resources

From pilot to AI-native: the four stages and how to know which one you’re in

An AI maturity model for mid-market companies, mapped to Gartner, MIT CISR, BCG and Microsoft’s frameworks, with the one test that reveals your real stage and a ten-statement self-assessment.

September 10, 2026

11 min read

WHY STAGES MATTER

THE FOUR STAGES

MOVING UP

TL;DR

We use four stages: Exploring, Piloting, Scaling and AI-native. One question separates the middle two, and it is the one that matters: did the second AI use case reuse the first one’s data foundation, or rebuild its own?

Every major framework draws the same lopsided ladder. MIT CISR puts 28% of companies experimenting, 34% piloting, 31% industrializing and 7% future-ready; BCG finds 60% laggards, 35% scalers and 5% future-built; Gartner counts 22% scaled across more than one business unit.

The value concentrates at the top. BCG’s future-built companies report 1.7x the revenue growth and 1.6x the EBIT margin of the rest, and Gartner’s high-maturity organizations keep AI in production for three or more years at more than twice the rate of low-maturity ones (45% vs 20%).

The move from Piloting to Scaling is a foundation project, not an AI project: systems of record, written definitions, maintained integrations, designed permissions and a review cadence. A ten-statement self-assessment below places a company in about twenty minutes.

Two companies, the same tools, a different next move

Picture two 800-person software companies. Both run Salesforce. Both switched on Agentforce this year, both have a service agent handling tier-one cases, and both have sellers drafting outreach with AI. From the outside they are indistinguishable. Inside, one is about to spend the next year building a second and third agent on the same data layer as the first. The other is about to spend the next year discovering that the first agent’s data layer was a spreadsheet.

That difference is what a maturity model is for. An AI maturity model describes how deeply AI is embedded in a company’s data, workflows and decisions, from individual experimentation to a business designed around it. The point is not to score well. The point is that the correct next move is different at every stage, and most failed AI investments are a reasonable move made at the wrong stage: scaling before the foundation exists, or building foundations for a use case nobody owns.

The evidence for sequencing is consistent across firms that rarely agree. Accenture’s May 2025 study of 2,000 companies found that scaling one strategic bet made a company roughly three times likelier to beat its ROI expectations. BCG’s October 2024 research on AI leaders found they pursue half as many opportunities as everyone else and scale twice as many. Both describe companies that knew which stage they were in and acted on it.

Every survey draws the same ladder

There is no shortage of maturity frameworks, and they agree more than their authors might like. Gartner’s AI maturity model uses five levels; its June 2025 survey scored high-maturity organizations at 4.2 to 4.5 on that scale and low-maturity ones at 1.6 to 2.2, with a wide and mostly empty middle. MIT’s Center for Information Systems Research uses four stages and, unusually, published the distribution. Across 721 companies in December 2024: Experiment and Prepare, 28%; Build Pilots and Capabilities, 34%; Industrialize AI, 31%; AI Future-Ready, 7%.

BCG’s September 2025 Widening AI Value Gap study sorted the companies of 1,250 executives into three tiers: 60% laggards, 35% scalers and 5% future-built. The future-built group reported 1.7 times the revenue growth, 1.6 times the EBIT margin and 3.6 times the three-year shareholder return of the rest, and a third of them were already running agents against 12% elsewhere. Microsoft’s 2025 Work Trend Index, drawn from 31,000 workers, named its top tier the Frontier Firm: 24% of surveyed companies had deployed AI company-wide, 12% were still piloting, and 71% of employees at Frontier Firms said their company was thriving, against 39% everywhere else. Microsoft’s July 2026 strategy roadmap adds a five-step sequence (Exploring, Planning, Implementing, Scaling, Realizing), and BCG describes the journey as Deploy, Reshape, Invent.

Read side by side, the frameworks share three properties. The bottom rung is always individual or experimental use. The top rung is always a company where AI changes what the business attempts, not just how fast it works, which is McKinsey’s June 2026 test for AI-native companies. And the population thins out sharply above the middle: 7% in MIT’s top stage, 5% in BCG’s, and about 1% of leaders in McKinsey’s Superagency research willing to call their own company mature.

We use four stages rather than five or six because a 300–2,000-person company does not need a finer scale, and because four stages put the most important transition, from Piloting to Scaling, in the center of the picture with one test attached. The ladder below is that picture.

Diagram: The one test: did the second use case reuse the first one’s foundation?

The four stages, from the inside

What follows is each stage as it looks in a company running Salesforce or HubSpot, with the tell that identifies it and the trap that keeps companies there.

Stage 1: Exploring

Employees use AI individually. Reps draft emails with ChatGPT, support runs a summarizer, someone in finance built a custom GPT for board-deck prose. There is no shared data foundation, no owned use case and no metric. The value is real and entirely personal. Gartner’s May 2026 finding that 88% of employees with enterprise AI tools also use personal ones is a portrait of this stage: the tools arrived, and the work did not change.

