The 14 Levels of an AI Rollout, Translated for a Mid-Market Budget

A rollout map did the rounds among AI operators recently: fourteen numbered levels, running from the first company-wide audit all the way to agents that guide their own work. It is a good map. It is also written for a company with a chief data officer, an enablement team, and a budget that treats a seven-figure transformation as a rounding error.

Most of the businesses we work with have none of that. They sit between $1M and $50M in revenue, the leadership team already works weekends, and exactly one person half-understands the CRM. The fourteen levels still apply to them. They just cost a fraction as much, and they almost never arrive in the tidy numbered order the map implies. This post walks the whole progression the way it actually plays out at mid-market scale, with a link to the deeper method at each stage.

One thing to hold onto before we start: the levels are a checklist of things that need to be true, not a Gantt chart. We have watched companies run a hackathon (level nine) before they finished their audit (level one), and get away with it. The value is in knowing what each stage is for, so you can tell which one you are actually stuck on.

Levels 1 and 2: Audit First, Then Sequence by ROI

The map starts with a company-wide audit: map the key processes, interview the leadership team, survey the people doing the work. Then a readout that lays out the transformation as a sequence, prioritised by return, risk, and how much cultural support each initiative has.

This is the half everyone wants to skip, and skipping it is why so many AI programmes are a graveyard of half-built pilots. You do not need the sixty-page enterprise version. You need a week: sit with the five people who run operations, watch where the hours actually go, and leave with a shortlist of processes ranked by value and readiness. The readout is not a strategy document, it is a queue. We wrote the mid-market version of this as its own piece: how to run an AI audit that actually leads somewhere.

The sequencing matters as much as the finding. An initiative with real ROI that the team will quietly sabotage is worse than a smaller win they will champion. Rank on all three axes, not just the spreadsheet.

Level 3: The Company Brain Is a Retrieval Layer, Not a Warehouse

Level three is where companies realise their data ducks are not in a row and reach for the "company brain". This is the point where a lot of budgets die, because the reflex is to buy an eighteen-month data platform rebuild.

You do not need one. A modern agent does not need your data in one place, it needs permission to read the places your data already lives. A thin retrieval layer over your top three systems, connected read-only, gets agents producing useful answers in weeks. The grand unified schema can wait until the returns pay for it, which they mostly never will. The full argument, including how to capture the knowledge that only lives in people's heads, is in building the company brain without rebuilding your data stack.

Level 4: Coding Agents Are the Cheapest Place to Learn

The map puts coding agents for the engineering and product team early, in tandem with the data work, and that ordering is right for a reason that has nothing to do with code. Developers are the one group who can read what an agent did and catch it when it is wrong. That makes engineering the safest place in the company to build the muscle of working with agents.

At mid-market scale you may not have a product org. That is fine, the principle holds: start where the people can verify the output. The risk is real, though, because coding agents touch your source and your systems. Roll them out with access tiers and review gates rather than a company-wide free-for-all. We laid out how in rolling out Claude Code and Cowork without betting the company data.

Levels 5 to 8: Access and Training Follow Capability, Not Job Title

The next four levels are all access and enablement: enterprise access for a subset of non-technical champions, then everyone, then workshops for the leadership team and champions, then company-wide training.

The instinct is to grant access by seniority. Grant it by capability instead: start with the people who can tell a good output from a confident wrong one, and widen the circle as your review tooling matures. Access without the judgement to check the work is how you get plausible nonsense shipped to clients. The governance model that makes this concrete, freedom inside guardrails, is here.

Training is not a one-off event either. People move from AI-literate to AI-enabled to AI-first at different speeds, and the enablement has to meet them where they are. The roadmap that actually sticks is in AI literate, AI enabled, AI first.

Levels 9 and 10: Run a Hackathon to Find the Work Worth Doing

Level nine is an internal hackathon that lets employees surface solutions to problems from the audit, plus new ones nobody wrote down. Level ten is leadership reviewing those prototypes and deciding which ones graduate to production.

This is one of the highest-return two days a mid-market company can spend, because the best AI use case in your business is almost never the one the leadership team guessed. It is the workaround the accounts clerk built in her lunch break. A hackathon drags that knowledge into the open. The trap is treating the demo as the deliverable, when the real work is turning the winners into something that survives Monday morning. The full playbook is in the internal AI hackathon playbook.

Level 11: Your First Build Should Be a Boring Win

The map is blunt about the first real build: a quick win to some back-office process, with attributable hard ROI, where cultural pushback is expected to be low. Not the ambitious client-facing thing. The boring one.

This is the most under-rated instruction in the whole progression. Your first AI project is not really about the ROI on that project, it is about buying credibility and runway for the next ten. A visible, uncontested win in the back office does that. A stalled moonshot in sales does the opposite. How to pick it, and what to avoid for build number one, is in how to choose your first AI project.

Level 12: Token Spend Becomes the New Cloud Bill

Level twelve is the one nobody plans for: cost optimisation becomes a major focus as the AI budget starts to balloon, usually in the engineering org first. The finance team notices, and the honeymoon ends.

Token spend behaves like your cloud bill did a decade ago: invisible until it is not, then suddenly a line item with a name attached. The fix is the same discipline, applied early: know which workflows cost what, cache the expensive context, and route each job to the cheapest model that clears your quality bar. The full treatment is in token spend is the new cloud bill.

Levels 13 and 14: From Deterministic Workflow to Self-Guided Agent

The last two levels describe a loop, not a destination: identify a problem, map the process, prototype, test and harden and secure and measure, then scale, drawing on the audit and the hackathon findings each time round. And as the muscle grows, the company moves along the spectrum from a fixed deterministic workflow toward a self-guided agent.

The word that matters there is spectrum. Most of the value for a mid-market business lives in the boring middle, well short of full autonomy, and handing an agent more control is a decision to make on purpose, not a milestone to rush. When to make that move, and the guardrails that have to be in place first, is in workflow or agent, and when to graduate. The measurement discipline that keeps the whole loop honest is in how to evaluate AI agents.

Where to Start This Quarter

Do not try to climb all fourteen levels this year. Find the one you are actually stuck on. Most companies we meet are stuck at level one, mistaking activity for a plan, or at level eleven, arguing about an ambitious build while a boring win sits unclaimed.

Run the audit this week, pick one back-office win, and put it through end to end. That is levels one, two, and eleven in a fortnight, and it earns you the right to the rest. If you want an outside read on where you sit and what to do next, our Deep Dive maps your processes, scores your use cases, and hands you the sequenced plan, grounded in what your data and your team can support today.