Playbook

The 6-Step Build

How AI-First Companies Actually Get There.

May 2026 · Synaptrix

Read

Executive summary · ~30 sec read

Six technical steps and a parallel People Track — drawn from how JPMorgan, Goldman, Klarna, Mercado Libre, Booking.com, and Meta actually got AI to production. The build sequence is universal: audit fragmentation → build the substrate → deploy internally first → pick a wedge workflow → embed engineers → re-architect the operating model. For a 50–2,000 person company, full transformation runs ~$3M–$10M over 18–24 months, with first measurable P&L impact in 6–9 months. The People Track runs in parallel from week 1. The McKinsey 70% failure rate is what skipping it looks like.

Contents

Summary
Opener
Step 1 — Audit
Step 2 — Substrate
Step 3 — Internal Agent
Step 4 — Wedge
Step 5 — Embed
Step 6 — Operating Model
People Track
Failure Modes
Summary Table
Closer

Opener

The 6-Step Build

The thesis is in the manifesto: software won AI first because software has a repo. The rest of the company has nothing comparable, and the unlock isn’t smarter models — it’s building the missing memory. The Company Brain.

This is the practical playbook: six technical steps drawn from how the live AI cohort (JPMorgan, Goldman, Klarna, Mercado Libre, Booking.com, Meta) actually got to production, a parallel People Track running alongside all six, and a closing failure-modes section.

The cost and time ranges below are industry-pattern estimates from comparable transformation projects (cloud migration, ERP, prior AI deployments), not figures the cited companies have disclosed. Directional, not diligence-grade. Validate against the specific company before underwriting.

Each step is specified along eight dimensions: What it is, Why it comes here, What success looks like, What fails it, Who runs it, Timeline, Cost (50–2,000-person mid-market), and Make vs Buy.

Six steps in sequence. People Track in parallel.

Step 01 — Audit Fragmentation

Audit Fragmentation

What it is: A catalog of every place company knowledge lives — Slack, email, Salesforce, Jira, Notion, meeting recordings, shared drives, and people’s heads — with explicit attention to decisions, contracts, customer histories, and operating procedures held only in tribal memory: the pricing exception approved in a DM, the renewal logic carried by one senior PM. The audit maps the substrate the business actually runs on, not the org chart of the systems it pays for.
Why it comes here: Cloud migration’s first phase was portfolio discovery. AWS’s “6 Rs” (Rehost, Replatform, Repurchase, Refactor, Retire, Retain) only worked because someone first cataloged what existed. A substrate cannot be built without knowing what feeds it; a wedge cannot be picked without knowing where its inputs live. AI projects that start with model selection instead of inventory repeat the cloud-era mistake.
What success looks like: A tagged inventory across four dimensions: systems, work-product types, decisions not captured anywhere, and per-source curation difficulty — concrete enough to plan integration sequencing and sign off on scope without a follow-up workshop.
What fails it: Skipping the audit because “we already know where everything is.” The companies that say this are reliably the ones whose pricing exceptions get approved in Slack threads no inventory contains. “We already know” is itself the strongest signal that the audit is required.
The “GIGO” question — do we need to clean our data first?: The most common objection: “Our Salesforce is a mess. Do we need 12 months of cleanup first?” No — the audit is part of the cleaning, and the Step 2 substrate is designed to tolerate messy sources by capturing provenance and flagging ambiguity. Klarna’s Kiki ingested Slack and email through curation, not pre-cleaning. Companies that wait until data is clean never ship.
Who runs it: An internal champion (chief of staff, COO direct report, VP Ops) at 0.25–0.5 FTE, paired with one or two external consultants who have done discovery work before. Not an IT-led project — the work is operational discovery, not data engineering.
Timeline: 4–8 weeks mid-market; 8–16 weeks at enterprise scale. The compressing factor is interview-cycle scheduling, not headcount.
Cost: $75K–$200K mid-market; $300K–$800K at enterprise scale where data-catalog tooling gets adopted.
Make vs Buy: Process, not product. A data catalog (Atlan, Collibra, Alation) helps where the engagement goes formal, but most of the work is workshops, interviews, and the tagged inventory. There is no “fragmentation auditor” SaaS product — the work is interpretive, not extractive. A vendor can crawl Salesforce; a vendor cannot tell you which exceptions get approved in Slack and never make it back into the record.