The tell: nobody can name a business number AI is supposed to move. The trap: comfort. Risk is close to zero here, and so is progress. This is MIT CISR’s Experiment and Prepare (28%), Gartner’s first two levels, and the bulk of BCG’s laggards.

Stage 2: Piloting

A team has connected AI to one real workflow. There is a sponsor, a demo and a Slack channel. The data is usually prepared by hand: an export from the CRM, cleaned in a sheet, loaded somewhere the model can reach it. There is rarely a system of record everyone agrees on, a written definition set, or an integration that will survive the pilot’s champion changing jobs.

This is where most companies live. MIT CISR’s Build Pilots and Capabilities stage holds 34% of companies; add Stage 1 and it is 62%. Deloitte’s 2026 State of AI in the Enterprise found only 25% of companies had moved more than 40% of their pilots into production, and McKinsey’s June 2026 research describes two-thirds as not yet scaled beyond pilots. MIT NANDA’s finding that 95% of generative-AI pilots show no measurable P&L impact is this stage measured at the bottom line.

The tell: the second use case rebuilt its own data foundation. The trap: activity looks like maturity from the inside. Three pilots running in parallel feel like scaling. They are three Stage 2 projects sharing a budget.

Stage 3: Scaling

Several workflows run on AI inside governed systems. Each has a named owner and a recorded baseline. The distinguishing feature is reuse: the second use case ran on the first one’s pipelines, definitions and permission sets, and the third was cheaper than the second. Governance exists as a rhythm, with someone reviewing what the agents did and someone able to switch them off.

This maps to MIT CISR’s Industrialize AI (31%), Gartner’s upper levels, BCG’s scalers (35%) and the 24% of companies Microsoft found with company-wide deployment. Gartner’s September 2026 survey adds a habit that separates this stage from the one below it. Companies that track AI ROI as a portfolio rather than project by project see positive returns on 81% of their initiatives.

The tell: cost per use case falls with each addition. The trap: the foundation stops being anyone’s job once it works. Data decays, integrations drift, and a Stage 3 company can slide back to Stage 2 without noticing until an agent acts on a stale Account in front of a customer.

Stage 4: AI-native

AI is assumed. Processes are designed with agents as participants, roles are written on the basis that routine decisions are automated, and the review of what AI did is a standing agenda item rather than an incident. The company builds workflows that would not exist without it. McKinsey’s June 2026 phrase is that AI is “not a tool, it’s a teammate,” and its test is whether the business now attempts things it could not attempt before.

This is MIT CISR’s AI Future-Ready (7%), BCG’s future-built (5%) and Microsoft’s Frontier Firm. Few mid-market companies are here, and the ones that are did not arrive by buying more tools. They arrived by finishing Stage 3 and then letting it change how they plan.

The one test

If you remember one thing from this post, make it a question: when you built your second AI use case, did it reuse the data foundation of the first one, or did it build its own? If it rebuilt, you are in Stage 2, whatever the count of agents running. If it reused, you are in Stage 3, and the path to Stage 4 is mostly discipline. The self-assessment below asks that question from several angles. First, the numbers behind the ladder, because the same shape shows up in four surveys that used four different rulers.

What we actually see

A pattern from our own engagements, offered as observation rather than data. Most companies we meet describe themselves as Scaling. Most are Piloting. The gap is not dishonesty; three active pilots, a steering committee and a vendor relationship feel like a program from the inside, and nobody has asked the reuse question out loud. RSM’s 2025 middle-market survey of 966 executives shows the same gap at scale: 91% use generative AI and 79% say they have a roadmap, but only 37% call that roadmap well-formulated and only 25% have AI fully integrated into core operations.

The second pattern is counterintuitive. Stage 2 companies often have an excellent pilot, and Stage 3 companies often have a mediocre one. What differs is whether the foundation under the first use case outlived the person who built it. We have watched a very good forecasting pilot die in the week its owner moved to another company, because the Opportunity export it depended on was a saved report in her personal folder. That is an inference about cause from a small number of cases, not a study, but it repeats often enough that we now check for personal-folder reports on day one.

A self-assessment you can run in one meeting

Ten statements. Answer each true or false with the executive team in the room. “Mostly” is false.

1. One person can name the system of record for customer, deal, invoice and ticket, and sales, marketing, finance and service give the same answer.
2. The definitions of lead, opportunity, customer, churn and revenue are written down and version-controlled.
3. At least one AI workflow runs in production on live CRM data, not on an export.
4. That workflow has a named owner whose goals contain the metric it is supposed to move.
5. A baseline for that metric was recorded before go-live.
6. The second AI use case reused the first one’s pipelines, definitions and permissions.
7. The most recent use case cost less and launched faster than the one before it.
8. Someone reviews what the agents did on a fixed cadence, and someone can switch them off.
9. Agent permissions were designed deliberately and match those of the human role each agent assists, and no more.
10. At least one process and one role have been redesigned assuming AI, rather than accelerated by it.

