% margin). AI‑first = software multiples (if margin + growth + architecture hold). Nothing else explains why "the AI of running FMCG" could be worth more than Hyro itself — if, and only if, the ops stack is agent‑rail by the time of the raise.

Diligence will not ask whether we use AI. It will ask: name one P&L line where turning the agent off would be visible in a week. If we can name three, we are a software company narrated as FMCG. If we can name zero, we are a brand with a ChatGPT subscription.

pre‑deck prep · map every HLX & Hyro process to a rail; mark 3 agent, 2 human, 1 verifier each

Carry these in. Don't carry anything else.

• Which one process will we make agent‑rail by 20 Jan 2027, and who owns the verifier?
• Which role did we redesign this quarter because of an agent — not just equip with one? If zero, we are 13.3%.
• If we had to show a US investor one chart proving we're software‑shaped not services‑shaped, which P&L line is it and does it exist yet?

You already use AI. 32% of your competitors do too. The 7.2 points are not in the usage. They're in the architecture. The firms that move this quarter will spend 2027 harvesting them; the firms that wait will spend 2027 explaining why their margins look like 2022.