Step 02 — Build the Substrate

Build the Substrate

“Nearly 90% of crucial enterprise data is trapped in unstructured formats across silos, making it inaccessible to AI.” — Salesforce, 2025

What it is

The Company Brain itself. A graph, not a chatbot — a continuously updated, queryable, version-controlled record of decisions, entities (people, accounts, products, contracts), and operating rules, with permissions and audit trails alongside the data, not bolted on after. Klarna’s Kiki runs on Neo4j; JPMorgan’s LLM Suite has a custom data layer; Meta calls its version Second Brain. Different stacks, same shape: the system answers “how do we handle X?” by retrieving the company’s own record, not by asking a foundation model to guess.

Why it comes here

Klarna’s 2025 reversal is the cleanest cautionary tale: the customer-facing chatbot shipped before the substrate was load-bearing, quality slipped, humans were rehired, marketing claims walked back. Agents without a curated substrate fail predictably. Step 2 makes Steps 3–6 possible.

What success looks like

A queryable system where an agent — or a new hire on day three — asks “how do we handle a renewal blocker on a Tier 1 account?” and gets the company’s actual answer with citations pointing back to the Slack thread, the Salesforce note, the contract clause. A reviewer can verify; legal can audit; quality is testable.

What fails it

Treating it as enterprise search. Glean, Onyx, Notion AI, Microsoft Copilot retrieve; they don’t organize, reconcile contradictions, or decide what to remember. Salesforce 2025: “Nearly 90% of crucial enterprise data is trapped in unstructured formats across silos, making it inaccessible to AI.” Search pulls fragments; curation turns fragments into a record.

Security and permissions are not later problems

Centralizing institutional memory makes permissions the first engineering constraint. Six mechanisms PE diligence will check:

Permission-aware retrieval. Every query inherits source permissions (Slack ACLs, Salesforce profile, Drive sharing, Jira). Substrate propagates existing access; it does not grant new.
Source-level access controls. HR, legal, M&A, and board sources gated by default; inclusion requires named approvers.
Audit logs. Every query, retrieval, and answer logged. SOC 2/HIPAA-grade.
Model isolation. No source data trains public models — no-training contracts, behind-firewall inference, or both.
Answer citations. Every response cites sources. A wrong answer with a citation is recoverable; without one, invisible.
Regression evals. Automated suites catch quality drift from schema changes and vendor model updates before users do.

Who runs it

2–4 engineers (ML/data plus backend) and a product lead — rarely an IT team’s spare capacity. Three structures: internal hire-up (hardest; talent scarce, comp competes with frontier labs), embedded engineers (vendor forward-deploys inside the company — the Anthropic/Blackstone JV model, the Synaptrix model; fastest to production), and hybrid (consultancy months 0–6, internal alongside, vendor phases out by month 12–18 — most common in mid-market).

Timeline

12–24 weeks for v1. Internal pilots by month 4–6. Production-grade (evals, audit logs, access controls) by month 6–9.

Cost

$500K–$2M for v1 at mid-market scale; $50K–$200K/month ongoing for iteration and curation. Enterprise scales to $2M–$5M+ for v1. The ongoing line is not optional: uncurated substrates decay, and decay is the failure mode that ends programs in year two.

Make vs Buy

Hybrid, with an unforgiving boundary. Foundations are buy: knowledge graph (Neo4j, Memgraph), vector DB (Pinecone, Weaviate, pgvector), LLM gateway (Anthropic, OpenAI), orchestration (Temporal, Airflow). The substrate itself is build: the integration layer wiring Slack, Salesforce, Jira, Notion, and meeting transcripts into the graph, plus the nightly curation jobs deciding what to remember. Klarna built on Neo4j; JPMorgan, Goldman, and Mercado Libre built custom. Any vendor pitch claiming end-to-end coverage is selling search and calling it a brain.