Scoring: zero to two true is Exploring; three to five is Piloting; six to eight is Scaling; nine or ten is AI-native or close to it. Two footnotes. Statement 6 outweighs the rest: if it is false, you are Piloting regardless of the total. And a false on statements 1 and 2 alongside a true on 3 is the most common pattern we see and the most fragile, because it describes a production workflow standing on a foundation nobody agreed.

Moving up

Each transition asks for something different, and the middle one asks for the most.

Exploring to Piloting: pick one workflow with a number the board already tracks, name an owner, and connect AI to real data even if that data is hand-prepared for now. The goal is a working demo with a recorded baseline, not a platform.
Piloting to Scaling: this is the transition that matters, and it is a foundation project. Declare the system of record per object. Write the definitions. Replace exports with maintained integrations. Design the agent’s permission set. Put a review cadence on the calendar. Then build the second use case on top of all of that and measure whether it was cheaper than the first.
Scaling to AI-native: redesign rather than accelerate. Deloitte’s AI Institute test applies here: is AI speeding up the existing process, or has the team rethought the process itself? Rewrite the roles, put AI review on the leadership agenda, and rank the next use cases by what the business could newly attempt rather than by what is easiest to automate.

Expect the middle transition to take about a quarter of foundation work in a mid-market company with Salesforce or HubSpot as its system of record, and expect the returns to lag the effort. Deloitte’s ROI research puts the typical AI use case at two to four years to satisfactory return against the seven to twelve months executives expect, with only 6% paying back inside a year. Companies with the foundation in place are the ones that beat those averages: PwC’s January 2026 survey of 4,454 CEOs found those with strong foundations were three times likelier to report returns, and Gartner’s high-maturity organizations keep AI in production for three or more years at 45% against 20% for the rest.

None of this requires a bigger AI budget. It requires knowing which rung you are standing on and refusing to skip the next one.

Frequently asked questions

What is an AI maturity model?

An AI maturity model is a staged description of how deeply AI is embedded in a company’s data, workflows and decisions, from individual experimentation to a business designed around it. Its practical use is sequencing: the right next investment is different at each stage, and most failed AI spending is a reasonable move made one stage too early.

What are the four stages of AI maturity?

Exploring (individuals use AI tools, nothing reaches the P&L), Piloting (one workflow on hand-prepared data with a sponsor and a demo), Scaling (several governed workflows sharing one data foundation, each with an owner and a baseline) and AI-native (processes and roles designed around AI, with governance as routine). They map closely to MIT CISR’s four stages and BCG’s laggards, scalers and future-built.

What does AI-native mean for a company?

An AI-native company assumes AI in how it designs processes, defines roles and reviews decisions, and it builds workflows that would not exist without it. McKinsey’s test is whether the business now attempts things it could not before. It is rare: MIT CISR counts 7% of companies as future-ready and BCG counts 5% as future-built.

How do I know which stage of AI maturity my company is in?

Ask whether your second AI use case reused the first one’s data foundation or rebuilt its own. Rebuilt means Piloting, whatever the number of pilots running; reused means Scaling. The ten-statement self-assessment in this post refines that answer, and the two statements that matter most are whether a system of record exists per object and whether definitions are written down.

What percentage of companies have actually scaled AI?

Between a fifth and a third, depending on the ruler. Gartner’s September 2026 survey finds 22% have scaled AI across more than one business unit; MIT CISR puts 31% at industrializing and 7% at future-ready; BCG counts 35% scalers and 5% future-built; Microsoft found 24% with company-wide deployment. About 1% of leaders in McKinsey’s research call their company mature.

How long does it take to move from pilot to scale?

For a 300–2,000-person company on Salesforce or HubSpot, the foundation work for the Piloting-to-Scaling transition typically fits in a quarter, with the second use case shipping in the quarter after. Returns take longer: Deloitte finds the typical AI use case needs two to four years to reach satisfactory ROI, and only 6% pay back within a year.

NEXT STEP

See where your organization stands before the next AI purchase.

A short conversation is enough to tell whether the AI Readiness Health Check, 36 checks on your Salesforce or HubSpot org, is the right first step. No decks, no obligations.

book a strategy session

Read the 36 checks →

WRITTEN BY

Tal Oryon

MORE NOTES

Perspective

The world is changing. Is your organization part of it?

In one year the platforms your company runs on rebuilt their pricing, their products and their partnerships around agents. A note to executives on the one decision that is still theirs.

AI transformation

An AI transformation roadmap for a 300–2,000-person company (12 months)

Four quarters, one workflow at a time. The deliverables, the owners and the failure mode of each quarter, written for mid-market companies that cannot afford an enterprise timeline.

ACTIVEAPPS