Step 03 — Deploy the First Internal Agent

Deploy the First Internal Agent

What it is: One internal workflow — onboarding, support triage, contract review, sales-call prep — handed to an agent for employees only. The agent reads from the Step 2 substrate, drafts an answer or action, and sends it to a human in Slack, Teams, or a thin web UI. Narrow on purpose: one workflow, one team, one metric the owner already tracks.
Why it comes here: JPMorgan deployed LLM Suite to employees first. Goldman ran a 10,000-user internal pilot before any external launch. Klarna’s Kiki served staff first. Mercado Libre’s Verdi reached 17,000 internal developers before touching a customer case. Internal users tolerate roughness, generate corrections that feed the substrate, and set the quality bar.
What success looks like: 30%+ daily active usage in the target population within 60 days. Measurable time savings on the workflow — JPMorgan reports 3–6 hours per week per user. Corrections flow back to the substrate as training signal. The workflow owner, not engineering, reports the metric to the sponsor.
What fails it: Going customer-facing first. Klarna’s 2025 reversal: humans rehired, claims walked back, a year of credibility lost. The customer surface is the worst environment to discover the substrate isn’t yet load-bearing.
Who runs it: 1–2 engineers (often drawn from the substrate team early on) plus the workflow’s actual owner — a support lead, ops manager, or sales ops director. Whoever owns the metric owns the agent. Product-led, not IT-led.
Timeline: 6–10 weeks to first internal users; 12–16 weeks to 30%+ DAU. The compressing factor is workflow specificity: “draft the first reply on Tier 2 billing tickets” ships faster than “help with support.”
Cost: $150K–$400K for the first agent — engineering, LLM spend, and frontend where the agent isn’t in Slack or Teams. Inference at 2026 rates runs roughly $0.10–$0.50 per non-trivial query; budget $5K–$30K/month during rollout. An order of magnitude smaller than substrate build because the hard work was Step 2.
Make vs Buy: Foundation is buy: LLM APIs (Anthropic, OpenAI, Google), Slack/Teams bot SDKs, optional orchestration (LangChain, LlamaIndex). Workflow logic and substrate integration is build. Most successful teams write their own orchestration rather than adopt a heavy framework — lock-in hurts exactly when the agent matures and needs non-standard moves.

Step 04 — Pick One High-Volume Workflow as the Wedge

Pick One High-Volume Workflow as the Wedge

What it is: Once internal trust exists, one high-volume, high-document workflow gets handed to an agent that touches the business outcome, not just employee productivity. The public examples cluster in three: customer-service mediation (Mercado Libre, 10% of cases on a major site, $450M in mediation value); legal review (JPMorgan COiN, 360K hours saved annually); trip planning (Booking.com’s AI Trip Planner, shipped in 10 weeks). Different industries, same shape: narrow scope, instrumented metric, real money in the workflow.
Why it comes here: Step 3 earned the right. The substrate is load-bearing, employees have corrected the agent, and the curation loop has caught its first drift. The wedge turns infrastructure into P&L. Benchmarks: 10-week prototype (Booking.com), 8-week iteration cadence after launch (JPM). Counterexample: Hershey’s $112M big-bang ERP cutover during the Halloween peak. Wedge first, then expand — the cheapest insurance policy in the playbook.
What success looks like: Measurable P&L impact in one workflow within 90 days of launch. The agent handles 10–30% of volume at human-level quality — the operating number the workflow owner reports to finance, not a demo metric. Below 10% the wedge isn’t load-bearing; above 30%, scope has crept past one workflow.
What fails it: Trying to “transform the whole company” at once. McKinsey’s 70% transformation-failure rate is the cost of fanning out before a single workflow has been proven. Steering committees pick five workflows because picking one feels timid; each gets 20% of attention; none reach the quality bar; sponsor patience dies by month nine.

One workflow, all the way through, beats five workflows halfway.

Who runs it: 2–4 engineers (often the substrate team) plus the workflow owner plus the operations team whose metrics move. A PM joins as scope grows. Embedded engineers work best here — the wedge depends on sitting next to the mediation specialist, the contracts paralegal, the trip-planning agent. The detail that makes it ship is rarely in the documentation.
Timeline: 10-week prototype (Booking.com); 8-week iteration cadence after launch (JPM). First production volume by month 4–6 of the wedge phase. Steps 3 and 4 often run in parallel.
Cost: $300K–$700K for the prototype; $75K–$200K/month ongoing. Customer-facing wedges (mediation, contracts, customer comms) budget to the top of the range; internal-adjacent wedges (support triage with human send) to the bottom.
Make vs Buy: Mostly build, on top of the substrate. Buy: observability (Datadog, Honeycomb), evals (Braintrust, LangSmith), volume/quality analytics. Build: agent, workflow logic, routing, human-in-the-loop, system-of-record integration. A wedge built mostly out of vendor packages is a wedge whose competitive advantage transfers to the next buyer of those packages.

Step 05 — Embed Engineers, Don’t Deliver Decks

Embed Engineers, Don’t Deliver Decks

What it is: Forward-deployed engineers who build inside the company, not advisors who write recommendations. Hands on the keyboard, shipping production code against the customer’s systems, pushing commits to the same repo the substrate team owns and measured on the same metric the workflow owner reports to finance.
Why it comes here: Paul David documented in his 1990 paper that wide-scale electrification in the 1890s required a cadre of factory architects familiar with the new approach — built up by working out details on site. The Anthropic/Blackstone JV is the 2026 version: engineers placed inside portcos to do the building, not the briefing. McKinsey’s Rewired calls this the “talent bench”; BCG: 70% of AI value comes from people and process, not the model. Companies that build a Company Brain hire or rent the cadre; ones that buy slides are the 70% failures.
What success looks like: Engineers shipping production code inside the customer’s environment, connected to the substrate, observable in the agent’s daily output. Skills transferring over the engagement, so the artifact is a running system plus a team that can extend it. The metric is whether the substrate gained an integration, the agent shipped a workflow, or the wedge moved a P&L number this quarter — not consultant burn rate.
What fails it: Traditional consulting. Decks without code. A 60-page “AI strategy” six months in, with building deferred to a “Phase 2” that never gets funded. When the contract is structured around documents, the engagement optimizes for documents.
Who runs it: 3–8 engineers ongoing — vendor (Anthropic JV, Synaptrix), internal hires, or hybrid. Hybrid is the most common path: vendor engineers months 0–12 to seed capability, knowledge transfer beginning month 6, internal team running steady-state by month 18–24. Pure-internal-from-day-one is feasible but slower; the talent is scarce in 2026, comp competes with frontier labs, and the learning curve is steep when nobody on staff has shipped a substrate before. Hybrid wins on time-to-production for the same reason cloud migrations leaned on AWS Professional Services: the vendor has done it ten times before; the company is doing it once.
Timeline: Continuous, not bounded. Vendor presence 12–24 months for full transformation, capability transfer from month 6. After month 24 the vendor shifts to advisory or specific high-leverage projects, not daily build cadence.
Cost: $1.5M–$5M/year for 3–8 engineers, fully loaded — consulting markup if vendor, comp + benefits + tools if internal. Range is wide because team mix matters and regulated industries add a clearance premium.
Make vs Buy: This step IS the buy/hire decision. Most companies do hybrid: buy first to learn what the work looks like, then hire when the requirement is specific enough to write a job description that doesn’t rhyme with every other AI job description on the market.
The maintenance tail (year 3 and beyond): Once the embedded relationship phases out (months 18–24), the company needs a steady-state internal team. Year-3 team for a 500-person mid-market: 1 AI/product lead, 2 backend/data engineers, 1 agent engineer, 0.5 FTE CHRO for curation governance — $1.2M–$2M/year fully loaded. Enterprise (2,000+) is 2–3× that. Not optional: uncurated substrates decay. PE buyers underwriting this tail matters — a portco that booked the productivity gain without funding maintenance gives it back inside 18 months.

Step 06 — Re-architect the Operating Model Around the Brain

Re-architect the Operating Model Around the Brain

What it is: Once the substrate is live and agents are producing value, workflows are redesigned so the Company Brain is the system of record, not a side-tool. Walmart frames Sparky as the “primary vehicle for discovery, shopping and for managing everything from reorders to returns” — the primary surface, not a feature. Shopify’s Tobi Lütke memo: “Reflexive AI usage is now a baseline expectation.” If the brain is the primary surface and reflexive use is the baseline, the org chart, workflow map, and comp structure all need to follow. Step 6 is that work.
Why it comes here: Electrification only moved the productivity statistics when factories were redesigned around the unit drive — single-purpose motors at each machine — rather than retrofitted to existing line shafts. The lag ran for decades; the gating factor was operating-model redesign, not technology. McKinsey’s Rewired names the same step under “new operating model.” Skipping it produces the 2010s “we did digital transformation” companies whose org charts looked identical to their pre-cloud ones.
What success looks like: Step-change EBITDA, not productivity-tool deltas. Workflows where the brain is upstream of human work: the agent drafts the contract, the human reviews; the agent triages the ticket, the human escalates. Roles redefined and titled honestly. Comp aligned to AI-driven impact. The brain’s reporting line sits at the executive table (Chief AI Officer, Head of AI Operations), not buried two levels deep in IT.
What fails it: Stopping at Step 4. A wedge ships, the metric moves, the executive team declares victory, and the redesign that would compound the wedge into a step-change never gets scoped. Two years later, the EBITDA line looks like a company that built a good internal tool.
Who runs it: CEO + CHRO + COO + heads of function. Not delegable. McKinsey, BCG, and Bain can advise on change management and benchmark peer moves; decisions about roles, comp, headcount, and reporting lines must be owned by the leadership team that will live with them. Runs in parallel with Steps 4–5 — comp and OKR cycles do not wait for the engineering roadmap.
Timeline: 6–18 months mid-market; longer at enterprise scale where union agreements and multi-region HR practice extend the cycle. The compressing factor is leadership decisiveness, not consulting throughput.
Cost: Change-management consulting plus program management for a 500-person mid-market: $500K–$2M. The unpredictable line is restructuring itself — severance, comp adjustments, hires for new positions, training — which runs into the tens of millions if the reorganization is significant.
Make vs Buy: All internal leadership work. McKinsey, BCG, and Bain shorten the cycle on benchmarking and change-management cadence. Strategic decisions stay internal — a firm that makes role and comp decisions on the leadership team’s behalf sells a deliverable the team will not defend the first time it’s challenged.

§4.A — The People Track (parallel)

The People Track

Why this is its own track. The 6 steps describe what gets built; this track describes what changes for the humans while it gets built. McKinsey: 70% of transformations fail. Microsoft WTI 2026: “The Transformation Paradox is, at its core, a systems problem. And systems don’t fix themselves—they have to be redesigned.” A company can ship a perfect substrate and a working agent and still hit zero P&L impact if the people layer breaks. The seven moves below are the practical mechanisms.

Move 1 — Visible executive sponsorship: leaders model the behavior

CEO and leadership must be visibly using AI before asking anyone else to. Without this, every other move is theater.

Backing data: Microsoft WTI 2026 — when managers actively model AI use, employees report a 17-point lift in AI value, a 22-point lift in critical thinking, and a 30-point lift in trust. Tobi Lütke’s “Reflexive AI usage is now a baseline expectation” Shopify memo is the canonical version.

How to make it real:

CEO publishes a first-person memo naming AI as a baseline expectation. The authority comes from voice — don’t outsource the writing.
Leadership uses AI live in all-hands. Share the prompts. Show the failures. Modeling beats messaging.
Monthly “what I built with AI” segments from CEO, CFO, head of product, head of sales. Vulnerable, not polished.
Explicit AI-leadership role at the executive table (Chief AI Officer, Head of AI Operations) reporting to the CEO. Without it, the function ping-pongs between CTO, CIO, and CHRO and dies in the gap.

Owner: CEO. When: from week 1, ongoing.

Move 2 — Communicate the why before the what

Substrate work feels invisible next to a flashy customer chatbot — that’s why it gets cut. Communicate the substrate strategy directly, before any tool ships.

How to make it real:

Three-part message in plain language: why now (Anthropic-Blackstone, JPM and Goldman in production), why this (substrate vs. more pilots), why us (the company’s specific competitive position).
All-hands kickoff with extended Q&A — budget 30+ minutes for questions and don’t cut them off.
A living FAQ updated weekly with the actual questions employees are asking. The common ones tell leadership where the people layer is failing.
Cascade through managers, not videos. Every people manager runs a 30-minute team session in week 2.

Owner: CEO + Internal Comms + CHRO. When: week 1, then quarterly cadence.

Move 3 — Address “AI will take my job” directly

Some roles will change, some will be eliminated, some new ones created. “AI augments humans” platitudes nobody believes produce silent disengagement six months later.

How to make it real:

Public commitment with teeth: “We will not lay off anyone who engages with the new tools and reskills within the next 12 months.” Only credible if leadership means it.
Show the data. JPMorgan and Goldman both net-hired during their AI rollouts — productivity gains funded growth.
Reskilling investment with a specific dollar amount per employee, published — not buried in HR talking points.
Manager–direct-report 1:1s in week 4 specifically about how the role will change. Manager-led, not HR-led.

Owner: CEO + CHRO. When: at major rollout milestones, especially before Step 4.

Move 4 — Training, not town halls

Most employees default to “AI as Google search.” The shift to AI-as-collaborator — Microsoft’s four modes (delegate, collaborate, ask, explore) — happens by running real workflows with real prompts, repeatedly. Not in a 45-minute lecture.

How to make it real:

Phase 1 (weeks 4–8): every employee gets 4–6 hours of hands-on training on real workflows. Track completion.
Phase 2 (weeks 8–16): functional champions in each department hold weekly open office hours.
Phase 3 (ongoing): peer learning circles of 5–7 people share weekly what they’re automating. Cross-functional rotation.
Internal AI Slack/Teams channel staffed by the substrate team, 4-hour SLA. Anything slower kills momentum.

Owner: CHRO + L&D + functional leaders. When: alongside Step 3.

Move 5 — Realign comp, recognition, and promotion

People won’t redesign their work if they’re rewarded for the old model. Comp, promo, and recognition have to align with the new operating model or employees will rationally avoid the work.

Backing data: Microsoft WTI 2026 — 65% of AI users fear falling behind, 45% feel safer focusing on current goals than redesigning work with AI, and only 13% are rewarded for AI-driven reinvention.

How to make it real:

Add “redesigned X workflow with AI” as a measurable OKR starting the quarter Step 3 ships.
Promotion criteria include AI-driven impact for senior IC and manager promotions. Document it.
Transformation bonus pool tied to wedge-workflow metrics. Pay out at completion of Step 4, visible to everyone who contributed.
Update job descriptions across the company to include AI fluency — new hires and current employees both.

Owner: CHRO + Comp Committee. When: Step 4 onward.

Move 6 — Manage the disruption window honestly

Months 3–9 are messy. Pretending otherwise destroys trust — the gap between the official narrative and what’s happening at people’s desks becomes the story they tell each other.

How to make it real:

Published transformation roadmap with month-by-month milestones, updated monthly. Slipped milestones addressed openly.
Pre-announcement before chaos peaks: “Months 3–9 will be messy. Here’s what’s changing. Here’s the support.”
Weekly internal post from the substrate team: what shipped, what broke, what’s next. Authored by the lead engineer, not comms.
Public wins leaderboard: workflows transformed, hours saved, CSAT changes. Makes the cost of disruption legible by showing the upside.

Owner: COO + CHRO + Internal Comms. When: throughout, especially months 3–9.

Move 7 — Manage tribal knowledge loss in real time

Senior people who leave during a transformation take undocumented knowledge with them. Capture it before it walks out the door — and build the muscle so the company never depends on individual heads again.

How to make it real:

Restructured exit interviews: a separate operational-knowledge session (60–90 min) distinct from the HR exit. Recorded with consent. Outputs into the substrate.
Knowledge-buddy program: every senior IC paired with someone who shadows 4–6 hours/week, structured to capture decision logic, not task replication.
Monthly “tell me how you decide X” sessions between the substrate team and senior leaders. 60 minutes, structured, captured.
Senior-leader legacy docs: every director-and-above contributes a “judgment patterns” doc — when I face X, I think Y, because Z. Content stays private until departure.

Owner: CHRO + substrate team. When: continuous from Step 1.

Cost / time framing. The People Track adds roughly $300K–$1.5M/year in change-management consulting, internal program management, and training for a mid-market company, plus comp and role restructuring (scope-dependent). It deserves its own track because who runs it (CEO, CHRO, COO, Internal Comms) is different from who runs the technical build.

What fails the People Track. Treating it as IT change management. Quarterly emails instead of leadership modeling. Training without comp realignment. Hiding the disruption window. Outsourcing the whole thing to McKinsey instead of operating it from inside.

§4.B — Common Failure Modes

Common Failure Modes

Ten patterns we see in companies that try and fail. Run this list as a self-check at every quarterly review.

CEO delegates it to IT. The substrate becomes a data project, the People Track gets skipped, 18 months later nobody uses it.
The first workflow has no clear owner. Nobody’s accountable for quality or business impact; metrics never improve.
The company buys tools before mapping workflow pain. Copilot, Glean, Notion AI rollouts — all bought before anyone asked which workflow needs help.
The substrate becomes search, not curated memory. “Enterprise search powered by AI” wins over the harder curation work; six months in it’s a smarter Google and P&L doesn’t move.
Permissions are treated as a later problem. A real or perceived data leak in month 9 freezes the program — sometimes permanently.
No baseline metrics before automation. When the wedge ships, nobody can prove impact because nothing was measured before.
Employees trained but not rewarded for changing behavior. Microsoft WTI 2026: only 13% are rewarded for AI-driven reinvention.
The wedge workflow is too broad. “Customer service” instead of “tier-1 refund eligibility checks.” Scope expands, prototype slips, momentum dies.
The team builds demos instead of production workflows. Impressive sprint reviews; no production traffic.
The board expects savings before adoption. Q1: “where’s the headcount reduction?” CEO panics, cuts, adoption breaks.

The shared root cause across all ten: treating this as a technology project instead of an operating-model change.

Summary

Summary table

Step	Who runs it	Time	Cost (mid-market)	Make vs Buy	Difficulty
1. Audit fragmentation	Internal champion (COO/CoS) + 1–2 consultants	4–8 weeks	$75K–$200K	Process, not product	Low
2. Build substrate	2–4 engineers (internal hire / embedded / hybrid)	12–24 weeks	$500K–$2M v1; $50K–$200K/mo ongoing	Foundations buy, substrate build	High
3. Deploy first internal agent	1–2 engineers + workflow owner	6–10 weeks	$150K–$400K	Mostly build on substrate	Medium
4. Pick wedge workflow	2–4 engineers + workflow owner + functional team	10 wks proto; 8-wk iteration	$300K–$700K + $75K–$200K/mo	Build on substrate	Medium
5. Embed engineers	3–8 engineers ongoing	6–24+ months continuous; year-3 internal team ~$1.2M–$2M/yr	$1.5M–$5M/year	Buy services; hire as capability transfers	High (ongoing)
6. Re-architect operating model	CEO + leadership team	6–18 months	$500K–$2M (consulting) + restructuring costs	Internal leadership work	Highest (org)
People Track (parallel)	CEO + CHRO + COO + Internal Comms	Continuous from week 1	$300K–$1.5M/year + comp/restructuring costs	Mostly internal + change-mgmt consulting	High (cultural — single largest cause of failure)

The realistic picture for a 50–2,000 person company:

Time to first measurable P&L impact: 6–9 months
Time to AI-first operating model: 18–24 months
Total investment over 24 months: ~$3M–$10M
Enterprise scale (5,000+ employees): 2–4× those numbers

Methodology note. Industry-pattern estimates based on comparable transformation projects (cloud migration, ERP, prior AI deployments). Directional, not diligence-grade. The cited companies have not publicly disclosed exact build costs. For PE diligence, validate against the specific company’s revenue, employee count, stack maturity, data infrastructure, and regulatory profile.

For PE Firms

Closer

The 6-step playbook is built for a single company. At portfolio level it becomes a repeatable operating playbook that compounds.

One PE firm running these six steps across 20–80 portcos generates operating alpha no individual portco can match. The reference architecture amortizes; vendor contracts consolidate; eval and security baselines reuse; change-management materials carry across. Each portco’s transformation gets faster than the last.

Pilots without a shared substrate scatter. Pilots with one compound.

The first move a thoughtful PE firm should make is not funding portco AI projects one at a time — it is standardizing the substrate pattern across the portfolio so every later project compounds.

Ready to actually start? The companion piece — How to Start — covers the first 30 days, three engagement options, the workflow scoring matrix, readiness scorecard, and buyer FAQ.

Read How to Start →

This is what we work on at Synaptrix. If it resonates, we’d like to talk.